Computational Neurogenetic Modeling (Topics in Biomedical Engineering. International Book Series)

Computational Neurogenetic Modeling

TOPICS IN BIOMEDICAL ENGINEERING INTERNATIONAL BOOK SERIES Series Editor: Evangelia Micheli-Tzanakou Rutgers University Piscataway, New Jersey

Signals and Systems in Biomedical Engineering: Signal Processing and Physiological Systems Modeling Suresh R. Devasahayam

Models of the Visual System Edited by George K. Hung and Kenneth J. Ciuffreda

PDE and Level Sets: Algorithmic Approaches to Static and Motion Imagery Edited by Jasjit S. Suri and Swamy Laxminarayan

Frontiers in Biomedical Engineering: Edited by Ned H.C. Hwang and Savio L-Y. Woo

Handbook of Biomedical Image Analysis: Volume I: Segmentation Models Part A Edited by Jasjit S. Suri, David L. Wilson, and Swamy Laxminarayan

Handbook of Biomedical Image Analysis: Volume II: Segmentation Models Part B Edited by Jasjit S. Suri, David L. Wilson, and Swamy Laxminarayan

Handbook of Biomedical Image Analysis: Volume III: Registration Models Edited by Jasjit S. Suri, David L. Wilson, and Swamy Laxminarayan

Complex Systems Science in Biomedicine Edited by Thomas S. Deisboeck and J. Yasha Kresh Computational Neurogenetic Modeling Lubica Benuskova and Nikola Kasabov

A Continuation Order Plan is available for this series. A continuation order will bring delivery of each new volume immediately upon publication. Volumes are billed only upon actual shipment. For further information please contact the publisher.

Computational Neurogenetic Modeling

Dr. Lubica Benuskova Senior Research Fellow Knowledge Engineering and Discovery Research Institute AUT, Auckland, New Zealand

and

Professor Nikola Kasabov Founding Director and Chief Scientist Knowledge Engineering and Discovery Research Institute AUT, Auckland, New Zealand

~ Springer

Dr. Lubica Benuskova Senior Research Fellow Knowledge Engineering and Discovery Research Institute, www.kedri.info AUT, Auckland, New Zealand [email protected]

Professor Nikola Kasabov Founding Director and Chief Scientist Knowledge Engineering and Discovery Research Institute, www.kedri.info AUT, Auckland, New Zealand [email protected]

Library of Congress Control Number: 200693690 I

ISBN-lO: 0-387-48353-5 ISBN-13: 978-0-387-48353-5

eISBN-lO: 0-387-48355-1 eISBN-13: 978-0-387-48355-9

Printed on acid-free paper. © 2007 Springer Science + Business Media, LLC All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science + Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights.

9 8 765 4 3 2 1 springer.com

Dedication

To the memory ofmy parents.

Lubica

To my mother and the memory ofmy father.

Nikola

Preface

It is likely that future progress in many important areas of science (e.g.

brain science, bioinformatics, information science, physics, communication engineering and social sciences) can be achieved only if the areas of computational intelligence, brain science and bioinformatics share and integrate their methods and knowledge. This book offers some steps in this direction. The book presents the background knowledge and methods for the integration of gene information and brain activity information with the purpose of the creation of biologically plausible computational models aiming at modeling and understanding the brain. The book is aiming at encouraging research in information sciences in the direction of human-like and human-oriented information processing. In this context, "human-like" means that principles from the brain and genetics are used for the creation of new computational methods. "Humanoriented" means that these machines can be used to discover and understand more about the functioning of the brain and the genes, about memory and learning, about speech and language, about image and vision, and about ourselves. This work was partially supported by the research grant AUTX0200l "Connectionist-based intelligent information systems", funded by the New Zealand Foundation for Research, Science, and Technology - FRST, through the New Economy Research Fund - NERF. There are a number of people whom we would like to thank for their encouragement and contribution to the book. These are several colleagues, research associates and postgraduate students we have worked with at the Knowledge Engineering and Discovery Research Institute in Auckland, New Zealand, in the period 2002- 2006: Dr Qun Song, Dr Zeke S. Chan, Dr Paul S. Pang, Dr Liang Goh, Vishal Jain, Tian-Min Ma (Maggie), Peter Hwang, Paulo Gottgtroy, Natalia Bedran, Joyce D'Mello and especially Simei Gomes Wysoski, who did the technical edition of the book and developed specialized simulation tools we used for our computer experiments. We appreciate the discussions we had with a number of colleagues from different laboratories and countries. Among them are Walter Freeman from University of California at Berkeley; Takeshi Yamakawa - Kyushu

VIII

Preface

Institute of Technology ; John G. Taylor - Kings College, London; Cees van Leuwen - RIKEN, Japan; Michael Arbib - University of Southern California ; Dimiter Dimitrov - National Cancer Institute in Washington DC; Alessandro E. P. Villa -University of Lausanne, Bassem Hassan Catholic University of Leuven, Gary Markus - New York. We thank the Springer, New York team and especially Aaron Johnson and Beverly Rivero for the encouragement and patience. When one of the authors, N. Kasabov, presented his talk at the ICONIP 2002 in Shanghai and suggested that gene interaction information should be used in biologically plausible neural network models, Walter Freeman commented "Yes, that makes sense, but how do we do that?" Michael Arbib when visiting Auckland in 2004 made a remark that integrating genes (molecular level) into neural networks may require to go to atom (quantum) level. This book is presenting some initial answers to these questions . It presents the foundations and the concepts of computational neurogenetic modeling (CNGM), initially introduced by the authors in 2004 (Kasabov and Benuskova 2004). The book was written by the two authors in a close collaboration, where Lubica Benuskova wrote chapters 2,3 ,8,9, 10 and compiled Appendix I , while Nikola Kasabov wrote chapters 1,4,5,6, 7 and compiled Appendix 2 and 3. Each of the authors also had a smaller contribution to the other chapters as well. The book is intended for postgraduate students and researchers in the areas of information sciences, artificial intelligence, neurosciences , bioinformatics, and cognitive sciences. The book is structured so that every chapter can be used as a reading material for research oriented courses at a postgraduate level. Additional materials, including: data, simulation programs, lecture notes, color figures, etc. can be found on the web site www.kedri.info. Dr Lubica Benuskova Prof. Dr. Nikola Kasabov Knowledge Engineering and Discovery Research (www.kedri.info), Auckland University of Technology Auckland, New Zealand I September 2006

Institute

Contents

Preface 1 Computational Neurogenetic Modeling (CNGM): A Brief Introduction 1.1 Motivation - The Evolving Brain 1.2 Computational Models of the Brain 1.3 Brain-Gene Data, Information and Know ledge 1.4CNGM: How to Integrate Neuronal and Gene Dynamics? 1.5 What Computational Methods to Use for CNGM? 1.6 About the Book 1.7 Summary 2 Organization and Functions of the Brain 2.1 Methods of Brain Study 2.2 Overall Organization of the Brain and Motor Control 2.3 Learning and Memory 2.4 Language and Other Cognitive Functions 2.4.1 Innate or Learned? 2.4.2 Neural Basis of Language 2.4.3 Evolution of Language, Thinking and the Language Gene 2.5 Neural Representation ofInformation 2.6 Perception 2.7 Consciousness 2.7.1 Neural Correlates of Sensory Awareness 2.7.2 Neural Correlates of Reflective Consciousness 2.8 Summary and Discussion

VII

1 1 .4 6 12 14 15 16 17 18 23 25 29 29 30 33 36 37 41 .41 .44 .49

3 Neuro-Information Processing in the Brain 53 53 3.1 Generation and Transmission of Signals by Neurons 3.2 Learning Takes Place in Synapses: Toward the Smartness Gene 56 58 3.3 The Role of Spines in Learning 3.4 Neocortical Plasticity 61

X

Contents 3.4.1 Developmental Cortical Plasticity 61 3.4.2 Adult Cortical Plasticity 64 3.4.3 Insights into Cortical Plasticity via a Computational Model... 66 3.5 Neural Coding: the Brain is Fast, Neurons are Slow 74 3.5.1 Ultra-Fast Visual Classification 74 3.5.2 Hypotheses About a Neural Code 77 Coding Based on Spike Timing 77 The Rate Code 77 3.6 Summary 78

4 Artificial Neural Networks (ANN) 4.1 General Principles 4.2 Models of Learning in Connectionist Systems 4.3 Unsupervised Learning (Self Organizing Maps - SOM) 4.3.1 The SOM Algorithm 4.3.2 SOM Output Sample Distribution Clustering Information Visualization of Input Variables Relationship Between Multiple Descriptors The Connection Weights Interpretation by the Fuzzy Set Theory 4.3.3 SOM for Brain and Gene Data Clustering 4.4 Supervised Learnin g 4.4.1 Multilayer Perceptron (MLP) 4.4.2 MLP for Brain and Gene Data Classification Example 4.5 Spiking Neural Networks (SNN) 4.6 Summary

81 81 84 93 93 95 95 95 96 96 96 97 97 98 98 99 99 102 105

5 Evolving Connectionist Systems (ECOS) 5.1 Local Learning in ECOS 5.2 Evolving Fuzzy Neural Networks EFuNN 5.3 The Basic EFuNN Algorithm 5.4 DENFIS 5.4.1 Dynamic Takagi-Su geno Fuzzy Inference Engine 5.4.2 Fuzzy Rule Set, Rule Insertion and Rule Extraction 5.5 Transductive Reasoning for Personalized Modeling 5.5.1 Weighted Data Normalization 5.6 ECOS for Brain and Gene Data Modeling 5.6.1 ECOS for EEG Data Modeling, Classification and Signal Transition Rule Extraction

107 107 108 112 116 118 119 120 122 122 122

XI

5.6.2 ECOS for Gene Expression Profilin g 5.7 Summary

124 126

6 Evolutionary Computation for Model and Feature Optimization .127 6.1 Lifelong Learning and Evolution in Biological Species: Nurture vs. Nature 127 128 6.2 Principl es of Evolutionary Computation 6.3 Genetic Algorithms 128 6.4 EC for Model and Parameter Optimization 133 6.4. 1 Example 133 6.5 Summary 136

7 Gene/Protein Interactions - Modeling Gene Regulatory Networks (GRN) 137 7.1 The Central Dogma of Molecular Biology 7.2 Gene and Protein Expression Data Analy sis and Modeling 7.2.1 Example 7.3 Modeling Gene/Prote in Regulato ry Networks (GPRN) 7.4 Evolving Connectionist Systems (ECOS) for GRN Modeling 7.4.1 General Principles 7.4.2 A Case Study on a Small GRN Modeling with the Use of ECOS 7.5 Summary

8 CNGM as Integration of GPRN , ANN and Evolving Processes 8.1 Modeling Genetic Control of Neural Development 8.2 Abstract Computational Neurogenetic Model 8.3 Continuous Model of Gene-Protein Dynamics 8.4 Towards the Integration of CNGM and Bioinformatics 8.5 Summary

9 Application of CNGM to Learning and Memory 9.1 Rules of Synaptic Plasticity and Metaplasticity 9.2 Toward a GPRN of Synaptic Plasticity 9.3 Putative Molecular Mechanisms of Metaplasticity 9.4 A Simple One Protei n-One Neuronal Funct ion CNGM 9.5 Application to Modeling ofL-LTP 9.6 Summary and Discussion 10 Applications of CNGM and Future Development 10.1 CNGM of Epilepsy 10.1.1 Genetically Caused Epilepsies

137 141 143 145 150 150 151 153

155 156 161 165 171 175

177 177 185 193 196 198 202

205 206 206

XII

Contents

10.1.2 Discussion and Future Developments 10.2 CNGM of Schizophrenia 10.2.1 Neurotransmitter Systems Affected in Schizophrenia 10.2.2 Gene Mutations in Schizophrenia 10.2.3 Discussion and Future Developments 10.3 CNGM of Mental Retardation 10.3.1 Genetic Causes of Mental Retardation 10.3.2 Discussion and Future Developments 10.4 CNGM of Brain Aging and Alzheimer Disease 10.5 CNGM of Parkinson Disease 10.6 Brain-Gene Ontology 10.7 Summary

209 210 212 214 217 2 18 219 223 224 229 232 235

Appendix 1.•................................................•..............•....•••..............•...... 237 237 A.I Table of Genes and Related Brain Functions and Diseases Appendix 2 A.2 A Brief Overview of Computational Intelligen ce Methods A.2.1 Probabil istic and Statistical Methods Stochastic Models A.2.2 Boolean and Fuzzy Logic Models Boolean Models Fuzzy Logic Models A.2.3 Artificial Neural Networks Evolving Classifier Function (ECF) A.2.4 Methods of Evolutionary Computation (EC)

247 247 247 250 250 250 251 253 254 256

Appendix 3 257 A.3 Some Sources of Brain-Gene Data, Information, Knowledge and Computational Models 257

References

259

Index

287

1 Computational Neurogenetic Modeling (CNGM): A Brief Introduction

This chapter introduces the motivation and the main concepts of computational neurogenetic modeling (CNGM). It argues that with the presence of a large amount of both brain and gene data related to brain functions and diseases, it is required that sophisticated computational models are created to facilitate new knowledge discovery that helps understanding the brain in its complex interaction between genetic and neuronal processes. The chapter points to sources of data, information and knowledge related to neuronal and genetic processes in the brain. CNGM is concerned with the integration of all these diverse information into a computational model that can be used for modeling and prediction purposes. The models integrate knowledge from mathematical and information sciences (e.g. computational intelligence - CI), neurosciences, and genetics. The chapter also discusses what methods can be used for CNGM and how. The concepts and principles introduced in this chapter are presented in detail and illustrated in the rest of the book.

1.1 Motivation - The Evolving Brain According to the Concise Oxford English Dictionary (1983), "evolving" means "revealing", "developing". It also means "unfolding, changing". The term "evolving" is used here in a broader sense than the term "evolutionary". The latter is related to a population of individual systems traced over generations (Darwin 1859, Holland 1975, Goldberg 1989), while the former, as it is used in this book, is mainly concerned with a continual change of the structure and the functionality of an individual system during its lifetime (Kasabov 2003, Kasabov 2006). In living systems and in the human brain in particular, evolving processes are observed at different levels (Fig. 1.1) (Kasabov 2006). At the quantum level, particles are in a complex evolving state all the time, being at several locations at the same time, which is defined by probabilities.

2

I Computational Neurogenetic Modeling(CNGM): A BriefIntroduction

At a molecular level, DNA, RNA and protein molecules, for example, evolve and interact in a continuous way. The area of science that deals with the information processing and data manipulation at this level is Bioinformatics. At the cellular level (e.g. a neuronal cell) all the metabolic processes, the cell growth, cell division etc., are evolving processes. 6. Evolutionary (population/generation) processes 5. Brain cognitiveprocesses (learning,thinking, etc.) 4. System level information processing(e.g. auditory system) 3. Information processing in a cell (neuron) 2. Molecularlevel of information processing(genes, proteins) 1. Quantum level of information processing Fig. 1.1. Six levels of evolving processes in the brain: evolution, cognitive brain processes,brain functions in neural networks, a single neuron functions, molecular processes, and quantumprocesses

At the level of cell ensembles, or at the neural network level, an ensemble of cells (neurons) operates in a concert, defining the function of the ensemble or the network, for instance perception of sound. In the human brain, complex dynamic interactions between groups of neurons can be observed when certain cognitive functions are performed, e.g. speech and language processing, visual pattern recognition, reasoning and decision making. At the level of population of individuals, species evolve through evolution (Darwin 1859)- the top level in Fig. 1.1. Evolutionary processes have inspired the creation of computational modeling techniques called evolutionary computing (EC) (Holland 1975, Goldberg 1989). A biological system evolves its structure and functionality through both lifelong learning by an individual and the evolution of populations of many such individuals. In other words, an individual is a result of the evolution of many generations of populations, as well as a result of its own developmental lifelong learning processes. There are many physical and information processes of dynamic interaction within each of the six levels from Fig. 1.1 and across the levels. Inter-

1.1 Motivation - The Evolving Brain

3

actions are what make an organism a living one, and that is also a challenge for computational modeling. For example, there are complex interactions between DNA, RNA and protein molecules. There are complex interactions between the genes and the functioning of each neuron, a neural network, and the whole brain. Some of these interactions are known to have caused brain diseases, but most of them are unknown at present. An example of interactions between genes and neuronal functions is the dependence of development of brain with human characteristics on expression of genes like FOXP2, the gene involved in speech production (Enard et al. 2002), ASPM and Microcephalin that affect the brain size (Evans et al. 2005, Mekel-Bobrov et al. 2005), and HARIF, that is of fundamental importance in specifying the six-layer structure of the human cortex (Pollard et al. 2006). Another example is the observed dependence between long-term potentiation (learning) in the synapses and the expression of the immediate early genes and their corresponding proteins such as Zif/268 (Abraham et al. 1994). Yet another example are putative genetic mutations for many brain diseases that have been already discovered (see Appendix 1). Generally speaking, neurons from different parts of the brain, associated with different functions, such as memory, learning, control, hearing and vision, function in a similar way. Their functioning is defined by several factors, one of them being the level of neurotransmitters. These factors are controlled both through genetics and external inputs. There are genes that are known to regulate the level of neurotransmitters for different types of neurons from different areas of the brain. The functioning of these genes and the proteins produced can be controlled through nutrition and drugs. This is a general principle that can be exploited for different models of the processes from Fig. 1.1 and for different systems performing different tasks (learning, hearing etc.). We will refer to the above as neurogenetics (Kasabov and Benuskova 2004). The evolving processes in the brain are based on several major principles (Arbib 1972, Grossberg 1982, Arbib 1987, Taylor 1999, Freeman 2000, Arbib 2003, van Ooyen 2003, Marcus 2004a), such as: • Evolving is achieved through both genetically defined information and learning. • The evolved neurons have a spatial-temporal representation where similar stimuli activate close neurons. • The evolving processes lead to a large number of neurons involved in each task, where many neurons are allocated to respond to a single stimulus or to perform a single task; e.g. when a word is heard, there are millions of neurons that are immediately activated.

4

1 Computational Neurogenetic Modeling (CNGM): A BriefIntroduction

• Memory-based learning, i.e. the brain stores exemplars of facts that can be recalled at a later stage. • Evolving is achieved through interaction with the environment and other systems. • Inner processes take place, such as sleep memory consolidation. • The evolving processes are continuous and lifelong. • Through evolving brain structures, higher-level functions emerge which are embodied in the structure, and can be represented as a level of abstraction (e.g. the acquisition and the development of speech and language). The advancement in brain science, molecular biology and computational intelligence, results in a large amount of data, information and knowledge on brain functioning, brain-related genetics, brain diseases and new computational intelligence methods. All these constitute a strong motivation for the creation of a new area of science that we call computational neurogenetic modeling (CNGM), with the following general objectives: 1. To create biologically plausible neuronal models. 2. To facilitate a better understanding of the principles of the human brain, the genetic code, and life in general. 3. To enable modeling of the brain. 4. To create new generic methods of computational intelligence and a new generation of intelligent machines.

1.2 Computational Models of the Brain A project, called The Blue Brain Project, marks the beginning of a study how the brain works by building very large scale models of neural networks (http://bluebrainproject.epfl.ch/index.html). This endeavor follows a century of experimental "wet" neuroscience and development of many theoretical insights of how neurons and neural networks function (Arbib 2003). The Blue Brain Project was launched by the Brain Mind Institute, EPFL, Switzerland and IBM, USA in May, 2005. Scientists from both organizations will work together using the huge computational capacity of IBM's Blue Gene supercomputer to create a detailed model of the circuitry in the neocortex - the largest and most complex part of the human brain. The neocortex constitutes about 85% of the human brain's total mass and is thought to be responsible for the cognitive functions of language, learning, memory and complex thought. The Blue Brain Project will also build models of other cortical and subcortical parts of the brain and models of

1.2 Computational Models of the Brain

5

sensory and motor organs. By expanding the project to model other areas of the brain, scientists hope to eventually build an accurate, computerbased model of the entire brain. The project is a massive undertaking because of the hundreds of thousands of parameters that need to be taken into account. EPFL' s Brain and Mind Institute's world most comprehensive set of empirical data on the micro-architecture of the neocortex will be turned into a working 3-dimensional model recreating the high-speed electrochemical interactions of the brain's interior. The first objective is to create a software replica of the neocortical column at a cellular level for real-time simulations. An accurate replica of the neocortical column is the essential first step to simulating the whole brain. The second and subsequent phases will be to expand the simulation to include circuitry from other brain regions and eventually the whole brain. In the Blue Column, a nickname for the software replica of the neocortical column, not only the cells but an entire microcircuit of cells will be replicated (like the duplication of a tissue). Neocortical column is stereotypical in many respects from mouse to man with subtle variations in different ages, brain regions and species. The Blue Column will first be based on the data obtained from rat somatosensory cortex at 2 weeks of age because these data are the most abundant. Once built and calibrated with iterative simulations and experiments, comparative data will be used to build columns in different brain regions, ages and species, including humans. The Blue Column will be composed of 104 morphologically complex neurons with active ionic channels to enable generation of electrical currents and potentials. The neurons will be interconnected in a 3dimensional (3D) space with 107 -108 dynamic synapses. The Blue Neuron will receive about 103 -104 external input synapses and generate about 103 -104 external output synapses. Neurons will transmit information according to dynamic and stochastic synaptic transmission rules. The Blue Column will self-adapt according to synaptic learning algorithms running on 107 _10 8 synapses, and according to metaplasticity, supervised and reward learning algorithms running on all synapses. The column project will also involve the database of 3D reconstructed model neurons, synapses, synaptic pathways, microcircuit statistics, and computer model neurons. Single synapses and whole neurons will be modeled with the molecular level details however a neocortical column will be modeled at the cellular level. In the future, the research will go in two directions simultaneously: 1. The first direction will be the simplification of the column and its software or hardware duplication to build larger parts of the neocortex and eventually the entire neocortex.

6

1 Computational Neurogenetic Modeling (CNGM): A BriefIntroduction

2. The second direction will stay with a single neocortical column, moving down to the molecular level of description and simulation. This step will be aimed at moving towards genetic level simulations of the neocortical column. A very important reason for going to the molecular level is to link gene activity with electrical activity, as the director of the project Henry Markram reckons. A molecular level model of the neocortical column will provide the substrate for interfacing gene expression with the network structure and function. The neocortical column lies at the interface between the genes and complex cognitive functions. Establishing this link will allow predictions of the cognitive consequences of genetic disorders and allow reverse engineering of cognitive deficits to determine the genetic and molecular causes. It is expected that this level of simulation will become a reality with the most advanced phases of Blue Gene and Blue Brain Project development. As is the model replica of the cortical column based on many computational models of neurons, channel kinetics, learning, etc., we can ask whether there are any models of the computational neurogenetic type to be employed to model the interaction between the genes and neural networks. In other words, are there already corresponding neurogenetic models for brain functions and processes? The goal of this book is to address this question and to introduce the development of such models, which we call computational neurogenetic models (CNGMs).

1.3 Brain-Gene Data, Information and Knowledge The brain is a complex system that evolves its structure and functionality over time. It is an information-processing and control system, collaborating with the spinal cord and peripheral nerves. Each part of the brain is responsible for a particular function, for example: the Cerebrum integrates information from all sense organs, motor functions, emotions, memory and thought processes; the Cerebellum coordinates movements, walking, speech, learning and behavior; the Brain stem is involved in controlling the eyes, in swallowing, breathing, blood pressure, pupil size, alertness and sleep. A simplified view of the outer structure of the human brain is given in Fig 2.1. The structure and the organization of the brain and how it works at a higher and a lower level are explained in Chaps. 2 and 3, respectively. Since the 50-ties of the 20th century, experimental brain data gathering has been accompanied by the development of explanatory computational brain models. Many models have been created so far, for example:

1.3 Brain-Gene Data, Information and Knowledge

7

- Brain models created at USC by a team lead by Michael Arbib, at http://www-hbp.usc.edu/Projects/bmw.htm; - Mathematical brain function models maintained by the European Bioinformatics Institute (EBI): http://www.ebi.ac.uk; - Wayne State Institute Brain Injury Models at http://rtb.eng.wayne.edu/braini; - The Neural Micro Circuits Software: www.lsm.tugraz.at; - Neuroscience databases (Koetter 2003); - Genetic data related to brain (Chin and Moldin 2001) and many more (see Appendix 3). None of these brain models incorporate genetic information despite of the growing volume of data, information and knowledge on the importance and the impact of particular genes and genetic processes on brain functions. Brain functions, such as learning and memory, brain processes, such as aging, and brain diseases, such as Alzheimer, are strongly related to the level of expression of genes and proteins in the neurons (see Appendix 1, Appendix 3 and also Chaps. 9 and 10). Both popular science books and world brain research projects, such as the NCB I (the National Center for Biomedical Information), the Allen Brain Institute, the Blue Brain Project, the Sanger Centre in Cambridge, and many more, have already revealed important and complex interactions between neuronal and genetic processes in the brain, creating a massive world repository of brain-gene data, information and knowledge. Some of the information or/and references to it, are given in Appendix 1 and 3 and Chaps. 9 and 10. The central dogma ofmolecular biology states that DNA, which resides in the nucleus of a cell or a neuron, transcribes into RNA, and then translates into proteins. This process is continuous, evolving, so that proteins in turn cause genes to transcribe, etc. (Fig. 1.2). The DNA is a long, double stranded sequence (a double helix) of millions or billions of 4 base molecules (nucleotides) denoted as A, C, T and G, that are chemically and physically connected to each other through other molecules. In the double helix, they make pairs such that every A from one strand is connected to a corresponding C on the opposite strand, and every G is connected to aT. A gene is a sequence of hundreds and thousands of bases as part of the DNA, that is translated into a protein or several proteins. Only less than 5% of the DNA of the human genome contains protein-coding genes, the other part is a non-coding region that may contain useful information as well. For instance, it contains the RNA genes, regulatory regions, but mostly its function is not currently well understood.

8

1 Computational Neurogenetic Modeling (CNGM): A BriefIntroduction

The DNA of each organism is unique and resides in the nucleus of each of its cells. But what makes a cell alive are the proteins that are expressed from the genes, and define the function of the cell. The genes and proteins in each cell are connected in a dynamic regulatory network consisting of regulatory pathways - see Chap. 7. Normally, only few hundreds of genes are expressed as proteins in a particular cell. At the transcription phase, one gene is transcribed in many RNA copies and their number defines the expression level of this gene. Some genes may be "over-expressed", resulting in too much protein in the cell whereas some genes may be "under-expressed" resulting in too little protein. In both cases the cell may be functioning in a wrong way that may be causing a disease. Abnormal expression of a gene can be caused by a gene mutation - a random change in the code of the gene, where a base molecule is either inserted, or - deleted, or - altered into another base molecule. Drugs can be used to stimulate or to suppress the expression of certain genes and proteins, but how that will affect indirectly the other genes related to the targeted one, has to be evaluated and that is where computational modeling of gene regulatory networks (GRN) and CNGM can help. Output Cell Function Translation mRNA into protein production Transcription Genes copied as mRNA

DNAgenes

RNA

Proteins

Protein-gene feedback loop through Transcription Factors

Fig. 1.2. The genes in the DNA transcribe into RNA and then translate into proteins that define the function of a cell (The central dogma of molecular biology). Gene information processing in moredetails is presented in Chap. 7

It is always difficult to establish the interaction between genes and proteins. The question "What will happen with a cell or the whole organism if one gene is under-expressed or missing?" is now being answered by using a technology called "Knockout gene technology" (Chin and Moldin 2001). This technology is based on a removal of a gene sequence from the DNA

1.3 Brain-Gene Data, Information and Knowledge

9

and letting the cell/organism to develop, where parameters are measured and compared with these parameters when the gene was not missing. The obtained data can be further used to create a CNGM as described in Chap. 8. Information about the relationship between genes and brain functions is given in many sources (see Appendices 1 and 3). In the on-line published book "Genes and Diseases" (www.ncbi.nlm.nih.gov/books) the National Center for Biological Information (NCB!) has made available a large amount of gene information related to brain diseases, for example: - Epilepsy: One of several types of epilepsy, Lafora disease (progressive myoclonic, type 2), has been linked to a mutation of the gene EMP2A and EMP2B found on chromosome 6 (see Chap. 10). - Parkinson disease: Several genes: Parkin, PARK7, PTEN, alpha synuclein and others, have been related to Parkinson disease, described in 1817 by James Parkinson (see Chap. 10). - Huntington disease: Mutation in the HD gene on chromosome 4 was linked to this disease. - Sclerosis: A gene SODI was found to be related to familial amyotropic lateral sclerosis (see Chap. 10). - Rett syndrome: A gene MeCP2, on the long arm of chromosome X (Xq28), has been found to be related to this disease. The gene is expressed differently in different parts of the brain (see Fig. 1.3). The Allen Brain Institute has completed a map of most of the genes expressed in different sections of the brain of a mouse and has published it free as the Allen Brain Atlas (www.alleninstitute.org). In addition to the gene ontology (GO) of the NCBI, a brain-gene ontology (BGO) of the Knowledge Engineering and Discovery Research Institute KEDRI (www.kedri.info) contains genes related to brain functions and brain diseases, along with computational simulation methods and systems (Fig. 1.4). The BGO allows users to "navigate" in the brain areas and find genes expressed in different parts of it, or for a particular gene - to find which proteins are expressed in which cells of the brain. Example is given in Fig. 1.5. Gene expression data of thousands of genes measured in tens of samples collected from two categories of patients - control (class 1) and cancer (class 2) using micro-array equipment in relation to the brain and the central nervous system have been published by (Ramaswamy et al. 2001) and (Pomeroy et al. 2002). The first question is how to select the most discriminating genes for the two classes that can possibly be used as drug targets. The second question is to build a classifier system that can correctly

10

I Computational Neurogenetic Modeling (CNGM): A BriefIntroduction

classify (predict) for a new sample which class it is likely to belong to that can be used as an early diagnostic test. The answer to the latter question is illustrated in Chaps. 4, 5 and 6, where classification and prediction computational models are presented. Average Difference Value

:

•

.

I

~

l

..

I I I I

I

!

Fig. 1.3. The gene MeCP2, related to Rett syndrome, is expressed differently in different parts of the human brain (left vertical axis), the highest expression level being in the Cerebellum (source: Gene Expression Atlas at http://expression.gnf.org/cgi-bin/index.cgi)

Fig. 1.4. A snapshot of a structure of the Brain-Gene Ontology (BGO) of the KEDRI Institute (httpz/www.kedri.info/)

1.3 Brain-Gene Data, Information and Knowledge

11

Average 01fference Value

c....br
AtoPndQIQ Thal~'I

... iiiiiiiiiiiiliiiiiil-----------~---ri ... ... ."

Cor~s_callo ~ ~l

... l..c""' _ _

~

Fig. 1.5. The expression of the GABRA2 gene causes the production of the a2 subunit of GABA A receptor in the synapses of neurons, and it is differently expressed in different parts of the brain (source: Gene Expression Atlas at http://expression.gnf.org/cgi-bin/index.cgi)

To answer the former question, many techniques for gene selection have been developed (Baldi and Brunak 2001) and made available as part of Bioinfonnatics tools. As an example, here we will take a publicly available gene expression data of 60 samples of CNS cancer (medulloblastoma) representing 39 children patients who survived the cancer after treatment, and 21 - who did not respond to the treatment (Pomeroy et al. 2002). Fig.l.6 illustrates the selection of the top 12 genes out of 7129, as numbered in the original publication (Pomeroy et al. 2002), based on 60 samples, using a signal to noise ratio method (SNR) in a software environment NeuCom (www.theneucom.com). The SNR method used in the analysis from Fig.l.6 ranks all variables according to their discriminative power between the two class samples. The following formula is used to calculate each gene G; its SNR coefficient S;: S;= abs (M;(l) -

M;(2)) /

(Std. (1) + StdP))

(1.1)

where M;(I) and M;(2) are the mean values of the gene G; in class 1 and class 2 samples, where Std?) and StdP) are the corresponding standard deviations. The CNS cancer data set as shown in Fig.l.6 is used further in the book to illustrate different methods that can be used in building CNGM. The selected smaller number of genes, out of thousands, can be further analyzed and modeled in terms of their interaction and relation to the functioning of neurons, neural networks, the brain and the CNS, which is a subject of CNGM.

12

1 Computational Neurogenetic Modeling (CNGM): A BriefIntroduction

-.._-

... _ D.

jOl'l_.... w

.' I

1~L-..

.=.

- .

0'

."

I...·"....."'.....

:.J

I~ro ""'"

~

_n_

"

I

m•

~-

r--- o.."...-.-

Dl!I

I

0'

.,

.-

0. ........ .. . . - ...

J

I

i '"

' w.•

D>

D"

01

~

01.

•

(i,llD)JA

ow

;~1:»1l5..~1~

VM;';';;-

'::<J J

Fig. 1.6. 12 genes selected as top discriminating genes from the Central Nervous System (CNS) cancer data that discriminates two classes - survivals and not responding to treatment (Pomeroy et al. 2002). The NeuCom software system is used for the analysis (www.theneucom.com) and the method is called "Signal-toNoise ratio". See Color Plate 1

1.4 CNGM: How to Integrate Neuronal and Gene Dynamics? A CNGM integrates genetic, proteomic and brain activity data and performs data analysis, modeling, prognosis and knowledge extraction that reveals relationship between brain functions and genetic information. Let us look at this process as a process of building mathematical function or a computational algorithm as follows. A future state of a molecule M' or a group of molecules (e.g. genes, proteins) depends on its current state M, and on an external signal Em: (1.2) A future state N' of a neuron, or an ensemble of neurons, will depend on its current state N and on the state of the molecules M (e.g. genes) and on external signals En: (1.3) And finally, a future cognitive state C' of the brain will depend on its current state C and also on the neuronal- N, and the molecular- M state and on the external stimuli E;

1.4 CNGM: How to Integrate Neuronal and Gene Dynamics?

13

(1.4)

C'=Fc (C,N,M, Ec )

The above set of equations (or algorithms) is a general one and in different cases it can be implemented differently, e.g.: • One gene - one neuron/brain function (see Chaps. 9 and 10). • Multiple genes - one neuronlbrain function, no interaction between genes (see Chaps. 9 and 10). • Multiple genes - multiple neuron/brain functions where genes interact in a gene regulatory network (GRN) and neurons also interact in neural network architecture (see Chap. 8 and also Fig. I. 7). • Thousands of genes - complex brain/cognitive function/s where genes interact within GRN and neurons interact in several hierarchical neural networks (discussed in Chap. 8). GRN

ANN

output

.. - - - - - - - - - ,

Fig. 1.7. A more complex case ofCNGM, where a GRN of many genes is used to represent the interaction of genes, and an ANN is employed to model a brain function. The model output is compared against real brain data for validation of the model and for verifying the derived gene interaction GRN after model optimization is applied (see Chap. 8)

The most common models of brain functions so far are based on the methods of artificial neural networks (ANN) (Rolls and Treves 1998), which is not surprising, as ANN are design using principles of the brain. Other methods, such as differential equations, statistical regression, evolutionary computation (EC), etc. are also used to model the brain. Generally speaking, different methods of mathematical and information sciences can be applied for the development of CNGM. In the next section we briefly discuss these methods, their principles, advantages, limitations and roles in building CNGMs. A short description of the methods is given in Appendix 2 and a more detailed description of the main methods - ANN and evolutionary computation (EC) is given in Chaps. 4, 5 and 6.

14

1 Computational Neurogenetic Modeling (CNGM): A Brieflntroduction

1.5 What Computational Methods to Use for CNGM? In order to implement Eq. 1.2 and Eq. 1.3 and to build a CNGM for classification or prediction and for knowledge discovery from brain-gene data, different mathematical and information science methods can be used. Here we will present a list of methods, most of them falling in the area of computational intelligence (CI) and knowledge engineering (KE), with a brief comment about their applicability to build CNGM. • Experimentally derived analytical functions and statistically derived regression functions (Pevzner 2000) are applicable for instance to derivation of GRN when sufficient experimental data and expert knowledge are available. • Probabilistic learning methods, e.g. Hidden Markov Models HMM (Hunter 1994, Pevzner 2000, Somogyi et al. 2001) are applicable when a priori information is known as prior probabilities and the distribution of the data is also known in advance. • Statistical learning methods, e.g. Support Vector Machines (SVM), Bayesian classifiers (Vapnik 1998, Baldi and Brunak 2001) are applicable when statistically significant amount of data is available and certain probability distribution may be required. • Case-based reasoning (e.g. k-I\TN; transductive reasoning) (Baldi and Brunak 2001) are applicable when a smaller number of variables is used and a relatively small number of samples, that have reliable variable values. The method is adaptable to new data and applicable for personalized modeling (Song and Kasabov 2006). • Decision trees (Hunter 1994) are applicable for classification tasks and are not adaptive to new data. They are good to extract structured domain information. • Rule-based systems (propositional logic dated back to Aristotle) and fuzzy systems (introduced by (Zadeh 1965)) are applicable if domain knowledge on the problem in hand is available even in an imprecise or incomplete form. Usually fuzzy logic methods are combined with ANN methods for the creation of fuzzy-ANN (Kasabov 1996a). • Artificial neural networks (ANN) (Grossberg 1982, Kohonen 1984, Bishop 1995, Kasabov 1996a) in their different models such as self organizing maps (SOM), multilayer perceptrons (MLP), radial basis function neural networks (RBF), are "model -free" and are applicable when some data is available, but there is no knowledge on what analytical function would be appropriate to use (Chap. 4). Some of the ANN models are adaptive to new data, such as the evolving connectionist systems (ECOS) (Kasabov 2001).These methods are described in Chap. 5.

1.6Aboutthe Book

15

• Evolutionary computation (EC) methods, such as genetic algorithms (GA) (Holland 1975, Goldberg 1989) are applicable when neither data nor much knowledge is available on the problem. They are based on the "generate and test" approach using fitness (goodness) criteria to evaluate how good a generated solution is. For applications of the EC methods, see (D'Haeseleer et al. 2000, Fogel and Come 2003). For a more detailed description of the principles of the EC and some of their applications, see Chap. 6. • Hybrid systems, e.g. knowledge-based neural networks; neuro-fuzzy systems; neuro-fuzzy-genetic systems; evolving connectionist systems (Kasabov 1996a, Kasabov and Song 2002) have the advantages of both ANN and rule based systems (see Chap. 5).

1.6 About the Book This book consists of 5 main parts as shown in Fig. 1.8: 1. Background knowledge on brain information processing (Chaps. 2, 3); 2. Background knowledge on gene information processing (Chap. 7); 3. Background knowledge on computational methods and methods of computational intelligence (CI), mainly ANN and EC (Chaps. 4, 5, 6); 4. Methodologies for building CNGM (Chaps. 8, 9); 5. Applications of CNGM for modeling brain functions and diseases and as novel generic techniques of Cl (Chaps. 9, 10). Brain information processing (Chapters 2, 3, Appendices I and [ \ , 3) Gene information processing (Chapter 7 and Appendix 3) Computational modeling techniques (Chapters 4, 5, 6, and Appendix 2)

f----+

/

Methodologies for building CNGM (Chapters 8 and 9)

r-+

Applications of CNGM for modeling brain functions and diseases (Chapters 9 and 10)

Fig. 1.8. The five partsof the book and theirconnections

16

1 Computational Neurogenetic Modeling (CNGM): A Brief Introduction

1.7 Summary This chapter presents the motivations and the rationale behind the CNGM, along with some introduction to the main concepts, expectations and organization of the book. To conclude, we raise several questions that will be addressed in the book. Hopefully the readers will be able to answer them on the completion of the reading: • Is it possible to create a truly adequate CNGM of the whole brain? Would gene-brain maps help in this respect (see http://alleninstitute.org)? • How can a dynamic CNGM be used to trace over time and predict the progression of brain diseases, such as epilepsy and Parkinson disease? • Is it possible to use CNGM to model gene mutation effects? • Is it possible to use CNGM to predict drug effects? • How can CNGM facilitate better understanding of brain functions, such as memory and learning? • Whether and which generic problems of artificial intelligence (AI), such as classification, prediction, feature selection, pattern discovery, adaptation and visualization, etc. can be efficiently solved with the use of a brain-gene inspired CNGM (Bentley 2004)?

2 Organization and Functions of the Brain

This chapter gives an overview of the brain organization and functions performed by different parts of the brain. We will try to answer the following questions: How is the human brain organized at the macroscopic and microscopic levels? Which functions are performed by the brain? How is the organization of the human brain related to its functions? These and many more other questions about the brain are still under investigation of thousands of neuroscientists all over the world. The first Nobel Prize for pioneering discoveries related to the brain microscopic organization was given to the Spanish scientist Santiago Ramon y Cajal (1852-1934) and Italian Camillo Golgi (1843-1926). These scientists are considered to be the founders of neuroscience and modem brain study. German physicist Herman Ludwig Ferdinand von Helmholtz (1821-1894) is the founder of psychophysics, that is a quantitative experimental and theoretical research of relations between mental and brain functions. Up to these days, the division of cerebral cortex based on its microstructure introduced by a neuroanatomist Korbinian Brodmann (1868-1918), is being used. Frenchman Paul Broca (1824-1880) and Russian Alexander Romanovich Luriya (1902-1977) pioneered research on brain localization of cognitive functions based upon cognitive deficits caused by brain lesions. Even nowadays, invaluable information on brain functions comes from the study of patients with mental and neurological deficits caused by injuries of particular brain areas. Currently however, these approaches are refined with noninvasive imaging techniques like functional magnetic resonance (fMRI), positron emission tomography (PET), electroencephalogram (EEG), and others, which provide a rich source of information about the dynamics and organization of the brain. We will begin this chapter with the description of methods of brain study. Next we will introduce the reader into the basics of how human brain works, how different parts specialize and how they are interconnected, how different parts form hierarchically organized neural networks, how they cooperate and lead to behavior. Although the focus will be on human brain, many features are shared with animals especially mammals and the closest evolutionary relatives (great apes) in particular. We will mention what is happening in the human brain during cognitive processes such as perception, learning, memory

18

2 Organization and Functions of the Brain

storage and recall, thinking and language processing. We will conclude with description of relationships between different neural levels of complexity in order to discover conditions that supposedly lead to consciousness that is the capability of reflecting the outer and inner world.

2.1 Methods of Brain Study At present, a number of techniques are available to investigate where in the brain particular cognitive and other kinds of functions are based. In general, these methods are divided as being invasive or noninvasive. In medicine the term invasive relates to a technique in which the body is entered by puncture, incision or other intrusion. Noninvasive means the opposite, the technique that does not intrude into the body. We will provide an overview of both kinds of methods of brain study starting with the invasive techniques. Information about functions of the brain is still being gathered from brain damaged subjects. Deficits in cognitive processing are observed in people who have suffered some kind of brain damage, due to an accident, stroke, tumor, etc. The damaged areas indicate their involvement in those mental processes or brain functions, which became disturbed. The main problem with this method is that observations are made after the event and therefore lack the proper experimental control before the accident. Very similar in nature to this first method are lesion studies. Comparison is made between cognitive performance before and after the deliberate removal or lesion of part of the brain. These types of studies are usually performed on animals. In humans they are performed only for therapeutic reasons like dissecting the corpus callosum connecting the two hemispheres to treat the life-threatening epilepsy or removal of a lifethreatening tumor. The problem with this approach is that lesions may damage other systems which happen to be next to or pass through the target part being damaged and thus the proper involvement of the target part in a given function may be misjudged. Another invasive method of the brain study is the direct stimulation. Researchers perform electrical, magnetic or chemical stimulation of some neural circuit or part of it, and observe the consequences. Electrical stimulation is delivered through microelectrodes inserted into the brain. This type of research is done routinely on animals. It can be done on human subjects during the brain surgery when the skull has to be opened anyway and surgeons have to map the functions of the operated area and its surrounded parts. Electrical stimulation ofthe brain (ESB) can be also used to

2.1 Methods of BrainStudy

19

treat chronic tremors associated with Parkinson disease, chronic pain of patients suffering from back problems and other chronic injuries and illnesses. ESB is administered by passing an electrical current through a microelectrode implanted in the brain. With chemical stimulation, a particular chemical compound is administered into a chosen part of the brain that is supposed either to stimulate or inhibit neurons within it. The least invasive methods of the stimulation methods is magnetic stimulation, called the Transcranial Magnetic Stimulation (TMS). TMS and rTMS (repetitive TMS) are simply the applications of the principle of electromagnetic induction to get electric currents across the insulating tissues of the scalp and skull without the tissue damage. The electric current induced in the surface structure of the brain, the cortex, activates nerve cells in much the same way as if the currents were applied directly to the cortical surface. However, the path of this current is complex to model because the brain is a non-uniform conductor with an irregular shape. With stereotactic, MRIbased control (see below), the precision of targeting TMS can be as good as a few millimeters. However, besides the invasiveness there are other problems with the methods of direct stimulation. Intensity of an artificial stimulation can be stronger or weaker than the level of spontaneous activity in the target circuit. Therefore artificial stimulation can engage more or respectively less of brain circuitry than is normally involved in the studied function. Thus, there are difficulties in determining which brain circuitries have been actually affected by the stimulation and thus which brain structures actually mediate the studied function. Often used methods of brain study in animal research are the single- and multi unit recordings. Microelectrode recordings from individual neurons or from an array of neighboring neurons indicate specific neural networks dedicated to processing of particular stimuli (e.g. bars of a certain orientation, movement in a particular direction, particular objects like faces, and so on). The problem with this method albeit very precise is that it is an invasive method, i.e. requires an invasion into the brain and into the brain cells. Moreover, without post-mortem histology, it is almost impossible to tell where exactly the recordings were actually made from. Classical anatomical methods, by means of which Cajal and other pioneers made their discoveries, are histology and staining. Anatomists still use to dissect dead brains, stain their cells with different dyes (Golgi stain, lucifer yellow, etc), and study them under the microscope. Thus they can reveal the microscopic structure of the brain in terms of cell types and neural connectivity between cells. The biggest disadvantage is that this study can be performed on dead specimen only.

20

2 Organization and Functions of the Brain

The oldest noninvasive method to measure electrical activity of the brain is the electroencephalography (EEG). An EEG is a recording of electrical signals from the brain made by attaching the surface electrodes to the subject's scalp. These electrodes record electric signals naturally produced by the brain, called brainwaves. EEGs allow researchers to follow electrical potentials across the surface of the brain and observe changes over split seconds of time. An EEG can show what state a person is in (e.g., asleep, awake, epileptic seizure, etc.) because the characteristic patterns of brainwaves differ for each of these states. One important use of EEGs has been to show how long it takes the brain to process various stimuli. A major drawback of EEGs, however, is that they cannot show us the structures and anatomy of the brain and tell us which specific regions of the brain do what. In recent years, EEG has undergone technological advances that have increased its ability to read brain activity from the entire head from up to 128 sites simultaneously. The greatest advantage of EEG is that it can record changes in the brain activity almost instantaneously. On the other hand, the spatial resolution is poor, and thus should be combined with CT or MRI (sec below). Related method to EEG, called magnetoencephalography (MEG) measures millisecond-long changes in magnetic fields created by the brain's electrical currents. MEG is a rare, complex and expensive neuroimaging technique. MEG machine uses a non-invasive, whole-head, 248-channel, super-conducting-quantum-interference-device (SQUID) to measure small magnetic signals reflecting changes in the electrical signals in the human brain. The incorporation of liquid helium creates the incredibly-cold conditions (4.2 degrees of Kelvin) necessary for the MEG's SQUIDS to be able to measure these brain magnetic fields that are billions of times weaker than the earth's magnetic force. Investigators use MEG to measure magnetic changes in the active, functioning brain in the speed of milliseconds. Besides its precision another advantage of MEG is that the biosignals it measures are not distorted by the body as in EEG. Used in conjunction with MRI or fMRI (see below), to relate the MEG sources to brain anatomical structures, researchers can localize brain activity and measure it in the same temporal dimension as the functioning brain itself. This allows investigators to measure, in real-time, the integration and activity of neuronal populations while either working on a task, or at rest. The brains of healthy subjects and those suffering from dysfunction or disease are imaged and analyzed . The oldest among the noninvasive methods to study brain anatomy is Comput er Tomography (CT). It is based on the classical X-ray principle. X-rays reflect the relative density of the tissue through which they pass. If a narrow X-ray beam is passed through the same point at many different

2.1 Methods of BrainStudy

21

angles, it is possible to construct a cross-sectional visual image of the brain. A 3D X-ray technique is called the CAT (Computerized Axial Tomography). CT is noninvasive and shows only the anatomical structure of the brain, not its function. Positron Emission Tomography (PET) is used for studying the living brain activity. This noninvasive method involves an on-site use of a machine called cyclotron to label specific drugs or analogues of natural body compounds (such as glucose or oxygen) with small amounts of radioactivity. The labeled compound (a radiotracer) is then injected into the bloodstream which carries it into the brain. Radiotracers break down, giving off sub-atomic particles (positrons). By surrounding the subject's head with a detector array, it is possible to build up images of the brain showing different levels of radioactivity, and therefore, cortical activity. Thus, depending on whether we used glucose (oxygen) or some drug, PET can provide images of ongoing cortical or biochemical activity, respectively. Among the problems with this method are expense including the on-site cyclotron and also technical parameters like the lack of temporal (40 seconds) and spatial (4 mm - 1 em) resolution. Usually the PET scan is combined either with CT or MRI to correlate the activity with brain anatomy. Single-Photon Emission Computed Tomography (SPECT) uses gamma radioactive rays. Similar to PET, this noninvasive procedure also uses radiotracers and a scanner to record different levels of radioactivity over the brain. SPECT imaging is performed by using a gamma camera to acquire multiple images (also called projections) from multiple angles. A computer can then be used to apply a tomographic reconstruction algorithm to the multiple projections, yielding a 3D dataset (like in CT). Special SPECT tracers have long decay time, thus no on-site cyclotron is needed, which makes this method much less expensive than PET. However, the temporal and spatial resolution of brain activity is even smaller than in PET. Magnetic Resonance Imaging (MRI) uses the properties of magnetism instead of injecting the radioactive tracers into the bloodstream to reveal the anatomical structure of the brain. A large (and loud) cylindrical magnet creates a magnetic field around the subject's head. Detectors measure local magnetic fields caused by alignment of atoms in the brain with the externally applied magnetic field. The degree of alignment depends upon the structural properties of the scanned tissue. MRI provides a precise anatomical image of both surface and deep brain structures, and thus can be combined with PET. MRI images provide greater detail than CT images. Problems: Expense, cannot be used in patients with metallic devices, patient must hold still for 40-90 min.

22

2 Organization and Functions of the Brain

Functional MRJ (fMRI) combines visualization of brain anatomy with the dynamic image of brain activity into one comprehensive scan. This noninvasive technique measures the ratio of oxygenated to deoxygenated hemoglobin which has different magnetic properties. Active brain areas have higher levels of oxygenated hemoglobin than less active areas. An fMRI can produce images of brain activity as fast as every 1-2 seconds, with very precise spatial resolution of about 1-2 mm. Thus, fMRI provides both an anatomical and functional view of the brain and is very precise. FMRI is a technique for determining which parts of the brain are activated by different types of brain activity, such as sight, speech, imagery, memory processes, etc. This brain mapping is achieved by setting up an advanced MRI scanner in a special way so that the increased blood flow to the activated areas of the brain shows up on fMRI scans. The subject in a typical experiment lies in the magnet and a particular form of stimulation is set up (auditory, visual, etc). For example, the subject may wear special glasses so that pictures can be shown during the experiment. Then, MRI images of the subject's brain are taken. Firstly, a high resolution single scan is taken. This is used later as a background for highlighting the brain areas which were activated by the stimulus. Next, a series of low resolution scans are taken over time, for example, 150 scans, one every few seconds. For some of these scans, the stimulus (sound, picture) will be presented, and for some of the scans, the stimulus will be absent. The low resolution brain images in the two cases can be compared, to see which parts of the brain were activated by the stimulus. After the experiment has finished, the set of images is analyzed. Firstly, the raw input images from the MRI scanner require mathematical transformation to reconstruct the images into space, so that the images look like brains. The rest of the analysis is done using a series of tools which correct for distortions in the images, remove the effect of the subject moving their head during the experiment, and compare the low resolution images taken when the stimulus was off with those taken when it was on. The final statistical image shows up bright in those parts of the brain which were activated by this experiment. These activated areas are then shown as colored blobs on top of the original high resolution MRI scan, for interpretation of the experiment. This combined activation image can be rendered in 3D, and the rendering can be calculated from any angle. By means of fMRI a very comprehensive picture of brain in action can be derived. Comparisons between healthy and ill brains can be made and correlated with structural changes if these are present. There is a serious interpretation problem with all the methods that measure cortical activity (PET, fMRI, EEG, MEG). How does an experimenter decide which cortical activity is specifically related to the psychological process in question? It is done by the so-called subtraction method.

2.2 Overall Organization of the Brain and Motor Control

23

The experimenter calculates the difference image between that of the process and that of a control situation. The difference images from individual subjects are averaged to produce a group mean difference image. It can be quite problematic to determine which situation should represent the control background. It is also problematic to align individual or average functional images with anatomical structures, when comparing the coordinates of activation to a standard atlas which by no means has been proven to reflect the borders of particular areas for all people. In fact, the converse is true; most Brodmann areas differ between individuals. Nevertheless, comparing images taken during some cognitive processing to those taken before or after it, scientists are gaining many new insights about the brain structure and function.

2.2 Overall Organization of the Brain and Motor Control It is estimated that there are io'' - 1012 of neurons in the human brain (Kandel et al. 2000). Three quarters of neurons form a 4-6 mm thick cerebral cortex that constitutes a heavily folded brain surface. Cerebral cortex is thought to be a seat of cognitive functions, like perception, imagery, memory, learning, thinking, etc. The cortex cooperates with evolutionary older subcortical nuclei that are located in the middle of the brain, in and around the so-called brain stem (Fig. 2.1). Subcortical structures and nuclei are comprised for instance of basal ganglia, thalamus, hypothalamus, limbic system and dozens of other groups of neurons with more or less specific functions in operations of the whole brain. For example, the input from all sensory organs comes to the cortex preprocessed in thalamus. Emotions and memory functions depend upon an intact limbic system. When one of its crucial parts, hippocampus, is damaged, humans (and animals) loose their ability to store new events and form new memories. When a particular cortical area has been damaged, a particular cognitive deficit follows. However, all the brain parts, either cortical or subcortical, are directly or indirectly heavily interconnected, thus forming a huge recurrent neural network (in the terminology of artificial neural networks). Thus, we cannot speak of totally isolated neuroanatomic modules. Fig. 2.1 shows a schematic functional division of the human cerebral cortex. One third of the cortex is devoted to processing of visual information in the primary visual cortex and higher-order visual areas in the parietal cortex and in the inferotemporal cortex. Association cortices take about one half of the whole cortical surface. In the parietal-temporal-occipital as-

24

2 Organization and Functions of the Brain

sociation cortex, sensory and language information are being associated. Memory and emotional information are associated in the limbic association cortex (internal and bottom portion of hemispheres). The prefrontal association cortex takes care of all associations, evaluation, planning ahead and attention. Language processing takes place within the temporal cortex, parietal-temporal-occipital association cortex, and frontal cortex.

limbic association cortaX

Fig. 2.1. Gross anatomical and functional division of the human cerebral cortex. The same division applies for the right hemisphere. Dashed curves mark the position of evolutionary older subcortical nuclei in and around the brainstem of the brain. Each of the depicted areas has far more subdivisions

At the border between the frontal and parietal lobes, there is a somatic sensory cortex, which processes touch and other somatosensory signals (temperature, pain, etc.) from the body surface and interior. In the front of it, there is a primary motor cortex, which issues signals for voluntary muscle movements including speech. These signals are preceded by the preparation and anticipation of movements that takes place in the premotor cortex. The plan of actions and their consequences, inclusion and exclusion of motor actions into and from the overall goal of an organism, are performed within the prefrontal association cortex. Subcortical basal ganglia participate in preparation and tuning of motor outputs, in the sense of initiation and the extent of movements. Cerebellum executes routine automatic movements like walking, biking, driving, etc. We want to point out that

2.3 Learning and Memory

25

there are far more anatomical and functional subdivisions within each of the mentioned areas. Functions, or better, dominances of the right and left hemispheres in different cognitive functions are different (Kandel et al. 2000). It was shown by Roger Sperry and Michael Gazzaniga in the studies of the so-called "split-brain" patients to whom connections between the two hemispheres were cut because of the therapeutic reasons. The dominant hemisphere (usually the left one) is specialized for language, logical reasoning, awareness of cognitive processes and awareness of the results of cognitive processes. Although the non-dominant hemisphere (usually the right one) is able to carry out cognitive tasks, it is not aware of them or their results. It is specialized for emotional and holistic processing, intra- and extrapersonal representation of space. Its intactness is crucial for the awareness of the body integrity (Damasio 1994). Lesion of the parietal cortex including the somatosensory cortex leads to the so-called anosognosia. The limbs and the body are intact but the cortical and mental representations become missing. Patients who have undergone a stroke to the right parietal lobe, neglect the left half of their body, in spite they can see it. It is not a consequence of the left hemiparalysis. Mirror damage to the left parietal lobe does not lead to anosognosia. It seems that the right hemisphere is dominant in mental representations of intra- and extrapersonal space. In other words, subjective experience of the body self depends upon specific brain mechanisms, namely an integrity of primary and higher-order somatosensory cortical areas in the right hemisphere (Damasio 1994). Although, right part of the body is represented in the right somatosensory cortex and the left half of the body in the right hemisphere, the latter seems to have a special role in the integral self-awareness.

2.3 Learning and Memory Capability of learning and memory formation is one of the most important cognitive functions. Our identity largely depends upon what we have learned and what we can remember. We can divide the study of learning and memory into two levels: 1. The system level (where?) that attempts to answer the question which brain parts and pathways the memory trace is stored in - the top-down approach, which will be the topic of this section, and 2. Molecular level (how?), which is devoted to investigation of the ways of coding and storage of information at the cellular and molecular level the bottom-up approach, which will be introduced in the next chapter.

26

2 Organization and Functions of the Brain

It is interesting that learning and memory is independent from other cognitive functions and as such can be studied independently. It was

shown for the first time at a famous patient from 1953, which underwent a bilateral removal of the middle portion of his temporal lobe including hippocampus to treat severe epilepsy. He has lost the ability to store new memories (the so-called anterograde amnesia). He could not keep a memory of new people, objects, facts, or places no longer than few minutes. Moreover, he lost all memories from about two years before the operation (the so-called retrograde amnesia), the period which scientists consider to be the consolidation period during which memories enter the long-term memory. However, all these heavy cognitive deficits did not affect other mental functions like language processing and thinking, including IQ and implicit memory (see below).

Facts

-j-

Conditioning

Events

Priming

!

Neocortex

hippocampus and neocortex

Habituation and sensitisation

Skills

j Reflex pathways

Emotional

Classical

j

j

Amygdala

Cerebellum

Basal ganglia

Fig. 2.2. Different kinds of long-term memory fall under two general categories: explicit and implicit

It has been long recognized that there is a short-term memory and a long-term memory. Short-term memory lasts for a few minutes and is also called the working memory. It occurs in the prefrontal cortex, although

2.3 Learning and Memory

27

other parts of the cortex relevant to the memory content are activated too (Roberts et al. 1998). The learning process and the process of long-term memory formation can be divided into these four stages: 1. Encoding. Attention focus and entering of new information into the working memory. Finding associations with already stored memories. 2. Consolidation. The process of stabilization of new information, transformation into a long-term memory by means of rehearsal. 3. Storage. Long-term storing of information in memory. 4. Recall. Retrieval of information into the working memory. Based on clinical, imaging and animal studies we can divide long-term memory into two main categories that have different subtypes with different mechanisms and different localizations in the brain (Fig. 2.1). Explicit (declarative) memory is a memory of facts (semantic memory) and a memory of events (episodic memory). Recall from explicit memory requires conscious effort and stored items can be expressed in language. Hippocampus is a crucial but only a transitory stage in the explicit memory. How is the explicit memory formed? Information comes to brain through the sensory organs (visual, auditory, olfactory, tactile), and proceeds through subcortical sensory nuclei and sensory cortical areas into multimodal association areas, like for instance the parieto-temporooccipital association cortex, limbic association cortex and the prefrontal association cortex. From there the information is relayed through parahippocampal cortex, perirhinal cortex and entorhinal cortex into the hippocampus. From hippocampus the information is relayed to subiculum from where it returns back to entorhinal cortex and all the way back to association cortical areas. Thus the brain circuit for the long-term explicit memory storage forms a re-entrant closed loop. According to experimental data, the "synaptic re-entry reinforcement" or SRR hypothesis and the corresponding computational model have been formulated and simulated (Wittenberg et al. 2002, Wittenberg and Tsien 2002). According to this hypothesis, after initial learning, reactivation of hippocampal memory traces repeatedly drives cortical learning. Thus, a memory trace (engram) is stored after many repetitions. Repeated reinforcement of synapses during the reactivation of memory traces could lead to a situation in which memory traces compete, such that the strengthening of one memory is always at the expense of others, which are either weakened or lost entirely. In other words, a single memory stored in a neural network is either lost (owing to synaptic decay) or strengthened and maintained by repeated rounds of synaptic potentiation each time the memory is reactivated. Once cortical connections are fully consolidated and stabilized, the hippocampus itself becomes dispensable. Differences in the frequency with which memory traces are

28

2 Organization and Functions of the Brain

either consciously or subconsciously recalled could be another factor affecting the selection of which memories are consolidated. An increasing amount of evidence suggests a role of sleep in memory consolidation by means of learning-induced correlations in the spontaneous activity of neurons and replaying the patterns of wake neural activities during sleep (Maquet 2001, Stickgold et al. 2001). Although others point out that people lacking REM sleep do not show memory deficits and that a major role of sleep in memory consolidation is unproven (Siegel 2001). An interesting question is how the degradation of out-dated hippocampal memory traces occurs after memory consolidation is finished. The most recent hypothesis is that memory clearance may actually involve newborn neurons. Neurogenesis in the dentate gyrus of the hippocampus persists throughout life in many vertebrates, including humans. The progenitors of these new neurons reside in the subgranular layer of the dentate gyrus (Seri et al. 2001). Deletion of the Presenilin-l gene in excitatory neurons of the adult mouse forebrain led to a pronounced deficiency in enrichment-induced neurogenesis in the dentate gyrus (Feng et al. 2001). This reduction in neurogenesis did not result in appreciable learning deficits, indicating that addition of new neurons is not required for memory formation. However, the postlearning enrichment experiments led to a postulate that adult dentate neurogenesis may playa role in the periodic clearance of outdated hippocampal memory traces. The clearance can happen because these adult-born neurons are short-lived, with a life span of several weeks in rodents. The implicit or nondeclarative memory serves to store the perceptual and motor skills and conditioned reactions. Recall of stored implicit information occurs without a conscious effort, automatically and the information is not expressed verbally. Basal ganglia and cerebellum are important for acquisition of motor habits and skills that are characterized by precise patterns of movements and fast automatic reactions. Cerebellum is the key structure for conditioning. Conditioned emotional reactions require amygdala in the limbic system. Nonassociative learning like habituation and sensitization occur in primary sensory and reflex pathways. Priming is an increase in the speed or accuracy of a decision that occurs as a consequence of a prior exposure to some of the information in the decision context, without any intention or task related motivation, and occurs in neocortex. Although implicit and explicit learning concern different memory contents, they share cellular and molecular mechanisms (Bailey et al. 2004). These mechanisms will be one of the topics of the next chapter. Later we also introduce the genetics of learning and memory and the neurogenetic computational model.

2.4 Language and OtherCognitive Functions

29

2.4 Language and Other Cognitive Functions Language is distinctive from other forms of communications by its form, content, use and creativity (Mayeux and Kandel 1991): • Form. Language is formed from a limited set of "nonsense" elementary sounds (i.e. phonemes) that are arranged into various predictable sequences that signal content. • Content. Language provides means of communicating contents whose meanings are independent of the immediate situation. • Use. Through language we organize our sensory experience and express our thoughts, feelings, and expectations. • Creativity. With every new thought we speak we create original sentences. We readily interpret original sentences spoken by others.

2.4.1 Innate or Learned? Although the acquisition of language undoubtedly involves learning, developmental and neurobiological studies indicate a large innate part to it. First, both natural and sign language functions are dominantly localized to the left hemisphere, as revealed by lesion studies and brain imaging. Second, there are universal regularities in the acquisition of language across all human cultures (see Table 2.1). Children learn the words and rules of language effortlessly by simply listening to the speech around them. There exist about 6000 live languages and all of them have principal characteristics in common. Currently, there exist strong evidence from comparative linguistics and genetic anthropology that language has appeared just once, in Africa, and all other languages descend from it (Cavalli-Sforza 2001). Noam Chomsky assumes that humans have some innate program (generative grammar) that prepares them to learn language in general. In more modem terms, these innate neural mechanisms determine possible characteristics of acquired language, and the process of its acquisition. In addition, there is a critical period during postnatal development for learning language. Investigation of feral children has showed that it is impossible to acquire a fully developed language (especially syntax-vice and grammar-vice) when this critical period is over, which is at about 8 years. In summary, psychologists and linguists now believe that the mechanisms for the universal aspects of language acquisition are determined by the structures in the human brain. Thus, the human brain is innately prepared to learn and use language. The particular language spoken and the

30

2 Organizationand Functions of the Brain

dialect and accent are determined by the social environment. The questions now being debated are which language characteristics derive from neural structures specifically related to language acquisition and which from cognitive characteristics that are more general. Table 2.1. Stages of development in the acquisition oflanguage in humans Average Age

Language ability

6 months 1 year 1 VI years

Beginning of distinct babbling. Beginning of language understanding, one-word utterances. Words used singly, child uses 30-50 words (simple nouns, adjectives, action words) one at a time but cannot link them to make phrases, does not use functors (the, and, can, be) necessary for syntax. Two-word (telegraphic) speaker, 50 to several hundred words in the vocabulary, much use of two-word phrases that are ordered according to syntactic rules, child understands propositional rules. Three or more words in many combinations, functors begin to appear, many grammatical errors and idiosyncratic expressions, good understanding of language. Full sentences, few errors, vocabulary of around 1000 words. Close to adult speech competence.

2 years

2 VI years

3 years 4 years

2.4.2 Neural Basis of Language The basic model of language processing during the simple task of repeating the word that has been heard is the Wernicke-Geschwind model (Mayeux and Kandel 1991) (Fig. 2.3). According to this model, the language task involves transfer of information from the inner ear through the auditory nucleus in thalamus to the primary auditory cortex (Brodmann's area 41), then to the higher-order auditory cortex (area 42), before it is relayed to the angular gyrus (area 39). Angular gyrus is a specific region of the parietal-temporal-occipital association cortex, which is thought to be concerned with the association of incoming auditory, visual and tactile information. From here, the information is projected to Wernicke's area (area 22) and then, by means ofthe arcuate fasciculus, to Broca's area (44, 45), where the perception of language is translated into the grammatical structure of a phrase and where the memory for word articulation is stored. This information about the sound pattern of the phrase is then relayed to the facial area of the motor cortex that controls articulation so that the word can be spoken. It turned out that a similar pathway is involved in naming an object that has been visually recognized. This time, the input

2.4 Language and Other Cognitive Functions

31

proceeds form retina and LGN (lateral geniculate nucleus) to the primary visual cortex, then to area 18, before it arrives to the angular gyrus, from where it is relayed by a particular component of arcuate fasciculus directly to Broca's area, bypassing Wernicke's area. Lesions in different parts of the cerebral cortex cause selective language disturbances , called aphasias, instead of an overall reduction in language ability. Normal language is dependent not only on cortical but also on subcortical structures and connections. Lesions that do not affect the cerebral cortex, typically vascular lesions in the basal ganglia and/or thalamus, can also result in aphasia. Basal ganglia take part in motor output and thalamus in perception. Furthermore , the damage to the brain language areas often affects also other cognitive and intellectual skills to some degree.

Occipital cortex

Eye

Temporal cortex

Fig. 2.3. Lateral view of the exterior surface of the left hemisphere with the main language processing areas. Broca's area (Brodmanu's areas 44/45) is adjacent to the regions of premotor (6) and motor (4) cortices that control the movements of facial expressions, articulation and phonation. Wernicke's area (area 22) lies in the posterior superior temporal lobe near the primary and higher-order auditory cortices in the superior temporal lobe (areas 41/42). Wernicke's and Broca's areas are joined by a fiber tract called the arcuate fasciculus

Wernicke's aphasia is characterized by a prominent deficit in language comprehension. The lesion primarily affects area 22 (Wernicke's area), and often extends to the superior portions of the temporal lobe (areas 39/40), and inferiorly to area 37. Comprehension of both auditory and visuallanguage inputs is severely impaired, accompanied with severe reading

32

2 Organization and Functions of the Brain

and writing disabilities. Speech is fluent, grammatical, but lacks meaning. Phenomena like empty speech, neologisms, and logorrhea occur. Patients are generally unaware of these speech failures, probably because of the lack of their own language comprehension. Occasionally, a right visual field defect is encountered. Conduction aphasia is the result of damage to the arcuate fasciculus. Symptoms resemble those of the Wernicke's aphasia. Many patients with conduction aphasia have some degree of impairment of voluntary movement.

j

Speech

Writing

Visual processing Visual cortex (areas 17, 18, 19)

Auditory processing Auditory cortex (areas 41/42)

j Language comprehension Temporo-parietal areas (areas 39, 22, 37, 40) Semantic association Left anterior inferior frontal cortex Motor control Left prem otor cortex (areas 44, 45, 6)

Speech

1

Writing

Fig. 2.4. Recent model of the neural processing of language build upon the Wernicke-Geschwind original model. Simplified scheme shows the relationships between various anatomical structures and functional components of language. Connections are actuallyreciprocal

Broca's aphasia is characterized by a prominent deficit in language production. Lesions affect areas 44 and 45 (Broca's area), and in severe cases also other prefrontal regions (8, 9, 10, 46) and premotor regions (area 6). The most severe case is the complete muteness. Usually, speech contains

2.4 Language and OtherCognitive Functions

33

only key words, nouns are expressed in singular, verbs in the infinitive. Articles, adjectives, adverbs and grammar are missing altogether. Unlike Wernicke's aphasia, patients with Broca's aphasia are generally aware of these errors. Reading and writing are also impaired, because they include also motor components. Some defects in comprehension related to syntax may be encountered. Right hemiparesis and loss of vision is almost always present in this type of aphasia. Lesions to the prefrontal cortical regions other than Broca's area or to the parietal-temporal cortical regions other than Wernicke's area can result in various language deficits in production or comprehension, respectively. When certain portions of higher-order visual areas are damaged, specific disorders of reading and/or writing follow (dyslexias, alexias and agraphias). Homologous language areas in the right hemisphere process affective components of language like musical intonation (prosody) and emotional gesturing. Disturbances in affective components of language associated with damage to the right hemisphere are called aprosodias. The organization of prosody in the right hemisphere seems to mirror the anatomical organization of the cognitive aspects of language in the left hemisphere. Thus, patients with posterior lesions do not comprehend the affective content of other people's language. On the other hand, lesion to the anterior portion of the right hemisphere leads to a flat tone of voice whether one is happy or sad. To sum up, recent cognitive and imaging studies have revealed that language processing involves a larger number of areas and a more complex set of interconnections than just a serial interconnection of Wernicke's area to Broca's area. Thus, a more realistic scheme illustrating the neural processing of language is shown in Fig. 2.4. 2.4.3 Evolution of Language, Thinking and the Language Gene

In most individuals the left hemisphere is dominant for language and the cortical speech area of the temporal lobe (the planum temporale) is larger in the left than in the right hemisphere. Since important gyri and sulci often leave impression upon the skull, it is possible to examine human fossils in search for such impressions. Marjorie LeMay searched the fossil skulls for the morphological asymmetries associated with speech and has found them besides in the modem Homo sapiens also in the Neanderthal man (Homo sapiens neanderthalensis, dating back 30,000 to 50,000 years) and in Peking man (Homo erectus pekinensis, dating back 300,000 to 500,000 years). The left hemisphere is also dominant for the recognition of species-

34

2 Organization and Functions of the Brain

species cries in Japanese macaque monkeys, and asymmetries similar to those of humans are present in brains of modem-day great apes. G. Rizzolatti et al. (DiPellegrino et aI. 1992, Rizzolatti et aI. 1996) have found out that neurons in the ventral premotor cortex of macaque monkeys are active not only when the monkey executes motor actions, but also when she watches others, either monkeys or humans, to perform the same actions. Thus, these mirror neurons follow or imitate what others are doing. They may form a neural basis for learning by imitation, which is very important for language acquisition. In humans, the ventral premotor area includes Broca's area (areas 44/45), which is a specific cortical area associated with expressive and syntactical aspects of language. Thus maybe, evolution of the ventral premotor area with its mirror neurons played an important role in evolution of neural basis for language (Rizzolatti and Arbib 1998). It is also intriguing to see areas responsible for contemplation of (motor) actions and areas processing language being at the same place in the brain . Thus thinking can be hypothesized to be a contemplation of actions in the real or abstract spaces. According to Rizzolatti and Arbib (Rizzolatti and Arbib 1998) and Corballis (Corballis 2003) speech and language have evolved from the communication gestures and not from the vocal communication of primates. Vocal production of primates is controlled in the emotional cortical and subcortical centers and serves mainly to communicate emotional state (anger, fear, content, etc.). On the other hand, communication modality of gestures is visual and motor and involves the premotor cortex. Our ancestors probably started to use the gestures and mirror neurons to communicate nonemotional contents. Brachial and manual gestures were probably accompanied with the oro-facial movements and differentiating sounds. Communication gestures started to be associated with these accompanying sounds on a regular basis, which has led to the liberation of these sounds in language. Although the anatomical structures that are prerequisites for language may have arisen early, many linguists believe that language per se emerged rather late in the prehistoric period of human evolution (about 100,000 years ago). There exist strong evidence from comparative linguistics and genetic anthropology that language arose only once, in Africa, and all other languages descend from a single original language (CavalliSforza 2001). Modem Homo sapiens, which evolved in Africa, began to leave Africa some 100,000-80,000 years ago, while taking this first language with them. This theory fits well with the estimated time of fixation of the so-called language gene FOX?2 in modem humans (Enard et aI. 2002). FOX?2 (fork-head box P2) is located on human chromosome 7q31. It codes a protein containing 715 amino-acid bases, which belongs to the class of the so-called fork-head transcription factors that control transcrip-

2.4 Language andOther Cognitive Functions

35

tion of DNA. So far it is not known which transcription is controlled by FOXP2. This gene is present in all mammals however humans gained two mutations compared to chimpanzees, gorillas and macaques, which have identical proteins FOXP2. It seems that two functional copies of the human gene FOXP2 must be present to acquire a full language. People with the point mutation of one of the gene copies have serious problems with articulation, grammar, expression and comprehension of language. One of the consequences of this mutation is the disorder of sequencing of subtle oro-facial movements. This motor disorder is accompanied with serious mental problems with sequencing the syllables into words and words into grammatically correct sentences. Over the years much evidence has accumulated to support the idea that aspects of our genetic makeup are critical for acquisition of language (Marcus and Fisher 2003). To gain insights into the evolution of language and also into its neural basis, studies of communication systems and language abilities of great apes like chimpanzees and bonobo have been very helpful (SavageRumbaugh and Lewin 1994). In short summary, bonobos' capacity for processing grammar even after many years of proper training remains limited. With respect to the language production, they are at the level of a two-year child (see Table 2.1). With respect to the language understanding, they can be at the level of 2.5-3-year child. Thus, so far it seems that fully developed language is an exclusive form of human communication. An important topic in the study of language is the relation of language and its evolution to other cognitive functions and their evolution, respectively (Gardenfors 2000, Marcus 2004b). Language is to communicate about something that is not here and not now. Thus, a more general cognitive abilities such as being able to create detached representations and being able to make anticipatory planning (planning about future needs, goals, events, etc.) can be necessary (but not sufficient) cognitive prerequisites for language. Grammar is an enhanced formal mean for organization of language. Except the evolution of general cognitive abilities, the evolution of language may go hand in hand with the development of advanced forms of cooperation. Without the aid of symbolic communication about detached contents, we would not be able to share visions about the future. We need language in order to convince each other that a future goal is worth striving for (Gardenfors 2000).

36

2 Organization and Functions of the Brain

2.5 Neural Representation of Information The first principle of representation of information in the brain is redundancy. Redundancy means that every information (meant in any sense) is stored, transmitted and processed by a redundant number of neurons and synapses so that it does not become lost when neural networks undergo damage, for instance due to aging. When neural networks get damaged, their performance does not drop down to zero abruptly, like in a computer, but instead it degrades gracefully. Computer models of neural networks also confirm the idea that a degradation of performance with the loss of neurons and synapses is not linear but instead neural networks can withstand quite substantial damage, and still perform well. Next, the contemporary view on the nature of neural representation is such that information (in the sense of content or meaning) is represented by place in the cortex (or in general in the brain). However, this placing is a result of anatomical framework and shaping by input, i.e. by experience-dependent plasticity. For instance, a sound pattern for the word "apple" is represented in the auditory areas of the temporal cortex. It is represented as a spatial pattern of active versus inactive neurons. This neural representation is associated (connected) through synaptic weights with the neural representation of a visual image of apple in the parietal cortex, with the neural representation of an apple odor in the olfactory cortex, with memories on the grandma garden and facts about apples, being represented in some other areas of the cortex, etc. Neural representations (that is distributions or patterns of active neurons) within particular areas and their associations between areas appear as a result of learning (i.e. synaptic plasticity). Different objects are represented by means of different patterns or distributions of active neurons within cortical areas. Therefore we speak about the so-called distributed representations. Current hypothesis states that recall from memory is an active process. Instead of passive processing of all electrical signals that arrive from hierarchically lower processing levels, cortical neural networks should be able to use fragments of activity patterns to fill in the gaps, and thus quickly recreate the whole neural representation. The filling-in process can be nicely modeled by means of model neural networks (Fig. 2. 5). Neural representations (patterns of activities) are stored in the matrix of synaptic weights through which neurons in the network are interconnected. The weight distribution storing a particular object representation is created due to an experience-dependent synaptic plasticity (learning). When a sufficiently large portion of this neural representation is activated from outside the network,

2.6 Perception

37

few electric signals along all the synapses in the network quickly switch on the correct remaining neurons in the representation. Neural representations in the sense of patterns of activity have a holistic character. Patterns of activity are being recalled (restored) as a whole. Thus, we can see a nice relation between the character of neural representations and gestalts. Gestalt psychology was developed at the beginning of the 20th century by Max Wertheimer, Kurt Koffka and Wolfgang Kohle in Germany. Gestalt psychology considers holistic mental gestalts (shapes, forms) to be the basic mental elements. For the gestalt to be stored and recalled, certain rules must be fulfilled, like the rules of proximity, good continuation, symmetry, etc. These rules have been experimentally verified.

.. • ....••

lIiiii••

• :=iii=:::

-·-·:-1

11111111· .Iiiiil•

•;:iiii:: .

iiiiiiii :. Fig. 2.5. Illustration of spontaneous re-creation of neural representation after few input impulses(figure in the uppermost left corner). Black pixel represents a firing neuron while blank pixel represents a silent neuron. Between each pattern of activity from left to right (I ms time frame), neurons in the network exchange only one impulse. Thus, basically after exchanging only two-three spikes, the memory pattern is re-created. Network remains in the restored memory pattern until a different external input arrives

To conclude, neural representations of objects are stored in the matrix of synaptic weights as a whole. We are not able to trace down a sequence of steps leading to the holistic percept. Synaptic weights implicitly bind together parts of the pattern.

2.6 Perception Perception is accompanied by sensory awareness, and therefore we will describe the underlying neural processes in relation to the next section on consciousness. We will concentrate on visual perception and visual awareness since similar principles apply to all sensations. Neurons in different areas of the visual cortex respond to various elementary features, like ori-

38

2 Organization and Functions of the Brain

ented edges of light intensity (bars), binocular disparity, movement, color, etc. (Kandel et al. 2000). Visual areas in the occipital, parietal and inferior temporal cortex, though reciprocally connected, are hierarchically organized. Results of processing at lower hierarchical levels are relayed to higher-order areas. Neurons in higher-order areas respond to various combinations of elementary features from lower-order areas. In primates, based on matching psychophysical and physiological data, three main visual systems, relatively independent but mutually heavily interconnected, have been identified: the "magno", "parvo" and the color system (Livingstone and Hubel 1988). The "magno" system is responsible for perception of movement, depth and space, and separation of objects. Several cues leading to the depth perception have been identified: stereopsy, depth from perspective, depth from mutual movement and occlusion, etc. The "parvo" system is responsible for shape recognition. For separation and recognition of objects, we use separation based on movement, separation from background, filling in of borders, shape from shading, etc. The color system is responsible for color perception. With respect to cortical neurons belonging to these three systems, they possess different combinations and ranges of these four physiological properties: sensitivity to color (small/ large), sensitivity to the light contrast (small/large), temporal resolution (small/large), spatial resolution (small/large). These are the so-called elementary features of visual objects. Elementary features belonging to one visual object activate different and spatially separated groups of neurons within the cerebral cortex. Scientists at the Max Planck Institute in Germany under the leadership of Wolf Singer analyzed trains of spikes of neurons within the visual cortex. They have proposed an intriguing hypothesis about the neural correlate of perception and sensory awareness (Gray et al. 1989, Singer 1994, Roelfsema et al. 1997). Binding of spatially separated neurons coding for features belonging to one visual object could be performed by transient synchronization of firing of these neurons (Fig. 2.6). Similar synchronous oscillations of neurons were detected also in auditory, somatosensory, parietal, motor, and prefrontal cortices in the case of auditory, tactile and other perceptions, respectively (Traub et al. 1996). Oscillations of neurons with frequencies around and above 40 Hz (long known as gamma oscillations) have been detected in the cerebral cortex of humans, primates and other investigated mammals, in particular as a result of sensory stimulation. This synchronization occurs over relatively long distances (mm to em), between different cortical areas, between cortex and thalamus, between the two hemispheres. Synchronization means that neurons discharge with the same frequency and the same phase (Fig. 2.6). This results in a distributed pattern of simul-

2.6 Perception

39

taneously firing neurons. Neural correlates of different objects can differ in (a) which neurons are members of the pattern, (b) which is the particular frequency of their synchronization, and (c) which is the phase of their synchronization. Thus, transient synchronous gamma oscillations have been suggested as a possible candidate for the mechanism of binding many elementary features belonging to one object to one transient whole corresponding to a percept. Establishment of transient synchrony is based upon the underlying synaptic connectivity. In the laboratory ofW. Singer, an interesting experiment was performed to demonstrate that when these synchronizations are disturbed, perception is also disturbed (Konig et al. 1996). In normal rearing, kittens develop normal sharp vision. Neurons in their primary visual cortex are sharply orientationally selective. Distribution of ocular dominance favors binocular neurons; however there are still also monocular neurons present. Firing of neurons responding to the left and right eye is synchronized when they are visually stimulated with the same object. After birth, the eyesight of kittens was disturbed in such a way that surgically they were made strabismic to one eye. They developed a syndrome typical for a strong uncorrected strabismus. Objects are fixated only with the good eye, whereas the strabismic eye is connected with a perceptual deficit called strabismic amblyopia (something like a blurred vision). Analysis of spike trains in response to visual stimuli in the primary cortex of these cats have revealed several facts: (i) neurons that respond only to the stimulation of the strabismic eye have normal physiological characteristics in terms of orientation selectivity, (ii) however synchronization of discharges of these neurons is very weak compared to synchronization of discharges of neurons responding to the stimulation of a good eye. (iii) There is no synchronization between firing of neurons responding to the good and strabismic eye. Thus, these animals have a perceptual deficit connected to the strabismic eye and this perceptual deficit is accompanied by the desynchronization in the visual cortex. Absence of synchronization means troubles in simultaneous binding of features belonging to one object which results in a blurred vision. Another experimental phenomenon strongly suggesting a one-to-one correspondence between transient synchronizations and perception is binocular rivalry. During binocular rivalry, each eye is constantly stimulated with a different pattern. Visual percept is neither an average of these two patterns nor their sum. Instead, a random alternation between the two percepts occurs as if they were competing with each other, hence the term binocular rivalry. Fries et al. (Fries et al. 1997) discovered that neurons which respond to one or the other pattern are synchronized only during the corresponding percept. Thus, although the pattern is constantly stimulating

40

2 Organization and Functions of the Brain

an eye, cortical neurons get synchronized only when the pattern is perceived.

a

• • c

e

b

' . ·/;::····I ·~~>· · ·

. ', _. \', : • •

'- '

d

..

:

I

.:.

-::

....

. ( , .. • •

\:

. ,

.....

-

Fig. 2.6. (a) Scheme illustrating spatially separated specialized neurons in the visual cortex. Each of them responds to a different elementary feature of one object; (b) Record of firing of these neurons before, during and after the presentation of that object. During the presence of object that activates a given set of neurons, their spikes are synchronized. These synchronizations repeat several times with some period, thus we speak about synchronous oscillations. Period of object presentation is denoted by a thick bar under the spikes. Activities before and after the object presentation are desynchronized spontaneous discharges of neurons; (c) The same as in (a), however this time a different object is presented, which activates a different set of neurons (features); (d) The same as in (b), the spikes of a different set of neurons are synchronized thus binding together features belonging to their object; (e) Illustration of different frequencies and phases of oscillations. Full and dashed curves denote oscillations of the same frequencies but occurring with different phases. The dotted curve denotes an oscillation with the frequency three times frequency of the first two oscillations

An important study of Rodriguez et a1. (Rodriguez et a1. 1999) has demonstrated that perception of faces in humans is accompanied by a transient (~180 ms) synchronization of gamma activity in hierarchically highest visual areas in the parietal cortex and premotor areas in the frontal cortex. They used sophisticated computational procedures for analysis of EEG

2.7 Consciousness

41

signals recorded during the task of face recognition. Humans were presented with the so-called Mooney faces either in the canonic upright position or in an upside down position. These black-and-white pictures are made out of the very high contrast photographs of human faces. It is very difficult to recognize a human face when they are presented upside down. Viewing these images, either in a normal or turned position is always accompanied with the increase of gamma activity in the visual areas. However, precise transient synchronization of gamma oscillations occurs only when the subject perceives a face. It is intriguing that this synchronization occurs only in the left hemisphere which is the so-called conscious hemisphere. When the subject did not perceive a face, but instead only a nonsense arrangement of black-and-white patches, no synchronization happened in the cortex. The second transient episode of synchronization occurred in the premotor and motor areas of both hemispheres during the motor response of subjects by which they indicated whether they saw a face or a no-face. Thus, transient synchronizations may accompany also other cognitive processes not only perception. W. Miltner et al. (Miltner et al. 1999) indeed detected synchronization of gamma oscillations during an associative learning. Humans were supposed to learn to associate a visual stimulus with the tactile stimulus . A selective synchronization occurred between the visual cortex and that part of somatosensory cortex which represented the stimulated hand, during and after the learning. When people forgot the learned association, synchronization between these two stimuli, or rather between neural responses to these two stimuli, disappeared.

2.7 Consciousness When we speak of consciousness we usually mean reflective or secondary consciousness, "the inner eye of our mind". Present neuroscience has a good reason to assume that neural mechanisms of reflective consciousness are derived from the mechanisms of sensory awareness that is related to perception (Singer 1999b). Thus, in building the picture of neural correlates of reflective consciousness we will proceed through assumed neural correlates of sensory awareness that is sometimes referred to as primary or phenomenal consciousness.

2.7.1 Neural Correlates of Sensory Awareness Currently , transient (l00-200 ms) synchronous gamma oscillations are being studied as a promising candidate for the mechanism of binding many

42

2 Organization and Functions of the Brain

elementary features belonging to one object to one transient whole corresponding to a percept of that object (Engel et al. 1999, Singer 1999a). Such synchronized activity summates more effectively than nonsynchronized activity in the target cells at subsequent processing stages, and the activity can spread to a longer distances. If so, synchronization could increase the effect that a selected population of neurons has on other populations with great temporal specificity (in the range of milliseconds). There is also evidence that synchrony is important for inducing changes in synaptic efficacies and hence facilitate transfer of information into memory. Different objects in one scene may be associated with different phase-locked synchronous oscillations within the gamma frequency band. Thus, increased coherence between brain areas confined to a narrow band around 40 Hz may denote a holistic perception of a complex stimulus. Based on experimental findings, crucial neural conditions for a conscious percept to be experienced is (Koch and Crick 1994, Singer 1994, Crick and Koch 1995, Koch 1996, Rodriguez et al. 1999): • Over the chain of primary and higher-order sensory areas with the areas that have direct connections to the frontal cortex being at the end of this chain (e.g. the posterior parietal cortex), and over the evolutionary youngest cortical areas, i.e. the frontal and prefrontal cortex, certain suprathreshold quantity (number) of neurons must be coherently active for a certain time of 100-200 milliseconds (see Fig. 2.7).

Fig. 2.7. Corticocortical connections between the posterior parietal cortex and the main subdivisions of the frontal cortex. Illustrated areas showed increased coherence within the 40 Hz band in the Rodriguez's et al. experiment on recognition of Mooney faces. When a human face was recognized, transient coherence occurred in the time window of 180-360 ms after the beginning of the picturepresentation

2.7 Consciousness

43

Let us go through this condition in greater detail. Why higher-order sensory areas? Because they code for invariant object features, and thus come closer to the invariant object identification (Engel et al. 1999). Primary sensory areas also take part in the chain and their activity must take part in a reciprocal reverberant interaction with higher-order areas (Silvanto et al. 2005). With respect to the quantitative condition, that a certain suprathreshold number of neurons must be synchronized within relevant cortical areas: Electrophysiological measurements on blind-sighted monkeys and fMRI on blind-sighted humans have shown that besides the superior colliculus, also the hierarchically higher visual cortical areas remain responsive to visual stimuli when VI is inactivated or damaged (Sahraie et al. 1997). Thus, blindsight seems to be mediated by both, intact relays within the extra-geniculostriate pathway which leads to superior colliculus, and also by the sparse and spared relays within the retino-geniculatecortical pathways themselves. However, neither subcortical structures nor an insufficient number of active cortical neurons can lead to a conscious percept. There is also an intuition from theory: in order for a large synchronization to occur in some physical system, a certain threshold number of elements must start the process otherwise synchronization does not spread over distance. Generating sensory awareness involves the process of attention. Several areas in the prefrontal cortex are crucially involved in attention, namely areas 8Av (major connections with the visual system), 8Ad (major connections with the auditory system) and 8B (major connections with the limbic system) (Roberts et al. 1998). Attentional selection may depend on appropriate binding (coherence) of neuronal discharges in sensory areas in two simultaneously active directions: an attentional mechanism in prefrontal cortex could induce synchronous oscillations in selected neuronal populations (top-down interaction) , and strongly synchronized cell assemblies could engage attentional areas into coherence (bottom-up interaction) (Singer 1994). Another prefrontal areas activated during sensory perception include areas 9, 10,45,46,47 (see Fig. 2.8). These prefrontal areas are known to be involved in an extended action planning. In addition, these prefrontal areas plus the posterior parietal cortex are known to be involved in the working memory. Posterior parietal cortex is also known to be involved in mental imagery. For planning of actions it is necessary to keep track of at least one sequence of partial actions, hence the overlap between planning and memory mechanisms . It might be that sensory contents reach awareness only if they are bound to prefrontal areas via the posterior parietal cortex and thus have a possibility to become part of the working memory and action planning (Engel et al. 1999). In tum, action planning may influence

44

2 Organization and Functions of the Brain

organization of attentional mechanisms and thus what is being perceived. Actually, the underlying action planning can occur at a subconscious level (Libet 1985,1999). Coherences in the involved areas are generated internally within the cortex and although they are phase-locked, they are not stimulus locked. They are superimposed upon global thalamocortical gamma oscillations which are generated and maintained during cognitive tasks (Ribary et al. 1991). Thalamocortical oscillations may provide the basic oscillatory modulation of cortical oscillations. Other cortical mechanisms are then responsible for a precise phase-locking of internal cortical synchronous oscillations. In particular, these are lateral inhibitory and excitatory interactions, regularly bursting layer V pyramidal cells, and spike-timing dependent rapid synaptic plasticity. In the latest, synapses and thus the inputs which do not drive the postsynaptic cell in synchrony are temporary weakened . Lateral

.~.!3." " " "''' '''' ' '''' ' 7'....~...''

..····8Ad

6

'..

.........·; ;: 6d / ..·8·;..:;;;·

..······)

..... ( ..··· 46

9/46v

88

.......................... , ......i:':

! 44\6

.................. 45A 458

Medi al

cc

~.~

.

9

,

.

/ ..

--'( ;;_ :;=!=3;=:'~ ~O

.

47;:j'2·..·..·

45A 47/12

C·..· ~·3. ·...

~

·11........ 10

....···14

Orb ito-frontal

..../

Fig. 2.8. Human prefrontal cortex. Lateral view (from outside), medial view (from inside) and the orbito-frontal view (from below) at the left hemisphere. The same divisions hold also for the right hemisphere. Numbers denote the corresponding Brodmann's areas. CC means corpus callosum

2.7.2 Neural Correlates of Reflective Consciousness

Since early childhood, we are engaged in learning, first through nonverbal and later through verbal communication, to assume what is going on inside of other people. Our reflective consciousness and our self-reflection develop gradually, step by step. Due to learning to assume what is going on

2.7 Consciousness

45

inside of other people, we can learn to assume what is going on inside of ourselves. Self-reflection is possible only thanks to communication and social interaction. However, it seems that our brains already possess certain structures that have been prepared and selected for this task - these are mirror neurons (Rizzolatti et al. 1996) and mentalization module (Frith 2001). In 1996, scientific community got stirred by the discovery of mirror neurons. G. Rizzolatti et al. (1996) recorded activity of neurons in the ventral premotor cortex of macaque monkeys. They have found out that these neurons are active not only when the monkey executes its own actions, but also when she watches others, either monkeys or humans, to perform the same action. Thus these neurons fire in the same way when others perform a given action as when a monkey performs the same action. The observation/execution matching system represents a given action irrespectively who performs it. Mirror neurons may be a part of the mind reading system. People have the ability to explain and predict behavior of others in terms of their presumed thoughts and feelings. The ability to attribute mental states to others and to ourselves is called "mentalization" or "theory of mind". Noninvasive brain imaging has shown that the ability to attribute various mental states, desires and beliefs to others and also to ourselves depends upon full functioning of a specific neurocognitive module (Frith 2001). The mentalization module includes in both hemispheres: (i) the medial prefrontal cortex (area 32), in particular the most anterior part of paracingulate cortex, a region on the border between anterior cingulate and medial prefrontal cortex (very medial), (ii) the temporal-parietal junction at the top of the superior temporal gyrus (stronger on the right), and (iii) the temporal poles adjacent to the amygdala (somewhat stronger on the left). Neural activity in all three or at least in the prefrontal part of this mentalization module as revealed by the brain imaging is significantly lower in autistic people (Frith 2001). Autistic people are not able to "read out" neither the mind of others nor the mind of themselves, while there is a whole spectrum of severity of autistic disorder. Mentalization module overlaps substantially with the brain higher-order emotional system. Medial parts of the prefrontal cortex (area 32) and orbitofrontal parts (i.e. areas 10, 11, 12, 13, 14) are evolutionary younger parts of the brain emotional system (Damasio 1994). These medial and orbital prefrontal areas are thought to be responsible for the so called secondary emotions. Secondary emotions are emotional feelings based on learned variety of associations between primary emotions and life situations. Hierarchy of these associations involves planning and strategies related to one's social role and personal goals in relation to the past and future. Evaluation and planning and feelings in the social and emotional spheres are therefore linked to be processed by the same structures in the prefrontal cortex, and

46

2 Organization and Functions of the Brain

these overlap with the mentalization module. According to Damasio (1994), the feeling of self, and consequently the awareness of self, would depend also on the intactness of the somatosensory system, on the signaling from the cortex down to the body and back. As we have said, it might be that sensory contents reach awareness only if they are temporarily synchronized with activity in the prefrontal areas, thus displaying a highly coherent joint activity. In such a way they can become part of conscious working memory and action planning. According to Singer (l999a) reflective consciousness would be based upon the same processes, i.e. highly coherent activity, happening over the prefrontal areas involved in planning and working memory and between areas devoted to representations of our inner world. Secondary consciousness or metaawareness would result from iteration of the very same processes that support primary consciousness, except that they are not applied to the signals arriving from the sensory organs, i.e. from the outer world, but to the outputs of previous cognitive operations (Singer 1999b). By means of recording electromagnetic activity of the brain it is possible to capture and visualize the fast semi-global coherent activity of the brain that accompanies conscious perception of a stimulus (Tononi and Edelman 1998). It is almost a magical view, because this semi global coherent activity changes with time as a burning fire which is boosted from the centre of the brain and its flames transiently engage currently synchronized brain areas (see Fig. 2.9). Why is the fire boosted from the centre of the brain? Clinical research has revealed that damaging intralaminar nuclei in the thalamus leads to the loss of consciousness and to coma. Neurons in intralaminar nuclei possess dense reciprocal connections to and from the brain cortex. Intralaminar nuclei are the source of arousal, without which the cortex cannot function. G. Edelman and G. Tononi (Edelman and Tononi 2000) call this ever changing semi global coherent activity, the dynamic core. The dynamic core corresponds to a large (semi global) continuous cluster of neuronal groups that are coherently active on a time scale of hundreds of milliseconds. Its participating neuronal groups are much more strongly interactive among themselves than with the rest of the brain. The dynamic core must also have an extremely high complexity as opposed to for instance convulsions. Each roughly 150 ms, a pattern of semi-global activity must be selected within less than a second out of a very large, almost infinite, repertoire of options. Thus, the dynamic core changes in composition over time. As suggested by imaging, exact composition of the core varies significantly not only over time within one individual, but also vary significantly across individuals.

2.7 Consciousness

47

According to Edelman and Tononi (2000), the dynamic core consists of a large number of distributed groups of neurons which enter the core temporarily based on their mutual coherence. Connecting groups of neurons into temporarily synchronized whole requires dense recurrent connections between brain areas, along which a reiterated reentry of signals occurs. Neural reference space for any conscious state may be viewed as an abstract N-dimensional space, where each axis (dimension) stands for some participating group of neurons that code for (represent) a given aspect of the conscious experience. There can be hundreds of thousands of dimensions. The distance from the beginning of the axis represents the salience of that aspect. It may, for instance, correspond to the number of firing neurons within a given group. We would like to point out the interesting similarity between this abstract N-dimensional neural space and the conceptual spaces introduced by Gardenfors (2000). Heanng

Imagery

Fig. 2.9. (a) Illustration of the dynamic core, a changing coherent semi global activity of the brain, which is supposed to be a neural correlate of consciousness. One configuration of the core lasts for about 150 ms; (b) Interpretation of the dynamic core as an N-dimensional neuronal reference space, where each axis (dimension) denotes some group of neurons which encodes (represents) a given aspect of the conscious experience. Each axis can be broken down into more elementary axes. There can be hundreds of thousands of dimensions

What would be, in this theory, a neural basis for subconsciousness? The same group of neurons may at times be part of the dynamic core and underlie conscious experience, while at other times it may not be part of it and thus be involved in subconscious processing. Koch and Crick (1994) have proposed that those active neurons which are not at the moment tak-

48

2 Organization and Functions of the Brain

ing part in the semi global activity keep processing their inputs, and results of this processing may still affect behavior. We would like to mention also the explanation of neural correlate of qualia or the hard problem of consciousness, according to Edelman and Tononi (2000). Qualia are specific qualities of subjective experiences, like redness, blueness, warmth, pain, and so on. According to the dynamic core hypothesis, pure redness would be represented by one particular state of the dynamic core that is by one and only one point in the N-dimensional neural space. This core state would certainly include large participation of neurons that code for the red color and a small participation of neurons that code for other colors and for anything else. Coordinates of a point in the Ndimensional reference neural space are determined by activities of all neuronal groups that are at the moment part of the core. And these activities vary in time and across individuals. Thus, the subjective experience of redness will be different in different people and can be different for the same individual for instance in the morning and in the evening. Another mystery is why consciousness fades as we fall asleep even when we nowadays know that the brain and especially the cortex remain highly active. Sleep research has revealed that during sleep, humans normally go through two-three cycles of two sleep phases. One of these two phases is the so called REM sleep, according to the accompanying Rapid Eye Movements. EEG activity of the brain during the REM phase is very similar to the EEG activity of the awake brain during cognitive activity. Hence the term paradoxical sleep for the REM sleep phase, as it was not sleep at all. We dream mostly during REM sleep phases. When awakened during the REM phase, we can recall the content of a dream. We experience self-awareness when we dream but not when we are in the deep sleep (Llinas and Ribary 1994). Thus "1" is preserved during dreaming as well as the awake-like EEG activity of the brain. When awakened around at the end of the REM phase, we can remember that we dreamt, not knowing about what. When awakened during the non-REM sleep phase, we mostly deny any experience of dreaming. The non-REM sleep phase is also called the deep sleep, and the brain activity occurs in typical slow large regular waves. Recently, experiment with the spread of activity within neocortex during sleep have revealed that different cortical areas stop communicating over distance with each other during the non-REM sleep - a stage of sleep for which people mostly report no or very little conscious experience on waking (Massimini et al. 2005). Thus, it seems that the coherent semi global activity is disrupted during the non-REM sleep, and so is the conSCIOUS awareness.

2.8 Summary and Discussion

49

2.8 Summary and Discussion Neurobiologists have made a great progress in trying to find the so-called neural correlates of consciousness. However, not every scientist finds this research compelling . In his influential book Shadows of the Mind, physicist Roger Penrose brings the problem of explaining consciousness to the domain of elementary laws of physics (Penrose 1994). According to him, at present any scientific (including physical) theory does not help us to come to terms with the puzzle of mentality including consciousness within such a physically determined universe. He has been calling for a radical upheaval in the very basis of physical theory. He sees the consciousness link in a new physical theory based upon the union of Einstein's general relativity with quantum theory. His critics question the competence of physics ever having anything of importance to say about mental phenomena in general, and consciousness in particular (http://psyche.cs.monash.edu.au/psyche-index-v2 .html#som). The grounds for this criticism are different, ranging from computational to neurobiological arguments (Koch and Hepp 2006). Nevertheless , some neurobiologists find quantum states of microtubules, tiny elongated organelles spanning the interior of neurons, to be the gate to consciousness (Hameroff and Penrose 1996). There is indeed a puzzling temporal aspect to the consciousness. Famous experiments of Benjamin Libet on human subjects showed that it takes a while, about 0.5s, for a conscious awareness to develop (Libet 1985). Acts initiated by a subject's free will are preceded by a specific electrical change in the brain, the so-called readiness potential (RP), which begins 550 ms before the act itself. Human subjects became aware of their own intention to act 350-400 ms after RP started, but 200 ms before the act execution (which is about the time of the process of initiation and generation of motor movements) . These data pose many questions including philosophical about the nature of free will, whether it exists at all (Libet 1999). While the author remains optimistic about this issue, the question still remains why it takes so long for a conscious awareness to develop. Why it develops at all? Which processes are responsible for it and how they lead to consciousness? Even concerning the mystery of mentality, there is not a general agreement among scientists. Some think that there is no mystery at all, and consciousness and other mental phenomena are equivalent to a particular underlying process, be it the specific computations (Dennett 1991), brainspecific processes (Searle 2002), or complexity (Edelman and Tononi 2000). In another influential book written on consciousness, philosopher

50

2 Organization and Functions of the Brain

and mathematician, David J. Chalmers clearly argues that consciousness and mentality are indeed genuinely puzzling and not explainable by present theories (Chalmers 1996). Chalmers asks: why should quantum processes (or any other specific physical processes) in microtubules (or any other brain substructures) or any specific computational processes give rise to consciousness? If one takes consciousness seriously, Chalmers says, one has to go beyond a strict materialist framework. The fundamental laws linking the physical and the experiential are yet to be discovered. The exercise for the reader can be to think about the nature of such fundamental laws. Although a lot is known about the brain, issues about its functioning, representation and processing of information are still subjects of an intense research. The nature of brain dynamics is still unknown. Some researchers find evidence of chaos, whereas some are doubtful (Theiler 1995). Main proponents of a chaotic dynamics, W.J. Freeman (Freeman 2003) and I. Tsuda (Tsuda 2001), argue in favor of chaotic itinerancy based on EEG and other neurophysiological data. According to the picture of chaotic itinerancy, a complex system such as the (human) brain evolves by steps along a trajectory in the state space. Each step corresponds to a shift from one basin of attraction to another. Attractors represent classes for abstraction and generalization. Thus, the brain states evolve aperiodically through sequences of attractors. In a closed system the next attractor would be chosen solely by internal dynamics. In an open system, such as the brain, external inputs interfere with internal dynamics. Moreover, due to the changes induced by learning, trajectories continually change. Chaotic itinerancy occurs in sequence of cortical states marked by state transitions that appear in temporal discontinuities in neural activity patterns (Freeman 2003). Experimental EEG data show that the entire cerebral cortex is constantly wandering in the fractal distributions of phase transitions that give the 1/j form of the temporal and spatial frequency spectra (with a E (1,3), (Freeman 2003)). From this type of frequency spectra it appears that the brain maintains a state of self-organized criticality (Bak et al. 1987). The self-organized criticality state can form the basis of the brain capacity to rapidly adjust to new external and internal stimuli. State changes resembling phase transitions occur continually everywhere in cortex at scales ranging from millimeters to ~O.l m. Local neural activity can trigger a massive state change. However, several issues of caution should be pointed out. In spite the compelling evidence for self-organized criticality in the brain, the nature of the critical state is still unknown in neurobiological interpretation. The spatial and temporal power spectral densities often show the l/j form, however more often this form is broken down due

2.8 Summary and Discussion

51

to the distortions by clinically defined peaks. Therefore the measurements of a vary widely. Aperiodic oscillations giving the llj power spectral densities are commonly referred to as chaotic. However the brain activity is not at all consistent with low-dimensional deterministic chaos (Theiler 1995, Benuskova, Kanich et al. 2001). It is high dimensional, noisy, nonGaussian, and nonstationary (Freeman 2003). Tremendous physical complexity of the brain arises also from the fact that it is not a homogenous tissue. Each part of the brain is morphologically different and has its own genetic profile as can be seen by analysis of large-scale human and mouse transcriptomes. Therefore the conditions for assessment of the type of dynamics are difficult to be met. Moreover brains are open systems driven by stochastic input. Thus it seems that the brain activity hardly can conform to the mathematical definitions of chaos. Whether the term chaotic itinerancy (or any other term from the chaotic vocabulary) is appropriate to describe state transitions in brain and cortex in particular remains open to challenge. Thus, the complex spatio-temporal activity data from the brain still awaits explanation.

3 Neuro-Information Processing in the Brain

While Chapter 2 presents the higher-level brain organization, this chapter presents a view on the low level information processing in the brain. Neuro-infonnation processing in the brain depends not only on the organization of the brain and properties of brain neural networks, but also on the properties of processing units - neurons and signal processing networks within neurons. These internal networks involved in signal processing are comprised of second and third messengers, enzymes, transcription factors and genes.

3.1 Generation and Transmission of Signals by Neurons A neuron (Fig. 3.1) receives and sends out electric and chemical signals. The place of signal transmission is a synapse. In the synapse, the signal can be nonlinearly strengthened or weakened . The efficacy of synaptic transmission is also called a synaptic weight or synaptic strength. One neuron receives and sends out signals through 103 to 105 synapses. Dendrites and soma, i.e. the body of a neuron, constitute the input surface for receiving signals from other neurons.

axon

basal dendrites dendritic tree

synapse

Fig. 3.1. Schematic illustration of a neuron and its parts. There is a synapse at every dendritic spine. Synapses are also formed on the dendritic shafts and on the soma

54

3 Neuro-Infonnation Processing in the Brain

Dendritic tree consists of thousands of dendrites which are covered by tiny extensions called spines. Most of synapses are formed on dendrites, particularly on spines. Spines are very important devices in relation to learning and memory, as we will see later. Electrical signals transmitted by synapses can have a positive or negative electric sign. In the former case, we speak about excitatory synapses and in the latter case about inhibitory synapses. When the sum of positive and negative contributions (signals) weighted by synaptic weights gets bigger than a particular value, called the excitatory threshold, a neuron fires, that is, emits an output signal called a spike. A spike is also called an action potential or a nerve impulse. Usually, as a result of synaptic stimulation and summation of positive and negative signals, a neuron fires a whole series (train) of spikes (Fig. 3.2). Mean frequencies of these spike trains range from 1 to102 Hz. The output frequency is proportional to the overall sum of positive and negative synaptic contributions. Spikes are produced at the initial segment of an axon, the only neuronal output extension. Then they propagate very quickly along the axon towards other neurons within a network. Propagation speed of nerve impulses is 5-100m/s. At its distant end, an axon makes thousands of branches, each of which is ended by a synaptic terminal (bouton). Spike train

11111 - -

Fig. 3.2. Generation and propagation of spikes in neurons. EPSP = excitatory postsynaptic potential, IPSP = inhibitory postsynaptic potential, LPSP = total postsynaptic potential equal to EPSP - IPSP, & = excitatory threshold

Transmission of signals from one neuron to another occurs in synapses. A synapse consists of a presynaptic terminal (bouton), synaptic cleft and postsynaptic membrane (Fig. 3.3). In the presynaptic terminal, there are dozens of vesicles filled with molecules of neurotransmitter (NT) ready to be released. When a presynaptic spike arrives into a terminal, calcium ions rush in and cause the fusion of vesicles with the presynaptic membrane. This process is also called exocytosis. Molecules of NT are released into the synaptic cleft (Fig. 3.3b), and diffuse towards the receptors within a postsynaptic membrane. Molecules of NT form a transient bond with the

3.1 Generation and Transmission of Signals by Neurons

55

molecules of receptors. This causes opening of ion channels associated with postsynaptic receptors. In the excitatory synapse, receptors are associated with sodium (Na+) ion channels, and a positive excitatory postsynaptic potential (EPSP) is generated. In the inhibitory synapse, receptors are associated with chlorine «(1) ion channels, and a negative inhibitory postsynaptic potential is generated. Alternatively, there can be ion channels for potassium (K+), which flows out and thereby lowers the interior potential. Eventually, NT releases its bond with receptors and diffuses back to the presynaptic membrane and out of the synaptic cleft. Special molecular transporters within a presynaptic membrane take molecules of NT back inside the terminal, where they are recycled into new vesicles. This process is called a reuptake of NT. The whole synaptic transmission lasts for about I millisecond. Such a synapse is called a chemical synapse, because the transmission of an electric signal is performed in a chemical way.

10 " m

pos tsynaptic membrane

Fig. 3.3. Scheme of synaptic transmission. (a) Synapse is ready to transmit a signal. NT = neurotransmitter, R = postsynaptic receptor. (b) Transmission of an electric signal in a chemical synapse upon arrival of an action potential into the terminal (plus signs around the terminal). AMPAR = AMPA-receptor-gated ion channel for sodium, NMDAR = NMDA-receptor-gated ion channel for sodium and calcium

The postsynaptic potential (PSP), either excitatory or inhibitory, has some amplitude and duration. The amplitude and duration of PSP depend upon the number of activated receptor-ion channels and upon the time for how long they stay open. This may last for milliseconds, tens of milliseconds or hundreds of milliseconds. The duration of channel opening depends upon the number of released NT molecules and upon the type of receptors that are associated with ion channels. The amplitude of PSP also

56

3 Neuro-Information Processing in the Brain

depends upon the electric input resistance for ions, which in tum depends upon the size and shape of a postsynaptic spine and dendrites, and upon the distance of synapse from soma. For instance, a short and stubby dendritic spine has a much smaller electric resistance than a long and thin spine. All these pre- and postsynaptic factors determine the weight (strength, efficacy) of a particular synapse. Within a postsynaptic membrane, there are also kinds of receptors that are not associated with an ion channel, but instead with an enzyme. When the overall amount of released NT reaches some critical concentration, these receptor-enzyme complexes activate particular cytoplasmatic enzymes, the so-called second messengers. Second messengers trigger chains of various biochemical reactions which may lead to a change in synaptic weight, or even to transient changes in gene expression leading to alteration in biomolecular synthesis of receptors, neurotransmitters and enzymes. Thus, second messengers may act locally within a synapse itself, or they may activate further (third and so on) messengers that carry the message to the genome of a neuron, thus causing a change in its biochemical machinery related to signal processing. Therefore, it is now widely accepted that the activity of a neuron itself, influences its processing of information, and even its life itself, whether it survives or not.

3.2 Learning Takes Place in Synapses: Toward the Smartness Gene For major discoveries in the field of synaptic mechanisms of learning, the 2000 Nobel Prize for medicine went to the neuroscientists Eric R. Kandel and Paul Greengard. The 3rd laureate, Arvid Carlsson, got his share of the prize for discoveries of actions of neurotransmitter dopamine. At present, it is widely accepted that learning is accompanied by changes of synaptic weights in cortical neural networks (Kandel et al. 2000). Changes of synaptic weights are also called synaptic plasticity. In 1949, the Canadian psychologist Donald Hebb formulated a universal rule for these changes: "When an axon of cell A excites cell B and repeatedly or persistently takes part in firing it, some growth process or metabolic change takes place in one or both cells so that A's efficiency as one of the cells firing B is increased", which has been verified in many experiments and its mechanisms elucidated (Hebb 1949). In cerebral cortex and in hippocampus of humans and animals, learning takes place in excitatory synapses formed upon dendritic spines that use glutamate as their neurotransmitter. In the regime of learning, glutamate

3.2 Learning Takes Place in Synapses: Toward the Smartness Gene

57

acts on specific postsynaptic receptors, the so-called NMDA receptors (Nmethyl-D-aspartate). NMDA receptors are associated with ion channels for sodium and calcium (see Fig. 3.3). The influx of these ions into spines is proportional to the frequency of incoming presynaptic spikes. Calcium acts as a second messenger thus triggering a cascade of biochemical reactions which lead either to the long-term potentiation of synaptic weights (LTP) or to the long-term depression (weakening) of synaptic weights (LTD). In experimental animals, it has been recorded that these changes in synaptic weights can last for hours, days, even weeks and months, up to a year. Induction of such long-term synaptic changes involves transient changes in gene expression (Mayford and Kandel 1999, Abraham et al. 2002). A subcellular switch between LTD and LTP is the concentration of calcium within spines (Shouval, Bear et al. 2002). We speak about an LTD/LTP threshold. In tum, the intra-spine calcium concentration depends upon the intensity of synaptic stimulation that is upon the frequency of presynaptic spikes. That is, more presynaptic spikes mean more glutamate within synaptic cleft. Release of glutamate must coincide with a sufficient depolarization of the postsynaptic membrane to remove the magnesium block ofthe NMDA receptor. The greater the depolarization, the more ions of calcium enters the spine. Postsynaptic depolarization is primarily achieved via AMPA (amino-methylisoxasole-propionic acid) receptors, however, recently a significant role of backpropagating postsynaptic spikes has been pointed out (Markram et al. 1997). Calcium concentrations below or above the LTD/LTP threshold, switch on different enzymatic pathways that lead either to LTD or LTP, respectively. However, the current value of the LTD/LTP threshold (i.e. the properties of these two enzymatic pathways) can be influenced by levels of other neurotransmitters, an average previous activity of a neuron, and possibly other biochemical factors as well. This phenomenon is called metaplasticity, a plasticity of synaptic plasticity (Abraham and Bear 1996). Dependence of the LTD/LTP threshold upon different postsynaptic factors is the subject of the Bienenstock, Cooper and Munro (BCM) theory of synaptic plasticity (Bienenstock et al. 1982) (for a nice overview see for instance (Jedlicka 2002)). The BCM theory of synaptic plasticity has been successfully applied in computer simulations to explain experience-dependent changes in the normal and ultrastructurally altered brain cortex of experimental animals (Benuskova et al. 1994). The easiness with which LTD and LTP can be evoked in the developing and in the adult brain are not the same. One of the factors responsible for this difference may be the genetically programmed difference in the NMDA receptor composition (Bliss 1999). The NMDA receptor is made up of an NR1 subunit, which is obligatory for channel function, and a se-

58

3 Neuro-Information Processing in the Brain

lection of developmentally and regionally regulated NR2 subunits (A to D). For example, the glutamate-evoked positive current has a longer duration in receptors containing NR2B subunits than in those containing NR2A subunits. The proportion of NR2B subunits is higher in young animals than in adults, and this probably accounts for the greater degree of synaptic plasticity seen in young animals that is accompanied by greater easiness of youngsters in formation of new memories (Bliss 1999). This conclusion was indeed confirmed in another study, in which scientists inserted an extra copy of the NR2B gene into mice (Ping et al. 1999). Mice with the NR2B gene insertion performed better than mice without the insertion on all of the tests used to evaluate their intelligence and memory. Adding a single gene to mice significantly boosted the animals' ability to solve maze tasks, learn from objects and sounds in their environment and to retain that knowledge. This strain of mice also retained into adulthood certain brain features of juvenile mice, which, like young humans, are widely believed to be better than adults at grasping large amounts of new information. The research proves that the NR2B gene is a key switch that controls the brain's ability to associate one event with another, the core feature of learning. The finding also shows that genetic improvement of intelligence and memory in mammals is now feasible, thus offering a striking example of how genetic technology may affect mankind and society in future. This research can ultimately lead to human gene therapy for use in areas such as dementia, mental retardation, etc, although more research is needed on comparison of regulation and functions of human and mouse NMDA receptors.

3.3 The Role of Spines in Learning Dendrites of cortical excitatory pyramidal neurons are abundant in tiny membrane extensions called spines. They are named so because they resemble in shape the spines on the rose stem. About 80% of all synaptic connections in the cerebral cortex are excitatory and vast majority of them is formed on the heads of synaptic spines. For many years the role of spines was a mystery. Nowadays it is accepted that they play several important roles in synaptic plasticity and learning. First, it was discovered that spines change their size, shapes and numbers during the induction and maintenance of LTP (Lee et al. 1980, Geinisman et al. 1991). There are growth changes on spines, like spine head swelling, spine neck thickening, and increase in appearance of spines with mushroom-shaped heads. Morphological properties of spines and changes

3.3 The Role of Spinesin Learning

59

in their shape were first supposed to playa role in affecting the efficacy of synaptic transmission by means of changes in the input resistance (Koch and Poggio 1983). Long, thin spines create a big input electrical resistance, while short, stubby spines create a smaller input resistance. Later, a role in sequestering and amplifying the calcium concentrations was suggested to be the main role of spines (Zador et al. 1990). Through this role a mechanisms for saturation and stopping the infinite growth of synaptic weights was proposed, as well as the role in the LTP/LTD threshold (Gold and Bear 1994). While all these effects can take place, another important role for spines was suggested in the transport of new receptors into the spine head (Benuskova 2000). This model is based on our older hypothesis that the changes in efficacy of excitatory dendritic spine synapses can result from the fusion of transport vesicles carrying new membrane material with the postsynaptic membrane of spines (Fedor et al. 1982). Spacek and Harris indeed found structural evidence for exocytotic activity within spines in hippocampal CAl pyramidal neurons (Spacek and Harris 1997). Smooth vesicles of the diameter around 50 nm occurred in the cytoplasm of spine heads, adjacent to the spine plasma membrane, and fusing with the plasma membrane. In addition, (Lledo et al. 1998) showed that inhibitors of membrane fusion blocked or strongly reduced LTP when introduced into CAl pyramidal cells. On the other hand, an increase in synaptic strength was elicited when membrane fusion was facilitated. In the CA 1 region, LTP requires the activation of the NMDA glutamate receptors and a subsequent rise in postsynaptic calcium concentration. Besides other roles, Ca2+ plays a crucial role in the final stage of vesicle fusion with the membrane, and the number of fused vesicles is proportional to [Ca2+] (Sudhof 1995). Since LTP in CA 1 neurons is accompanied by appearance of AMPA subclass of glutamate receptors (Liao et al. 1995), it is reasonable to assume that vesicles can be a mean of their insertion. Indeed, Kharazia et al. observed GluRI (a subunit of AMPA receptors) containing vesicles associated with the cytoplasmic side of some GluRI-containing cortical synapses (Kharazia et al. 1996). Moreover, tetanic stimulation induces a rapid delivery of GluRI into spines and this delivery requires activation of NMDA receptors (Shi et al. 1999). Another effect of the vesicle fusion with the spine membrane would be the shaping and growth of the spine, which were observed during the induction and maintenance of LTP. However, prior to fusion the vesicles must get very close to the plasma membrane. The main mechanism for displacement of vesicles within axons and dendrites is the fast active transport with the speeds of 0.001-0.004 um/ms (Schnapp and Reese 1986). Fast transport depends on the direct interaction of transported vesicles with microtubules via the translocator kinesin-like molecules

60

3 Neuro-Information Processing in the Brain

(Schnapp and Reese 1986). However, microtubules do not enter spines (Spacek and Harris 1997). Thus, while the fast transport can bring vesicles close to the walls of dendritic shafts, another mechanism must come into play within spines themselves. The first natural candidate for this mechanism can be the diffusion of vesicles. However, we have shown that an electrophoretically driven, directed motion of negatively charged vesicles towards the spine head, evoked by the synapse stimulation itself can be ten times faster (Benuskova 2000). We have estimated the intensity of intra-spine electric fields triggered by stimulation of excitatory spine synapses. We have shown that this electric force can cause fast electrophoretic movement of negatively charged vesicles which bring new postsynaptic receptors and membrane for insertion during the induction of LTP. Due to the direction of an intra-spine electric field, movement of vesicles is electrophoretically directed along the longitudal spine axis towards the spine head. Spinnulae, small extrusions in the middle of the spine head might be the witnesses of such a directed vesicle fusion. The number of fused vesicles may be proportional not only to the increased calcium concentration within the spine head during the induction of LTP but also to the magnitude of electric force which drives vesicles towards the postsynaptic membrane. The thicker the spine gets, the smaller the electric field becomes, thus none or far less vesicles would get to the postsynaptic membrane in time to catch up with the increased [Ca2+J during NMDAR stimulation. This effect is congruent with the effect of spine dimensions on [Ca2+] (Gold and Bear 1994), and thus may also be a part of saturation mechanism in LTP. Electrophoretic hypothesis is illustrated in Fig. 3.4.

Fig. 3.4. Schematic illustration of the electrophoretic hypothesis linking the stimu-

lation of the spine excitatory synapses with the morphological changes and insertion of new receptors during LTP. Arrows point in the direction of the intra-spine electric field E. E is getting smalleras the spine gets thickerdue to new membrane insertion. Spinulae in the middle of the spine head would be the place where the vesicles fuse most often due to the direction of E. SER = smooth endoplasmic reticulum

3.4 Neocortical Plasticity

61

Morphological changes on spines, dendrites and growth of synapses probably constitute the mechanism of long-term and permanent memory storage (Rhawn 1996, Bailey et al. 2004, Maviel et al. 2004). There is a local machinery for protein synthesi s at spines and dendrites consisting of polyribosomes, tRNAs , microRNAs, initiation factors and mRNAs for glutamate receptors, structural proteins and kinases like CaMKII (Steward 1997). MicroRNAs are small, non-coding RNAs that control the translation of target messenger RNAs (mRNAs). It has been shown that both sensory experience and synaptic stimulation leads to translation of a-CaMKII (Wu et al. 1998), the kinase that is crucially involved in the induction of LTP. In addition it has been shown that a brain-specific microRNA, miR134, is localized at the synapto-dendritic compartment of rat hippocampal neurons and regulates the size of dendritic spines. This effect is mediated by miR-134 inhibition of the translation of an mRNA encoding the protein kinase , Limkl (Lim-do main- containing protein kinase) , that controls spine development (Schratt et al. 2006).

3.4 Neocortical Plasticity The ability of the brain to wire and rewire itself in response to changes in experience has become known as experience-dependent plasticity. The brain is able to remodel its connections in order to adjust the organism's respon se to changing conditions. Experience-dependent neocortical plasticity means that cortical neurons change their response characteristics due to altered stimulation of cortical inputs (i.e. experience). Amount of experience-dependent cortical plasticity, i.e. the degree of changes, is the biggest in early development and decays with age, however it never ceases totally. Developmental plastic ity refers to the cortical changes during the early stages of postnatal brain development. 3.4.1 Developmental Cortical Plasticity With respect to the cortical developmental plasticity, the most studied and best known neural system is the visual system of mammals. It is assumed that the visual system of primates has around twenty hierarchical processing levels (areas) , starting from the eye retina, going through the LGN in thalamus and primary visual cortex in the occipital cortex, to tens of the higher-order visual areas in the parietal and temporal cortices. Neurons at every hierarchical level of the visual system respond to different abstract elementary features of visual objects. These elementary features are for

62

3 Neuro-Information Processing in the Brain

example, edges of light intensity, angles between edges, distances between points (disparity), colors, movements, direction and/or speed of the movement, and so on. Higher-order visual neurons respond to various combinations of elementary features processed at lower levels. The most studied cortical visual area is the primary visual cortex (visual one, VI), which is the hierarchically lowest visual cortical area. The Nobel laureates, David Hubel and Torsten Wiesel, discovered that neurons in VI have different orientation selectivity and ocular dominances. Orientation selectivity means that different neurons respond to the optical stimulation by the bars (edges) of different angular orientations that are like this: \ I/ / - , and so on, over the 360 0 circle. Ocular dominance means the degree to which a neuron is dominated by the right or left eye. When a neuron is dominated by both eyes equally, we say it is binocular. Hubel, Wiesel, and many other scientists have studied whether the distribution (pattern) of ocular dominances and orientation selectivity within neurons of VI is inborn or acquired by experience. Currently, it is widely accepted that these patterns are in part inborn however for the development of the normal sharp binocular vision, normal visual stimulation is necessary. Immediately after eye opening, neurons in VI are not silent but instead each one responds to an optical stimulation by bars of several close orientations. It is said that neurons are broadly tuned. There is also some initial ocular dominance distribution. Degrees of ocular dominance are equally distributed. It can be said that nature (i.e. innate factors) provides some initial "scaffold", in the sense of a "framework outlining parts to be formed on it later". In time, and experiencing normal (natural) visual stimulation, neurons become sharply tuned and the distribution of ocular dominance becomes skewed towards binocular neurons. The development of sharply tuned binocular neurons is completed during the so-called critical period of postnatal development. In other words, the critical period is a limited time period after birth, only during which the development of normal response characteristics of neurons is possible. After eye opening, the critical period of cats lasts for about 2-3 months the critical period of monkeys lasts for about 6 months, and in humans up to about 5-6 years after birth. Developmental plasticity, i.e. the degree of evoked long-lasting changes in response characteristics of cortical neurons, decays with time during the critical period and ceases by its end. Abnormal visual experience during the critical period causes the development of irreversible abnormal response characteristics of cortical neurons. Such an abnormal visual experience can change even the initial response properties of neurons. Thus, a "scaffold" itself, i.e. the innate basis for normal response

3.4Neocortical Plasticity

63

characterist ics, can be disrupted by an inappropriate experience (at least in the visual cortex). Examples of abnormal visual experience during the critical period that lead to the development of permanent and irreversible changes in response characteristics of cortical neurons in VI (the same outcomes were observed for monkeys and cats) (Clothiaux et al. 1991): • Binocular deprivation (BD). When both eyes are close shut during the whole critical period, both eyes are completely deprived of visual inputs. Neurons in VI remain broadly orientation selective and the distribution of ocular dominances remains unchanged . At the perceptual level, poor orientation selectivity results in blurred vision, which can not be corrected by any means. • Monocular deprivation (MD). One of the eyes is close shut during the whole critical period. Neurons in VI become dominated only by the open eye. Neurons dominated by the open eye develop normal, i.e. sharp tuning to oriented bars, and thus gain normal orientation selectivity properties. These response characteristics are irreversible, even when the closed eye is opened at the end of the critical period and remains open for the rest of life. At the perceptual level, an animal stays blind on the closed eye for the whole life and has a sharp vision connected only to the open eye that received a normal visual input. • Normal rearing (NR) after MD. When after cortical neurons loose their responsivity to the closed eye, this eye is opened and the critical period is not over yet, neurons in V 1 regain binocular properties and develop normal orientation selectivity. A normal sharp vision can be restored. • Reverse suture (RS). One of the eyes is close shut for some time since the beginning of the critical period. After cortical neurons loose their responsivity to the closed eye, this eye is opened and the formerly opened eye is closed (before an end of the critical period). Cortical neurons in VI loose their responsivity to the newly closed eye, and regain and retain responsivity to the newly opened eye. Properties and time course of cortical plasticity in the auditory cortex is not very well known. Only recently it has been shown that the development of normal cortical representations of sounds correlates with the capacity to hear. Prof. Klinke 's team from the Goethe University in Frankfurt has advanced our knowledge about the developmental plasticity in the auditory cortex (Klinke et al. 1999). First they demonstrated that the cats that are deaf because of the innate (congenital) degeneration of the organ of Corti do not have normal representations of sound frequencies in their auditory cortex, although there is some weak representation of sounds be-

64

3 Neuro-Information Processing in the Brain

cause some sounds can be transmitted through bones in the skull. Scientists implanted the so-called cochlear implant into the inner ear of deaf kittens which were 2 to 4 months old. The cochlear implant is an electronic device that enables the transformation of sound air waves into electric stimulation of an auditory nerve (which is spared in these cats) and thus enables to relay sounds into the auditory cortex. They used a cochlear implant that was a crude substitute for a true cochlea, in the sense it was far from the faithful replication of natural stimulation. In spite of that, after a period of I to 3 weeks the kittens, completely deaf before they received the implant, started to hear. Perceptual improvement was accompanied by the development of almost normal sound maps (representations of sound frequencies) within auditory cortices of these cats. Thus, the cats have not started to hear immediately after they received a cochlear implant. Instead, they started to hear only after almost normal neural sound representations have developed in their auditory cortex. Degree of plastic changes in the auditory cortex in response to auditory stimulation decreases with age, although it seems that there is not such a sharp critical period like in the visual system. 3.4.2 Adult Cortical Plasticity

For the first time, the experience-dependent neocortical plasticity of the adult cerebral cortex was demonstrated in the somatosensory cortex of monkeys. At any moment, millions of signals from the whole body, its internal and external surface, travel to the brain through myriads of peripheral nerves and the spinal cord. Processing of bodily signals, like touch, warmth, pain, takes place in the somatic sensory cortex, which occupies the postcentral gyrus of parietal cortex (see Fig. 3.5). Human (or monkey or any other animal) body is mapped upon the somatosensory cortex in a topographic order. Topography means that neighboring body parts occupy neighboring parts in the map. A body map is deformed, i.e. those body parts that are often actively used are represented by bigger neural areas. Jon Kaas from Vanderbilt University and Michael Merzenich from University of California were among the first to discover that none of the two somatosensory maps are the same. They have discovered that individual variations in body representations reflect variances in tactile experience of different individuals especially with respect to fingers (Kandel et al. 2000). Details of cortical maps represent individual experience. Many experiments support their conclusions. For instance, in one experiment, monkeys were trained to touch a rotating disk for one hour per day. Before the experiment and several weeks later, scientists

3.4 Neocortical Plasticity

65

measured an extent of finger representation in the somatosensory cortex. They discovered that the cortical area representing stimulated finger tips has enlarged. Each finger has its own discrete area of the hand map. Neurons in area 1 respond only when touching finger 1, and so on. After differential tactile stimulation of the tips of fingers 2, 3 and 4, their cortical area has enlarged. In another experiment researchers sutured the skin of the fingers 3 and 4. Thus, a monkey had to use these two fingers as a single one. In the somatosensory hand map, an entirely new representation has developed. On the border between finger 3 and 4 exclusive representations, a band of neurons responding only to the joint stimulation of these two fingers has emerged.

Fig. 3.5. Schematic illustration of the bodily representation in the somatosensory cortex. Body surface is represented topologically while more important parts like face, lips, hands and especially fingers take more cortical representation than other parts like back, legs, arms, etc.

Similar somatosensory plasticity has been revealed in humans by means of magnetoencephalography (Mogilner et al. 1993). Adult humans were studied before and after surgical separation of inborn webbed fingers (syndactyly). The presurgical map contained nonsomatotopic hand representation. Within weeks after surgery, cortical reorganization over distances of 3-9 mm was evident, correlating with the new functional status of their separated digits. On the other hand, 95-100 % of people who have lost their arm or leg experience the phantom limb (Melzack 1999). They feel their lost limb as if it was still there, they feel it moving when they move, and it is still part of their subjective body experience. They can feel pain in it, touches, warmth, cold, and so on. At the beginning, the phantom limb

66

3 Neuro-Information Processing in the Brain

has a normal shape and size, but in time it starts to change: it may float freely in the air, contract into to the stump, or acquire any other bizarre shape, size or connection with the rest of the body. Nevertheless, it always remains an integral part of the patient's self as if the reorganization of large somatotopic maps was not entirely possible. 3.4.3 Insights into Cortical Plasticity via a Computational Model As we have seen although some forms of experience-dependent cortical plasticity disappear at the end of a developmental critical period, the adult cortex retains a significant capacity to undergo functional changes in response to alterations in sensory input. We are interested in the rules that determine how the adult cortex changes its synaptic circuitry to adapt to changes in the pattern of afferent activity. Very well structured and suitable system for such studies is the rat whisker system. Whiskers can be thought of as tiny fingers, with which the animal palpates objects all around. The facial whiskers of rats are aligned in five rows (row A is dorsal and row E is ventral) and the whiskers within a row are numbered from caudal to rostral, like positions on a chessboard (see Fig. 3.6a). Each facial whisker projects via the trigeminal nuclei and "barreloids" in the ventral posterior medial nucleus (VPM) in the thalamus into a separate cluster of neurons in layer IV, called barrels (Jensen and Killackey 1987). Clusters of neurons in layer IV remind the barrels of neurons hence their name. Barrels form the morphological basis of a discrete one whisker-one column organization of the barrel cortex (Fig. 3.6b).

a

b

Fig. 3.6. (a) In the experiment to gain insights into an adult experience-evoked cortical plasticity, all whiskers on one side of the rat's face but two (D2 and D3) were cut close to the fur. Dots mark the positions of cut whiskers aligned in five rows, A to E; (b) Illustration of cortical one whisker-one barrel mapping and the positionof the recording electrode. Clusters of neurons in layer IV remindthe barrels of neurons hencetheir name

3.4 Neocortical Plasticity

67

To alter the pattern of sensory activity, all whiskers except two, D2 and one neighbor in the D row, were cut close to the fur on one side of the face. We can think about this procedure as a removal of all fingers but two (however neither pain nor nerve damage is involved in whisker clipping). After 3, 7-10 and 30 days of "whisker pairing" (rats could use only 2 whiskers for palpation, all other whiskers were re-clipped regularly), the activity of single neurons in barrel D2 was measured in response to controlled deflection of the two paired whiskers, D2 and "D-paired", and the three cut neighbors ("D-cut", C2, and E2). Prof. Ebner and his team recorded and documented progressive and complex changes in D2 barrel cell responses during the course of paired whisker experience (Diamond et aI. 1993, Armstrong-James et aI. 1994). The physiological study outlined above motivated us to develop a computational model of a barrel D2 neuron to gain a deeper insight into which synapses in the cortex modify, how they modify and why they modify (Benuskova et aI. 1994). The weights of the synaptic inputs to the modeled barrel neuron (see Fig. 3.7a) were modifiable according to the Bienenstock, Cooper and Munro (BCM) theory (Bienenstock et aI. 1982, Cooper et aI. 2004). The BCM theory postulates that the neuron possesses a synaptic modification threshold, ~ that changes as a nonlinear function of the time-averaged postsynaptic activity (see Fig. 3.7b). b

a

4»

0: potent iation c

I I

I

,

I

postsynapti c activity c

I I I

I I

TT r

wh iskers C20 1 0 2

n 0 3 E2

c 4> < 0: depression

Fig. 3.7. (a) Schematic illustration of the barrel cortex circuit used in our model. In our simplified circuitry, whiskers C2, DI, D2, D3, and £2 converge through polysynaptic pathways (broken lines) upon a cell in the VPM barreloid D2, and through separate polysynaptic intracortical pathways upon a model cell in the barrel D2. Synaptic weights (m) that were modifiable in our model are denoted by open triangles. (b) Schematic illustration of the BCM synaptic modification rule together with how the same level of postsynaptic activity c' can result in synapse potentiation or depression depending on the current value of Brvt. According to the BCM synaptic modification rule, active excitatory inputs d, > 0 are strengthened when postsynaptic activity c > Brvt and ¢ > O. Active inputs weaken when c < Brvt and ¢< 0

68

3 Neuro-Information Processing in the Brain

Current position of 8M dictates whether a neuron's activity at any given instant will lead to strengthening or weakening of the synapses impinging on it. Whisker pairing was simulated by setting input activities of the model barrel cell to the noise level, except two inputs that represented untrimmed whiskers. Initially low levels of cell activity, resulting from whisker trimming, led to low values of~. As certain synaptic weights potentiated, due to the activity of the paired whiskers, the values of ~ increased and after some time their mean reached an asymptotic value. This saturation of ~ led to the depression of some inputs that were originally potentiated. The changes in cell response generated by the model replicated those observed in in vivo experiments (Benuskova et al. 1994). Previously, the BCM theory has explained salient features of developmental experience-dependent plasticity in kitten visual cortex (Bienenstock et al. 1982, Clothiaux et al. 1991). Our results suggested that the idea of a dynamic synaptic modification threshold, ~, is general enough to explain plasticity in different species, in different sensory systems and at different stages of brain maturity. Another test of BCM theory of synaptic plasticity was its application upon experience-evoked plasticity in the developmentally altered neocortex (Benuskova, Rema et al. 2001). If some synaptic plasticity rule captures real relations in the biological system then it should work even if this system is altered, provided corresponding parameters in the synaptic plasticity equations are changed accordingly. Numerous experimental results indicate that prenatal ethanol exposure impairs the development of synaptic plasticity mechanisms in the rat neocortex: namely, there is a decrease in the number of NMDA receptor-gated ion channels and the development of long, ineffectual dendritic spines (Miller and Dow-Edwards 1988, AIRabiai and Miller 1989, Rema and Ebner 1999). After the offsprings were born, they were fed normally, that is without ethanol, until they had reached adulthood. Then they were tested for ability of their cerebral cortex to undergo plastic changes in response to a new sensory experience. In these experiments again whisker D2 and D3 neighbor, "D-paired", were left intact while all other whiskers were cut. A novel sensory experience had caused plastic neural changes in both, normal and impaired cortices (i.e., cortices that were exposed to ethanol in utero), but these changes differed markedly from each other (Rema and Ebner 1999). As we will show below, the BCM theory can explain both the normal and impaired neocortical plasticity. The original BCM theory is a one cell theory that is one cell in this theory represents a population of cells with the defined properties, and inputs are expressed as instant frequencies of spikes. For instance here, one

3.4 Neocortical Plasticity

69

model cell represents the whole population of excitatory cells in the barrel D2 (Fig. 3.7a). Similarly, the inputs and their weights represent rather particular relay pathways with many synaptic contacts than a one individual synapse. Thus, a change in synaptic weight on the model neuron may represent any kind of plastic changes - i.e. biochemical, and/or morphological. Synaptic plasticity is modeled according to the BCM synaptic modification rule (Cooper 1987). If we consider the case of a linear cell, the modification of the /h synapse with the weight m, at time t is proportional to the product of input activity at the /h synapse, d(t), and a function ¢, in such a way that dm(t) [ lJ = TJ¢ e(t), eM (t)}I, (t) dt

- '-

(3.1)

The "modification rate", 17, is equal to the magnitude of the synaptic modification for the i" input in one time step, when ¢ = 1 and di(t)= 1. Modification function ¢ is a parabolic function of the cell's current firing rate e(t) and modification threshold ~(t), i.e. ¢[c(t),eM(t)] = e(t)[c(t) - eM (t)]

(3.2)

The dynamic synaptic modification threshold ~ is proportional to the squared postsynaptic response averaged over some recent past time t, such that (3.3)

The positive constant a determines how far to the right on the x axis we can place the actual or effective threshold for synaptic potentiation in the equation for the modification function ¢. It will tum out that the value of the constant a is the key to the differences in simulated experience-evoked cortical plasticity between normal and faulty cortex. Cell's current firing rate, e, for simulation of the model cell is calculated as a linear sum of the thalamocortical (VPM) input and intracortical input activities, such that

e(t) = m

vpm (t) Ld?m (t) + Lm~or (t)dt or (t)

(3.4)

where m vpm is the synaptic weight of the thalamocortical input from VPM upon the cell in the barrel-column and mi" is the weight of the lateral intracortical connection from neighboring barrel-column correspond-

70

3 Neuro-Infonnation Processing in the Brain

ing to the whisker i = D2, DI, D3, C2, E2 (see also Fig. 3.7a). The thalamocortical or intracortical input activities d?m(/) ot d," (I), respectively, relayed from the lh whisker, i = D2, DI, D3, C2, E2, are comprised of the sum of the evoked response plus random noise: (3.5) (3.6) In these equations, d(t) is equal to either 1 or 0, depending on whether or pm not the lh whisker is deflected. Parameter 0 < Ir < 1 is the input strength constant of the lh whisker input relayed through VPM to the model barrelcolumn D2. Parameter 0 < I;or < 1 is the intracortical input strength constant of the lh whisker input relayed through its own barrel-column to the model barrel-column D2. Qualitative and quantitative agreement with exm

cor »Irpmfor i = DI,

perimental data was obtained when Ib2 »I?m and I i m,

D3, C2, E2. For D2, and Ie;; '" Ib2 in accordance with experimental data (Armstrong-James et al. 1991). The thalamocortical noise n?m(/) and intracortical noise n~or (I) are defined as random variables which are uniformly distributed in the interval [-Ai(noise),+AiCnoise)], where Ai < ~ is the noise amplitude. They represent the random spontaneous neural activity. Whisker pairing was simulated by setting input activities to the noise levels, except for the two inputs that represented untrimmed whiskers. To match the simulated cell response evolution in time with the experimental testing procedure, the cell's response, c, of the model barrel-column cell was calculated as a sum of the short-latency « lOms) and long-latency (10-100 ms) responses to 50 deflections of the corresponding whisker (Armstrong-James and Callahan 1991, Armstrong-James et al. 1991). The short-latency response, CSL, is the number of spikes generated in response to activation of a multi-whisker thalamocortical input (see Fig. 3.7a). The long-latency response, CLL, is the number of spikes generated in response to activation of intracortical inputs. The total cell response reads (3.7)

At testing intervals, the weight values were fixed and

cor

dr = d;or = 1 for

deflected whisker, otherwise d?m=di = O. Since the response to intracor-

3.4 Neocortical Plasticity

71

tical inputs was measured over the time interval 9 times longer that the short-latency response, we multiply the second term by this number. For the values of simulation parameters and their ranges, as well as other details of the model, we refer the reader to (Benuskova, Rema et aI. 2001). In the following two figures, Fig. 3.8 and 3.9, we show the results of computer simulations of the above described model of neocortical experiencedependent plasticity. a

60

'3 60 E

:l ....40 0 00

.. 30

"-2"',,~

'

~30

-- ,-

II Do

.20 ..."i5.10

b

35

E

i26

-~..... , ........ -.... _.... _, ...... a

2

, ~20 ,,

f ,--r----------. .....

"",

.

............. ~

-

2 f

8,16

.ll.

-

..."t10

i

-::6

"IlL

10

15

20

25

Da", of whl"ker palMoa

30

5

10

15

20

25

30

D8'" of whl"ker pa1rlna

Fig. 3.8. Evolution of the long-latency responses evoked in the real and simulated barrel-column D2 by deflection of paired whiskers, D2 and D-paired (D3). Longlatency responses (10-100 ms poststimulus) are mediated by intracortical pathways. Lines correspond to simulation results, and discrete points with S.E. bars correspond to experimental data averagedover the whole population of cells. Evolution of responses to (a) whisker D2 and (b) paired whisker D3, for the normal cortex (dashed lines and squares) and for the impaired cortex (solid lines and triangles)

Our model based on the BCM theory of synaptic plasticity, explains the plastic changes in the barrel D2 as a result of the modification of excitatory synapses between (i) neighboring cortical barrel-columns, and (ii) between VPM and the model cell representing the barrel D2. In the model of normal barrel cortex: (1) short-latency «10 ms) responses to all whiskers increase, (2) long-latency responses to the paired whiskers, D2 and its spared D row neighbor (D3) initially increase and then decrease. In the model of barrel cortex of adult rats which were prenatally exposed to ethanol (impaired cortex): (1) short-latency «10 ms) responses to all whiskers decrease, (2) long-latency responses to the paired whiskers, D2 and its spared D row neighbor (D3) initially decrease and then increase. Compared to the values of parameters of the model simulating normal cortex, in the model simulating faulty cortex, three factors made a difference: (1) Levels of evoked intracortical activity and levels of noise were much lower than in the model of normal cortex. This is in accordance with

72

3 Neuro-Information Processing in the Brain

experimental findings in (Miller and Dow-Edwards 1988, Rema and Ebner 1999). (2) The thalamocortical noise had to be increased compared to the model of normal plasticity, thus reflecting alterations of thalamocortical projections induced by prenatal ethanol (Granato et al. 1995). (3) The value of the constant a in temporal dependence of 8M upon the past activity of neuron (equation 2.4) must have been twice as large as the value of a in the model of normal cortex (Benuskova, Rema et al. 2001). In spite of that the 8M behaves differently from 8M in the model of a normal cortex and determines a different course of plasticity. a

2S

-= •

~

E20 :;

~

--------------.

",,--------- -

.

S'S

y

..c OJ

.

to

OJ

,

.5

.E

8.10

<$2

..• =as

OJ

..

1

-e

0

OL

to

~

In

0

b

S ;., to "C4

0

S

10

15

OJ

;.

20

2S

Days or whisker pairing

50

0

5

10

15

20

25

Days or whisker pairing

50

Fig. 3.9. (a) Evolution of the short-latency responses evoked in the real and simulated barrel-column D2 by deflection of whisker D2. Short-latency response «10 ms poststimulus) is mediated by thalamocortical pathway. Dashed line corresponds to simulation results, and squares with S.E. bars correspond to experimental data averaged over the whole population of cells in the normal barrel D2. Solid line and triangles belong to the impaired cortex. (b) Illustration of evolution of the mean value of the modification threshold 8M for each day of whisker pairing. Note that in the model of impaired cortex the value of 8M decreases (triangles), whereas in the model of normal cortex 8M increases (squares), compared to the initial value of ~ immediately after whisker trimming

Theoretical and experimental results indicate that the search for the mechanism of 8M should be focused on the excitatory synapses that are formed on dendritic spines with NMDA receptor-gated ion channels (Bear et al. 1987, Bear 1995). According to the physiological model of 8M, the value and sign of ¢ are determined by the Ca2+ movement into dendritic spines. Synaptic efficacy will increase when presynaptic activity evokes a large postsynaptic calcium signal (¢ > 0). This will occur only when the membrane potential exceeds the level required to allow enough ions of Ca2+ to enter through the NMDA receptor-activated channels (c > 8M). When the amplitude of the evoked Ca2+ signal falls below a certain critical

3.4 Neocortical Plasticity

73

level, corresponding to ¢ = 0 and c = 8rvI, then active synapses will be weakened over time. This is in agreement with the fact that the induction of long-term synaptic depression (LTD) requires a minimum level of postsynaptic depolarization and rise in intracellular Ca 2+ concentration in the postsynaptic neuron (Artola et al. 1990). On the other hand, relatively high level of intracellular Ca 2+ concentration in the postsynaptic neuron leads to the induction of long-term synaptic potentiation (LTP). Thus, there is a LTD/LTP crossover point with a critical level of NMDA receptor activation and Ca 2+ entry, for changing the direction of synaptic modification. In the BCM theory, the LTD/LTP crossover point corresponds to 8rvI, or to the point where the ¢ function changes the sign from minus to plus (see e.g. Fig. 3.7b). In our simulations this LTD/L TP crossover point was influenced by the value of the constant a in equation 2.4. It is noteworthy that in the model of faulty cortex, the value of a must have been twice as large as the value of a in the model of normal cortex. This is important especially in the light of experimental finding that in the barrel cortex of adult rats, which were prenatally exposed to ethanol, the number of NMDA receptors is reduced to about one half of the number of NMDA receptors in the normal adult barrel cortex (with respect to all NRl, NR2A, and NR2B subunits) (Rema and Ebner 1999). Thus, we have arrived at hypothesis that the effective position of 8rvI on the x axis may depend on both the number of NMDA receptors and the average past activity. That is, the smaller number ofNMDA receptors is on the postsynaptic neuron the higher is the value of 8rvI, and vice versa. On the other hand, the small average activity lowers the 8rvI value therefore the value of 8rvI is in fact lower in the impaired cortex than in the normal cortex (Fig. 3.9). It is still however too high for extremely low levels of activity in the impaired cortex. Exposing the prenatal alcohol-exposed rats to enriched rearing conditions significantly improves all measured cortical functions but does not restore normal values (Rema and Ebner 1999). The results predict that combinations of interventions will be necessary to completely restore cortical function after exposure of the fetal brain to alcohol. Maybe insertion of the gene for NR2B would help in addition to enriched environment. Relevant studies indicate that not only the number of NMDA receptors may influence 8rvI value, but there are also other factors that are related to the cascade of intracellular events triggered by Ca 2+ influx which can affect 8rvI. We will investigate these factors more thoroughly in the chapter on computational neurogenetic model ofleaming and memory.

74

3 Neuro-Infonnation Processing in the Brain

3.5 Neural Coding: the Brain is Fast, Neurons are Slow Neurons within and between different brain areas send messages to each other by means of output spikes. Thus, neural representations of objects "communicate with each other" since neurons within and between these representations send messages to each other. Within the brain, in the cortex in particular, the principle of topographic mapping of input features is obeyed all over. Thus, the information about the features of stimulus is encoded in the places where neurons responding to object fire. These places get allocated due to interplay between the inborn scaffold and early experience. The principle of topology preservation holds in the somatosensory cortex, auditory and visual cortices. The principle of topology preservation means that features of the object that are close together in the object are close to each other in its neural representation too. For instance, ascending sound frequencies are mapped onto ordered neural stripes in the primary auditory cortex, and bars of similar orientations are mapped onto neighboring columns within the primary visual cortex. Every object, comprised of number of features, is thus represented by a spatial pattern of activated columns within a map. Neurons within one cortical column (there may be 10000 of them) redundantly represent one elementary feature. When activated, neurons are supposed to send a message of their activation and about the salience of the present feature up and down the hierarchy of processing areas, and also within the same processing area. Thus, the messages exchanged by neurons are about the salience of features that they represent. At present the nature of these messages, or the nature of a neural code, is a mystery. Is it the number of spikes sent? Is it the timing of spikes? Is it the frequency of spikes that conveys the message? In this section we will introduce possible solutions to this problem.

3.5.1 Ultra-Fast Visual Classification Simon Thorpe and his team at the University of Toulouse, France, performed an experiment with humans, in which subjects were supposed to classify pictures into two categories, either an animal or a non-animal category (Thorpe et al. 1996). Subjects were not supposed to name an object, they were just asked to classify it. They have not seen the pictures before. For each category they should press one of the two response buttons. The problem was that pictures were shown to subjects for an extremely short period of only 20 milliseconds (1 second = 1000 ms). Exact timings were controlled by a computer. Hundreds of pictures were used for statistical evaluation.

3.5 NeuralCoding: the Brainis Fast, Neurons are Slow

75

Every two seconds, a picture from one or the other class was drawn at random. Each time, it was a different picture, thus there was no repetition. During an experiment the brain activity of subjects was recorded. Reaction time, i.e. time of pressing the button, was equal to 250 ms, on average. Activity in the inferotemporal (IT) cortex occurred on average in 150 ms, so the preparation and execution of motor response took on average 100 ms. After an experiment was over, people used to say that they did not have enough time to realize (to be aware of) what was actually on the picture. Decision to press one or the other button was made on feelings. In spite of that, they correctly classified the pictures in 94% of instances. The same experiment was carried out with monkeys (Thorpe and FabreThorpe 2001). Their average reaction time was 170 ms, IT activity occurred after 100 ms, thus the motor response was prepared in about 70 ms after visual processing. (monkey's brain size is about 1/3 of human size.) Their classification accuracy reached 91%, so they were almost as good as humans were. Experimenters varied an experimental protocol to discover that the ultra-fast classification did not depend on classes of objects, did not depend on color, and did not depend on attention or eye fixation. It is amazing that after such an extremely short presentation of complex stimuli upon eye retina, the primate brain is able to perform a correct classification in more than 90% of instances in a matter of less than 200 ms. 250ms

I

1 Fig. 3.10. Serial processing of the visual stimulus in the image classification experiment with humans. Location of illustrated cortical areas is only schematic. VI = primary visual cortex, V2 = secondary visual cortex, V4 = quartiary visual cortex, IT = inferotemporal cortex, PFC = prefrontal cortex, PMC = premotor cortex, MC = motorcortex

Let us trace the visual brain processing in this experiment (see Fig. 3.10.). Projected image stimulates retina for 20 ms. In about 80 ms, neu-

76

3 Neuro-Information Processing in the Brain

rons in the thalamic LGN (lateral geniculate nucleus) respond. Thalamic neurons activate neurons in the primary visual cortex (V1). Then, activation proceeds to and through higher-order visual areas, V2, V4 and IT. We speak about the so-called "WHAT" visual system, which is assumed to be responsible mainly for classification and recognition of objects. In the highest-order area of this system, i.e. the inferotemporal (IT) cortex, activity appears after 150 ms since the picture onset (on average). It is thought that here, in the IT area, the classification process is completed (Thorpe and Fabre-Thorpe 2001). If we divide 150 ms since the picture onset by the number of processing areas (i.e. retina, thalamus, VI, V2, V4), on average each of them has only 30 ms for processing of signals. The frontal areas, PFC, PMC and MC, are responsible for preparation and execution of motor response, for what they need only 100 ms. Divided by three, again we get about 30 ms for each area. Since each of the mentioned areas has further subtle subdivisions, each sub area can have only 10 ms to process signals and send them higher in the hierarchy of processing. At the same time, neurons in each area send signals up and down in the stream of hierarchical processing. Whether 10 or 30 ms, it is an extremely short time for processing in one single area. Cortical neurons, when naturally stimulated fire with frequencies of the order of 10 to 100 Hz. A neuron firing with an average frequency of 10 Hz (i.e., 10 impulses in 1000 ms), may fire the first spike in 100 ms from the beginning of stimulation. Thus, during the first 10-30 ms there will be no spikes from this neuron. Another neuron firing with the frequency of 100 Hz fires 1-3 spikes during the first 10-30 ms. In each of the above-mentioned areas, there are millions perhaps milliards of neurons, then these neurons exchange only 1-3 spikes, and the result of this processing is sent higher to higher-order areas, and lower to the lower-order areas. Each neuron receives signals from 10000 other neurons and sends off signals to the next 10 000 neurons. Synaptic transmission delay in one synapse is about 1 ms. A neuron cannot wait 10 000 ms to receive signals from all its presynaptic neurons. Thus, the signals ought to come almost simultaneously, and not one after another. Another complication in the neuronal processing of inputs is the fact that firing is a stochastic process. A good model for it is a Poisson stochastic process where the value of dispersion is equal to the value of the mean, thus the dispersion is large. Speaking about firing frequencies of 10 or 100 Hz, we mean average frequencies over relatively long time periods, let us say 500 ms (half of a second). Thus, a neuron firing with the average frequency of 100 Hz does not have to fire a single spike during the first 10-30 ms from the beginning of stimulation, and a neuron firing with the average frequency of 10Hz may fire four spikes. Thus, to summarize, it is

3.5 NeuralCoding: the Brainis Fast,Neurons are Slow

77

really a problem how neurons code information. So far, this problem has not been solved. In the following section we will introduce several current hypotheses. 3.5.2 Hypotheses About a Neural Code

These hypotheses can be divided into two categories: (A) spike timing hypotheses and (B) rate code hypotheses (Maass and Bishop 1999). Coding Based on Spike Timing 1. Reverse correlation. The first option is that the information about the salience of the object feature is encoded in the exact temporal structure of the output spike train. Let us say that two neurons fire three spikes within 30 ms. The first neuron fires a spike train with this temporal structure I II and the second neuron with this temporal structure I[I . By means of the techniques of reverse correlation, it is possible to calculate which stimulus exclusively causes which temporal pattern of which neuron. The main proponents of this theory are Bialek and his coworkers who have made its successful verification in the fly visual system (Rieke et al. 1996). 2. Time to the first spike. Let at time instant to a stimulus arrives to the neural network. Neurons that fire the first (let us say in a window of 10 ms) carry the information about the stimulus features. The rest of neurons and the rest of impulses are ignored. This theory is favored by S. Thorpe (Thorpe et al. 1996, Thorpe and Fabre-Thorpe 2001). 3. Phase. Information about the presence of the feature is encoded in the phase of neuron's impulses with respect to the reference background oscillation. Either they are in a phase lead or in a phase lag. The information can also depend on the magnitude of this phase lead (lag). This coding is preferred by people investigating hippocampus (Jensen 2001). 4. Synchronization. Populations of neurons that represent features belonging to one object can be bound together by synchronous firing. Such synchronization was discovered in the laboratory ofW. Singer in the cat visual cortex to accompany percepts (Fries et al. 1997). It was also detected in the human cortex during perception of meaningful stimuli (faces)(Rodriguez et al. 1999). The Rate Code 1. Temporal average rate. In this respect, works of an English physiologist Adrian from the 30-ties of the 20 th century are being cited. Adrian found

78

3 Neuro-Information Processing in the Brain

out that the average frequency of a neuron in the somatosensory cortex is directly proportional to the pressure applied to its touch receptor. Similar dependencies have been discovered in the auditory and visual cortices. That is, in the auditory cortex, the heard frequency is encoded by the average firing frequency of auditory neurons, and in the visual cortex, the average frequency of neurons encodes for the salience of its visual elementary feature. This coding is still being considered for stationary stimuli that last up to around 500 ms or longer, so that neurons have enough time to count (integrate) impulses over long time. Neurons that have the highest frequency signalize the presence of the relevant feature. 2. Rate as a population average. An average frequency is not calculated as a temporal average but rather as a population average. One feature is represented by a population of many (10 000) neurons, for instance in one cortical column. Upon presence of a feature, most of them are activated. When we calculate the number of spikes in a 10 ms window of all these neurons and divide this number by the number of neurons, we will get approximately the same average frequency as when calculating a temporal average rate of any of these neurons (provided they all fire with the same average rate). This idea has been thoroughly investigated by Shadlen and Newsome (Shadlen and Newsome 1998). They showed on concrete examples, that by means of population averaging we can get a reliable calculation of neuron's average rates even in the case when they have a Poisson-like distribution of output spikes. Populations that relay the highest number of spikes signalize the presence of the relevant feature.

3.6 Summary The problem of coding of neural messages is not new. It was thought, however, that we already have its solution. That this not the case became apparent several years ago, when researches evaluated the temporal intervals of neural processing. Before this, it was thought that in the neural system, the messages are coded by the rate code, that is, by the time average of frequency. The rate code still takes place for the stationary stimuli that last for up to 500 ms or more. Hubel and Wiesel and many others, made their discoveries in many different brain areas, using an idea of the rate code. However, the brain is able to process extremely complex stimuli at much shorter time scales, while the processing takes place in large and hierarchically neural networks comprised of milliards of neurons. Neurons in

3.6 Summary

79

these networks do not have time to calculate average frequencies. Upon seeing all the options for neural coding introduced in the previous section, one gets inevitably confused. Every hypothesis has some experimental support, thus which one is correct and which one is not? Is it possible that different brain areas use different codes? In one area it is a spike timing code and in another area it is a rate code? Or does the brain switch between different codes according to the task to be performed? These possibilities remain to be explored. Or, let us reason that all these hypotheses do not have to be in fact mutually exclusive. For instance, as an exercise, let us imagine that a population of neurons representing one feature has a synchronized firing in response to their feature to be saliently present in the input. Those neurons that have the highest output rates have also the shortest time to the first spike. Their synchronization guarantees that all of them have the same phase difference with respect to the background oscillation. Actually, the background oscillation might have helped to synchronize them. The information about the intensity of a feature in the stimulus (feature salience) is encoded in the average frequency of the whole population of neurons (within 10 ms) and is relayed upon the next population of neurons in the processing hierarchy. And, an actual temporal pattern of spikes relayed by one population of neurons indeed corresponds only to one stimulus. It may be that the mentioned options only different angles under which we can view the same process. Experimental developmental neuroscience brings abundant evidence that in the developing brain there exist strong genetic programs determining the overall pattern of hierarchical organization and connections between brain areas. Nature provides an anatomical and physiological "scaffold" in the sense of a framework that outlines the structures to be formed later. A newborn brain is not a "tabula rasa" however nurture and experience can shape it dramatically. Information needed to specify precise subtle differentiation of neurons and subtle twists of interneuronal connectivity far more surpasses that contained in genetic programs. Genetic programs provide for a vast overproduction of abundant and redundant synaptic connections in the developing brain. Individual differences in experience early in life cause selective pruning of majority of these synapses. Only synapses which mediate a genuine individual experience of an individual remain. Process of experience-dependent synaptic pruning during early stages of brain development may constitute the basis of brain and mind individuality. Early developmental overproduction of redundant synapses lasts only for some time after birth. Time windows, i.e. beginnings and durations, are different for different brain systems. They also differ for different animal species. In

80

3 Neuro-Information Processing in the Brain

general, what lasts weeks and months in rats, cats and monkeys, usually lasts for years in humans. Later in adolescence and adulthood, a new experience is being "burned" into the brain mainly by a selective creation of new synapses and by changing the efficacies of synaptic transmission of existing connections. This does not mean that synapses cannot be removed due to experience later in life, or that new connections cannot be created due to experience later in life. They can, but the prevailing process of experience-dependent adult cortical plasticity is not based on pruning some kind of permanently created abundant connections like in the early development. Some cortical areas retain the capacity of synaptic plasticity for the whole course of life. These are association cortical areas, highest-order sensory areas, premotor and emotional cortical areas of the brain. Capacity of the brain to be plastic and change its microstructure due to experience is of a profound importance for discovering the rules of the mindlbrain relation. Brain plasticity also relates to the genetic code and future modeling of brain functions will require this information.

4 Artificial Neural Networks (ANN)

This chapter introduces the basic principles of artificial neural networks (ANN) as computational models that mimic the brain in its main principles. They have been used so far to model brain functions, along with solving complex problems of classification, prediction, etc. in all areas of science, engineering, technology and business. Here we present a classification scheme of the different types of ANN and some main existing models, namely self-organized maps (SOM), multilayer-perceptrons (MLP) and spiking neural networks (SNN). We illustrate their use to model brain functions, for instance the generation of electrical oscillations measured as LFP. Since ANNs are used as models of brain functions, they become an integral part of CNGM where gene interactions are introduced as part of the structure and the functionality of ANN (see e.g. Chap. 8).

4.1 General Principles ANNs are massively parallel computational systems inspired by biological neural networks. They can have different architectures and properties of their processing elements. Illustration of a general architecture of an ANN is shown in Fig. 4.1.

,

'._.. _.. .. _... _...

Feedlor wa rd connec tions

Feedbac!l;

c o n n ec lio n s

, _~

1

Lateral

c c nn e c nc e s

Fig. 4.1. A general architecture of ANN with three kinds of connectivity: feedfor-

ward, feedback, and lateral. Neurons are usually organized into layers

82

4 Artificial Neural Networks (ANN)

There are different variations derived from this general architecture: for instance there can be only feedforward connections between layers, or there can be only feedback connections in the one and only layer with the feedforward input, or there can be feedback connections allowed only between some layers, etc. In general, ANNs perform mappings of vectors from m-dimensional input space upon vectors from n-dimensional output space. We can also say that ANNs learn how to associate vectors from one space to vectors from another space. First of all, there are many models of ANNs, each of them having different particular properties. However, they all are adaptable systems that can reorganize their internal structure based on their experience in the process called training or learning. ANN are very often being referred to as connectionist systems (Kasabov 1996a, Arbib 2003). Basic processing elements (processing units) of ANNs are formal neurons based on the rate code hypothesis of a neural code according to which the information being sent from one neuron to another is encoded in the output rate of its spike train. Thus the input-output function of an {h element of an ANN is o;(t) =

g( ~

W ik

(t)x k

(4.1)

(t)J

where o.; xi, Wik E 91 are the output rate of the {h neuron, input rate of the J(h input, and Wik is the synaptic weight between the kth and {h element, respectively. The function g is the so-called activation or transfer function of a neuron. It can be linear or nonlinear, continuous and differentiable or binary, depending on which ANN model we are working with. w

b

::~ .~o ~

Fig. 4.2. (a) A simplifieddrawing of a real neuron; (b) a diagram of a simple artificial rate-based neuron model

Different models of ANNs differ with respect to their architecture, transfer functions of their elements and the rules used to modify the

4.1 General Principles

83

weights between neurons in the process of learning (Bishop 1995). A simple model of an artificial neuron and different activation functions are illustrated in Fig. 4.2 and Fig. 4.3, respectively. Different models of ANNs differ with respect to their architecture, transfer functions of their elements and the rules used to modify the weights between neurons in the process of learning (Bishop 1995). A simple model of an artificial neuron and different activation functions are illustrated in Fig. 4.2 and Fig. 4.3, respectively. o

-----------'!~-------+ w x

------------~-------+ ------------~

--7t\=--o

____!__::.;

.

w x

--------~:----

------------

---J-~--~-~~

------71 Fig. 4.3. Different typesof activation functions for artificial neuron models

Most of the known ANN training algorithms are influenced by a concept introduced by Donald Hebb (Hebb 1949). He proposed a model for unsupervised learning in which the synaptic strength (weight) is increased if both the source and the destination neurons become simultaneously activated. It is expressed as: (4.2)

r

where wij(t) is the weight of the connection between the lh and neuron in the network at the moment I, o, and OJ are the output signals of the neurons i and j at the same moment I. The weight wij(t+ 1) is the adjusted weight at the next time moment (1+ 1). Usually some kind of weights normalization is applied after each adjustment to prevent their growth to infinity (Miller and MacKay 1994). In general terms, a connectionist system {S, W, P, F, 1, L} that is defined by its structure S, its connection weights W, its parameter set P, its

84

4 Artificial Neural Networks (ANN)

function F , its goal function J, and a learning procedure L , learns, if the system optimizes at least part of its structure S and its function F when observing events z l, z2, z3, ... from the problem space Z. Through a learning process, the system improves its reaction to the observed events and captures useful information that may be later represented as knowledge . The goal of a learning system is defined as finding the minimum of an objective function J(S) named "the expected risk function" (Amari 1990, Amari and Kasabov 1998). The function J(S) can be represented by a loss function Q(Z,S) and an unknown probability distribution !l(Z). Most of the learning systems optimize a global goal function over a fixed part of the structure of the system. In ANN this part is a set of predefined and fixed number of connection weights, i.e. the set number of elements in the set W. As an optimization procedure, some known statistical methods for global optimization, are applied (Amari and Kasabov 1998), for example the gradient descent method. Final structure S is expected to be globally optimal, i.e. optimal for data drawn from the whole problem space Z. In case of a changing structure S and changing (e.g. growing) part of its connections W, where the input stream of data is continuous and its distribution is unknown, the goal function could be expressed as a sum of local goal functions 1' , each one optimized in a small sub-space of Z'c Z as data is drawn from this sub-space. Something more, while the learning process is taking place, the number of dimensions of the problem space Z may also change over time. The above scenarios would reflect in different models of learning as it is explained next. Let us introduce a general classification scheme of AJ\TNs which will lead to explanation of their properties, capabilities and drawbacks (Kasabov 2003).

4.2 Models of Learning in Connectionist Systems There are many methods for learning that have been developed for connectionist architectures (for a review see (Arbib 2003». It is difficult and quite risky to try to put all the existing methods into a clear classification structure (which should also assume "slots" for new methods) but this is necessary here in order to define the scope for further applications of ANN to buildCNGM. A connectionist classification scheme is explained below. On the one hand, this scheme is a general one as it is valid not only for connectionist

4.2 Models of Leaming in Connectionist Systems

85

learning models, but also for other learning paradigms, for example evolutionary learning, case-based learning, analogy-based learning and reasoning, etc. On the other hand, the scheme is not a comprehensive one as it does not present all existing connectionist learning models. It is only a working classification scheme needed for the purpose of this work. A (connectionist) system that learns from observations zl, z2, z3, ... from a problem space Z can be designed to perform learning in different ways. The classification scheme below outlines main questions and issues and their alternative solutions when constructing a connectionist learning system. Now, let us offer an explanation of individual issues and alternatives (Kasabov 2006). 1. What space has the learning system developed in? a) The learning system has developed in the original problem space Z. The structural elements (nodes) of the connectionist learning system are points in the d-dimensional original data space Z. This is the case in some clustering and prototype learning systems. One of the problems here is that if the original space is high dimensional (e.g. 6,000 genes expressed in the brain, or 64 EEG channels) it is difficult to visualize the structure of the system and observe some important patterns. For this purpose special visualization techniques, such as Principal Component Analysis (PCA), or Sammon mapping, are used to project the system structure S into a visualization space V. b) The learning system has developed in its own machine space M The structural elements (nodes) of the connectionist learning system are created in a system (machine) space M, different from the d-dimensional original data space Z. An example is the Self-Organizing Map (SOM) neural network (Kohonen 1997). SOMs develop in a two-, three-, or more dimensional topological spaces (maps) from the original data.

2. Is the space open? a) An open problem space is characterized by unknown probability distribution P(Z) of the incoming data and a possible change in its dimensionality. Sometimes the dimensionality of the data space may change over time, involving more, or fewer dimensions, for example adding new modalities to a person identification system. b) A closed problem space has a fixed dimensionality, and either a known distribution of the data or the distribution can be approximated in advance through statistical procedures. 3. Is learning on-line?

86

4 ArtificialNeural Networks (ANN)

a) Batch-mode, off-lin e learning In this case a pre-defined learning (training) set of data P={zl, z2, ..., zp} is learned by the system through propagating this data set several times through the system. Each time the system optimizes its structure S based on the average value of the goal function over the whole data set P. Many traditional algorithms, such as the backpropagation algorithm, use this type of learning (Rumelhart et al. 1986, Werbos 1990). b) On-line, pattern mode, incremental learning On-line learning is concerned with learning each data example separately as the system operates (usually in a real time) and the data might exist only for a short time. After observing each data example the system makes changes in its structure (e.g. the W connections) to optimize the goal function J. A typical scenario for on-line learning is when data examples are drawn randomly from a problem space and fed into the system one by one for training. Although there are chances of drawing the same examples twice or several times, this is considered as a special case in contrast to the offline learning when one example is presented to the system many times as part of the training procedure. Methods for on-line learning in ANN are studied in (Albus 1975, Fritzke 1995, Saad 1999). In (Hassibi and Stork 1992) a review of some statistical methods for on-line learning, mainly gradient descent methods applied on fixed size connectionist structures, is presented. Some other types of learning, such as incremental learning and lifelong learning are closely related to on-line learning. Incremental learning is the ability of an ANN to learn new data without fully destroying the patterns learned from old data and without the need to be trained on either old or new data. According to Schaal and Atkeson (Schaal and Atkeson 1998) incremental learning is characterized by the following features: • Input and output distributions of data are not known and these distributions may change over time . • The structure of the learning system W is updated incrementally. • Only a limited memory is available so that data have to be discarded after they have been used. On-line learning, incremental learning, and lifelong learning, are typical adaptive learning methods. Adaptive learning is aiming at solving the wellknown stability/plasticity dilemma, which means that the system is stable enough to retain patterns learned from previously observed data while it is flexible enough to learn new patterns from new incoming data.

4.2 Models of Leamingin Connectionist Systems

87

Adaptive learning is typical for many biological systems and is also useful in engineering applications, such as robotic systems and process control. Significant progress in adaptive learning has been achieved due to the Adaptive Resonance Theory (ART) (Carpenter and Grossberg 1991) and its various models that include unsupervised models (ARTl-3, FuzzyART) and supervised versions (ARTMAP, FuzzyARTMAP-FAM) (Carpenter et al. 1991). c) Combined on-line and off-line learning. In this mode the system may work for some of the time in an on-line mode, after that it switches to off-line mode, etc. This is often used for optimization purposes where a small "window" of data from the continuous input stream can be kept aside, and the learning system, which works in an on-line mode, can be locally or globally optimized through off-line learning on this window of data through "window-based" optimization of the goal function J(W). 4. Is the learning process lifelong? a) Single session learning The learning process happens only once over the whole set P of available data (even it may take many iterations during training). After that the system is set in operation and never trained again. This is the most common learning mode in many existing connectionist methods. b) Lifelong learning Lifelong learning is concerned with the ability of a system to learn from continuously incoming data in a changing environment during its entire existence. Growing, as well as pruning may be involved in the lifelong learning process, as the system needs to restrict its growth while always maintaining a good learning and generalization ability. 5. Is output data available and in what form? The availability of output data that can be used for comparison with what the learning system produces on its outputs defines four types of learning: a) Unsupervised learning There are no desired output data attached to the examples zl, z2, z3, ... The data is considered as coming from an input space Z only. b) Supervised learning There are desired output data attached to the examples zl, z2, z3, ...The data is considered as coming in (x, y) pairs from both an input space X and an output space Y that collectively define the problem space Z. The con-

88

4 Artificial Neural Networks (ANN)

nectionist learning system associates data from the input space X to data from the output space Y. c) Reinforcement learning In this case there are no exact desired output data, but some hints about the "goodness" of the system reaction are available. The system learns and adjusts its structural parameters from these hints. In many robotic systems a robot learns from the feedback from the environment that may be used as a qualitative indication of the correct movement of the robot.

d) Combined learning This is the case when a connectionist system can operate in more than one of the above learning modes. 6. Is evolution of populations of individuals over generations involved in the learning process? a) Individual, development-based learning A system is developed independently and is not part of a development process of a population of individual systems. b) Evolutionary learning, population-based learning over generations Here, learning is concerned with the performance of not only an individual system, but of a population of systems that improve their performance through generations. The best individual system is expected to emerge, to evolve from such populations. Evolutionary computation (EC methods, such as genetic algorithms (GA have been widely used for optimizing ANN structures (Yao 1993, Fogel 1995, Watts and Kasabov 1998). Such ANNs are called evolutionary neural networks. They utilize ideas from Darwinism. Most of the evolutionary computation methods developed so far assumes that the problem space is fixed, i.e. the evolution takes place within a pre-defined problem space and this space does not change dynamically. Therefore, these methods do not allow for modeling real, on-line adaptation. In addition they are very time consuming which also prevents them from being used in real world applications. 7. Is the structure of the learning system of a fixed size or is it evolving? Here we will refer again to the bias/variance dilemma, (see for example (Grossberg 1969,1982, Carpenter and Grossberg 1991)). With respect to an ANN structure, the dilemma states that if the structure is too small, the ANN is biased to certain patterns, and if the ANN structure is too large there are too many variances that may result in over-training, poor generalization, etc. In order to avoid this problem, an ANN structure should

4.2 Models of Leaming in Connectionist Systems

89

change dynamically during the learning process thus better representing the patterns in the data and the changes in the environment. a) Fixed-size structure This type of learning assumes that the size of the structure S is fixed (e.g. number of neurons, number of connections) and through learning the system changes some structural parameters (e.g. W - the values of connection weights). This is the case in many multi-layer perceptron ANNs trained with the backpropagation algorithm (Rosenblatt 1962, Arbib 1972, Rumelhart et al. 1986, Arbib 1987, Amari 1990, Werbos 1990, Amari and Kasabov 1998). b) Dynamically changing structures According to Heskes and Kappen (Heskes and Kappen 1993) there are three different approaches to dynamically changing structures: constructivism, selectivism, and a hybrid approach. Connectionist constructivism is about developing ANN that have a simple initial structure and grow during its operation through inserting new nodes. This theory is supported by biological facts (see (Saad 1999)). The insertion can be controlled by either a similarity measure of input vectors, or by the output error measure, or by both, depending on whether the system performs unsupervised or supervised mode of learning. A measure of difference between an input pattern and already stored ones is used for deciding whether to insert new nodes in the adaptive resonance theory models ARTl and ART2 (Carpenter and Grossberg 1991) for unsupervised learning. There are other methods that insert nodes based on the evaluation of the local error. Such methods are the Growing Cell Structure and Growing Neural Gas (Fritzke 1995). Other methods insert nodes based on a global error to evaluate the performance of the whole ANN. One such method is the Cascade-Correlation Method (Fahlman and Lebiere 1990). Methods that use both similarity and output error for node insertion are used in Fuzzy ARTMAP (Carpenter et al. 1991) and also in EFuNN (Evolving Fuzzy NN) (Kasabov 2003). Connectionist selectivism is concerned with pruning unnecessary connections in an ANN, that starts its learning with many, in most cases redundant, connections (Sankar and Manmone 1993, Rummery and Niranjan 1994). Pruning connections that do not contribute to the performance of the system can be done by using several methods: Optimal-Brain Damage (LeCun et al. 1990), Optimal Brain Surgeon (Hassibi and Stork 1992), Structural Learning with Forgetting (Ishikawa 1996). 8. How do structural modifications affect the partitioning of the problem space?

90

4 Artificial Neural Networks(ANN)

When a connectionist model is created, either in a supervised or in an unsupervised mode, the nodes and the connections divide the problem space Z into segments. Each segment of the input sub-space is mapped onto a segment from the output sub-space in case of a supervised learning. The partitioning in the input sub-space imposed by the model can be one of the following two types: a) Global partitioning (global learning) Learning causes global partitioning of the space. Partitioning hyperplanes can be modified either after every example is presented (in case of on-line learning), or after all examples are presented for one iteration (in case of a batch-mode learning). Through the gradient descent learning algorithm the problem space is partitioned globally. This is one of the reasons why global learning in multilayer perceptrons suffers from the catastrophic forgetting phenomenon (Miller and MacKay 1994, Robins 1996). Catastrophic forgetting (also called unlearning) is the inability of the system to learn new patterns without forgetting previously learned patters. Methods to deal with this problem include rehearsing of the ANN on a selection of past data, or on generated new data points from the problem space (Robins 1996). Other techniques that use global partitioning are Support Vector Machines (SVM (see (Kecman 2001)) for a comparative study of ANN , fuzzy systems and SVM). Through learning, the SVM optimize the positioning of the hyper-planes to achieve maximum distance from all data items on both sides of the plane . b) Local partitioning (local learning) In case of a local learning the structural modifications of the system affect the partitioning of only a small part of the space where the current data example is drawn from. Each subspace is defined by a neuron. The activation of each neuron is defined by local functions imposed on its subspace. As example of such local functions, the kernels K are defined by formula:

K(x)=exp(-x 2 / 2)/ .,j2; while IK(x) = 1 for all x s Z

(4.3)

Other examples of local partitioning are when the space is partitioned by hyper-cubes and fractals in a 3D space. Before creating a model it is important to choose which type of partitioning would be more suitable for the task in hand. In the evolving connectionist systems presented later in this book, the partitioning is local. Local partitioning is easier to adapt in an on-line mode, faster to calculate, and does not cause catastrophic forgetting. 9. What knowledge representation is facilitated in the learning system?

4.2 Models ofLeaming in ConnectionistSystems

91

It is a well-known fact that one of the most important characteristics of the brain is that it can retain and build knowledge. However, it is not known yet how exactly the activities of the neurons in the brain are transferred into knowledge. For the purpose of the discussion in this chapter, knowledge can be defined as the information learned by a system that the system can interpret in different ways and can use in inference procedures to obtain new facts and new knowledge. Traditional ANNs and connectionist systems have been known as poor facilitators of representing and processing knowledge despite of some early investigations (Hinton 1989,1990). However, some of the issues of knowledge representation in connectionist systems have already been addressed in the so-called knowledge based neural networks (KBNN) (Towell and Shavlik 1993, Towell and Shawlik 1994, Cloete and Zurada 2000). KBNN are ANNs that are pre-structured in a way that allows for data and knowledge manipulation, which includes learning, knowledge insertion, knowledge extraction, adaptation and reasoning. KBNN have been developed either as a combination of symbolic AI systems and ANN (Towell et al. 1990), or as a combination of fuzzy logic systems and ANN (Jang 1993, Yamakawa et al. 1993, Furuhashi et al. 1994, Hauptmann and Heesche 1995, Kasabov 1996b). Rule insertion and rule extraction operations are examples of how a KBNN can accommodate existing knowledge along with data, and how it can 'explain' what it has learned. There are different methods for rule extraction that are applied to practical problems (Hayashi 1991, Kasabov 1996c, Duch et al. 1998, Kasabov 1998, Mitra and Hayashi 2000, Kasabov 2001). Generally speaking, learning systems can be distinguished based on the type of knowledge they represent:

a) No explicit knowledge representation isfacilitated in the system An example for such connectionist system is the traditional multilayer perceptron network trained with the backpropagation algorithm (Rosenblatt 1962, Amari 1967, Rumelhart et al. 1986, Arbib 1987, Werbos 1990). b) Memory-based knowledge The system retains examples, patterns, prototypes, cases, for example instance-based learning (Aha et al. 1991), case-based reasoning systems (Mitchell et al. 1997), and exemplar-based reasoning systems (Salzberg 1990). c) Statistical knowledge

92

4 Artificial Neural Networks(ANN)

The system captures conditional probabilities, probability distribution, clusters, correlation, principal components, and other statistical parameters (Bishop 1995). d) Analytical knowledge The system learns an analytical function f X -+ Y, that represents the mapping of the input space X into the output space Y. Regression techniques and kernel regressions in particular, are well established (Haykin 1994, Bishop 1995). e) Symbolic knowledge Through learning, the system associates information with pre-defined symbols. Different types of symbolic knowledge can be facilitated in a learning system as discussed further below. f) Combined knowledge The system facilitates learning of several types of knowledge. g) Meta-knowledge The system learns hierarchical level of knowledge representation where meta-knowledge is also learned, for example, which piece of knowledge is applicable and when. h) "Consciousness" ofa system The system becomes "aware" of what it is, what it can do, and where its position among the rest of the systems in the problem space is.

i) "Creativity" ofa system An ultimate type of knowledge would be such knowledge that allows the system to act creatively, to create scenarios, and possibly to reproduce itself, for example, a system that generates other systems (programs) improves in time based on its performance in the past. 1o. What type of symbolic knowledge is facilitated by the system? If we can represent the knowledge learned in a learning system as symbols, different types of symbolic knowledge can be distinguished: • • • • • • • •

Propositional rules First-order logic rules Fuzzy rules Semantic maps Schemata Meta-rules Finite automata Higher-order logic

4.3 Unsupervised Learning (SelfOrganizing Maps- SOM)

93

11. If the systems' knowledge can be represented as fuzzy rules, what types of fuzzy rules are facilitated by the system? Different types of fuzzy rules can be used, for example: • Zadeh-Mamdani fuzzy rules (Zadeh 1965, Mamdani 1997) • Takagi-Sugeno fuzzy rules (Takagi and Sugeno 1985) • Other types of fuzzy rules, e.g., type-2 fuzzy rules (for a comprehensive reading, see (Mendel 2001). Generally speaking, different types of knowledge can be learned from a process or from an object in different ways, all of them involving the human participation. They include direct learning by humans, simple problem representation as graphs, analytical formulas, using ANN for learning and rule extraction, etc. All these forms can be viewed as alternative and possibly equivalent forms in terms of final results obtained after a reasoning mechanism is applied on them. Elaborating analytical knowledge in a changing environment is a very difficult process involving changing parameters and formulas with the change of the data. If evolving processes are to be learned in a system and also understood by humans, neural networks that are trained in an on-line mode and their structure is interpreted as knowledge are the most promising models at present. 12. Is the learning process active? Humans and animals are selective in terms of processing only important information. They are searching actively for new information (Taylor 1999, Freeman 2000). Similarly, we can have two types of learning in an intelligent system: • Active learning in terms of data selection, filtering and searching for relevant data. • Passive learning - the system accepts all incoming data.

4.3 Unsupervised Learning (Self Organizing Maps - SOM)

4.3.1 The SOM Algorithm

Self organizing maps belong to the unsupervised artificial neural network modeling methods (Kohonen 1984). The model typically projects a high dimensional dataset on to a lower dimensional space. The SOM network consists of two layers: the input and the output layers. The dataset pre-

94

4 Artificial Neural Networks (ANN)

sented to the network is comprised of samples characterized by p descriptors - variables. Each sample is represented by a vector that includes all p descriptors and there are as many sample vectors as samples. The input layer is comprised of p nodes (neurons) (Fig. 4.4). The output layer forms a d-dimensiona1 map, where d < p. In this study, the map is in the form of a rectangular 2D grid with I by m neurons laid out on a hexagona11attice ( C = I x m neurons in the output layer). Each neuron Cj of the output layer, also called a cell, is linked to the neurons i = 1,2,... ,p of the input layer by connections that have weights wij associated with them, forming a vector wij. These weights represent the virtual values for each descriptor in each output neuron such that each output neuron or cell of the output layer Cj stores a virtual vector of connection weights wij. These virtual vectors represent the co-ordinates of centers of groups of similar input vectors, where similarity is measured in terms of Euclidean distance:

D(x,wj)

=

[L(i=l....p)(x;-wijif'

(4.4)

for all neurons (cells) Cj and with x a sample vector. The aim of the SOM algorithm is to organize the distribution of sample vectors in a d- dimensional space (in our case - two dimensional) using their relationship to the virtual vector distribution thus preserving the similarity and the difference between the input vectors. Similar input vectors are allocated to the same virtual vector and the virtual vector changes with the addition of new input vectors to it. The virtual vectors that are neighbors on the map (neighboring neurons) are expected to represent neighboring groups (clusters) of sample vectors; consequently, sample vectors that are dissimilar are expected to be distant from each other on the map.

Fig. 4.4. (a) Self-organizing map architecture. (b) The input layer is linked to the cells of the output layer by connections called weightswhich define the virtual assemble of the input variables. Lateral connections are treated within the neighborhood function

4.3 Unsupervised Learning (Self Organizing Maps - SOM)

95

Two different learning algorithms could be used in a SOM: sequential or batch. The first one is an incremental algorithm that is commonly used but learning is highly dependent on the order of input. The batch algorithm overcomes this drawback. Furthermore, the batch algorithm is significantly faster (Kohonen 1997). The process involves presenting the whole sample vectors as input to the SOM at once. Using a distance measure, the sample vectors are compared to the virtual vectors that have been randomly assigned to the output neurons at the beginning of the algorithm. Each sample vector is assigned to the nearest virtual vector according to the distance results and the virtual vectors are modified to the mean of the sample vectors that are assigned to it. Details about the algorithm can be found in (Kohonen 1984,1990,1997). At the end of the training, an output neuron has been determined for each sample vector such that each sample is then assigned to a neuron or cell of the map. 4.3.2 SOM Output

Sample Distribution

A direct result of the SOM algorithm is a distribution of the samples on the SOM topological map. According to the properties of the algorithm, the samples that are in the same cell are very similar, and similar to those in neighboring cells. They are less similar to samples that are in distant cells. Each cell unit is at this stage, a cluster and the SOM training procedure constitutes a clustering method, clustering the samples in cells and similar cells together. The approximate number of cells in the output layer can be defined using the formula:

C = 5";;;

(4.5)

where C is the number of cells and n is the number of training samples (sample vectors). Clustering Information

Despite that a SOM clusters samples onto the cells of the map it is of interest to define larger clusters by regrouping the neighboring cells that contain similar samples. The definition of larger clusters can be achieved using several methods. A method well-known to experienced SOM users is the unified-matrix (U-matrix) approach (Utlsh and Siemon 1990). The U-

96

4 Artificial Neural Networks (ANN)

matrix displays the distances between the virtual sites and provides a landscape formed by light plains separated by dark ravines. Another method is a classical clustering analysis of SOM output using any of the classical distance measures and linkages, or the K-means method. These methods are applied to the results of the SOM model, or more precisely on the virtual vector for each neuron of the output layer. Visualization of Input Variables

To analyze the contribution of input variables to cluster structures of the trained SOM, each input variable and the connection weight of its associated descriptor calculated for each virtual vector during the training process, can be visualized in grey scale on the SOM map. Remembering that each cell of the map is represented by a virtual vector, and also that each virtual vector is composed of as many weight values as descriptors, it is possible to visualize each descriptor's weight values associated with each neuron or cell of the trained SOM map. A map can be visualized separately for each descriptor. Relationship Between Multiple Descriptors It may be important to investigate the relationship between sets of descriptors (input variables) and try to find meaningful patterns of their values in combination. For example, it might be important to investigate the relationship between biological and environmental variables across the samples. A second set of descriptors for each sample can be introduced into the SOM and trained along with the first set of descriptors. Initially, each descriptor set is submitted to the trained SOM, and then, the mean value of each descriptor in the descriptor sets in each output of the trained SOM is calculated. If a neuron was not occupied by input vectors, the value is replaced with the mean value of neighboring neurons. These mean values assigned on the SOM map can once again be visualized in grey scale and then compared with the map of the samples as well as other descriptor maps.

The Connection Weights

As described previously, the two layers of the SOM network are linked by connections that are called weights. The set of weight values for each output neuron comprises a virtual vector for that neuron. These weights represent the coordinates of each output neuron in a multidimensional space with as many dimensions as descriptors. But in case of binary data, be-

4.3 Unsupervised Learning (Self Organizing Maps - SOM)

97

°

cause the observed or real values are or 1, the virtual values are constrained between and 1. It should be immediately obvious that these values can be used as some sort of measurement, evaluation, gradient, or as an index depending on the content of the data set and the meaning of the descriptors. One interpretation theory is given in the next section.

°

Interpretation by the Fuzzy Set Theory In the classical approach of the set theory, if a subset A of a set E is considered, the characteristic function is XA . XA is a two-valued function taking its values in {O,l} and defined as follows: If

xEA,xAx)=l

(4.6)

If

x Il A,xAx)= 0

(4.7)

However, in some cases the membership of an element to a given subset is imprecise. (Zadeh 1965) proposed fuzzy set theory to explicitly account for this. In this case, the characteristic function is replaced by the membership functionji, wherejA is a real-valued function taking its values in [0,1]. Now the functionjA(x) gives an indication of the degree of truthfulness for x to be member to the fuzzy subset A. 4.3.3 SOM for Brain and Gene Data Clustering SOM can be trained on unlabeled data and can be used to identify groups (clusters) of data samples grouped based on their similarity. Such data can be 64 EEG channel data, for example, and the SOM can identify similar channel vectors, visualizing them in close topological areas on the map. Gene expression data of tens and hundreds of genes can also be clustered, visualized and explored in terms of how similar the gene expression vectors are. This is illustrated in Fig. 4.5, where the brain cancer data is used after the top 12 genes are selected (See Fig. 1.6). In Fig. 4.5, the top left map shows the SOM output derived from 60 samples and 12 input gene expression variables from the eNS data (Pomeroy et al. 2002). In the top right map the class labels are mapped (class survival in the left blob, and class fatal - on right side) and the bottom three maps show the contribution of gene Gl, G3 and G4 respectively. None of them on its own can discriminate correctly the samples and a good discrimination is achieved through their interaction and pattern formation.

98

4 Artificial Neural Networks (ANN)

. ..

00

0:

••

O'

O'

1(1

Fig. 4.5. SOM output derived from 60 samples and 12 input gene expression variables from the eNS data (Pomeroy et al. 2002) (See Fig. 1.6) - the top left map; in the top right map the class labels are mapped (class survival in the left blob, and class fatal - on right side); the bottom three maps show the contribution of gene G 1, G3 and G4 respectively - none of them on its own can discriminate correctly the samples and a good discrimination is achieved through their interaction and pattern formation. The software system ViscoverySOMine was used for this experiment (http://www.somine.info/). See Color Plate 1

4.4 Supervised Learning

4.4.1 Multilayer Perceptron (MLP)

Multilayer perceptrons (MLP) have a feedforward architecture (see Fig. 4.1). Multilayer perceptrons (MLP) trained with a backpropagation algorithm (BP) use a global optimization function in both on-line (pattern mode) training, and in a batch mode training (Rumelhart et al. 1986, Amari 1990, Werbos 1990, Saad 1999). In the on-line, pattern learning mode of the backpropagation algorithm, after each training example is presented to the system and propagated through it, an error is calculated and then all

4.4 Supervised Learning

99

connections are modified in a backward manner. This is one of the reasons for the phenomenon called catastrophic forgetting - if examples are presented only once, the system tends to forget to react properly on previously used examples (Robins 1996). MLP can be trained in an on-line mode, but they have limitations in this respect as they have a fixed structure and the weight optimization is a global one if a gradient descent algorithm is used for this purpose. A very attractive feature of the MLP is that they are universal function approximators (Cybenko 1989, Funihashi 1989, Hornik et al. 1989, Kurkova 1991) even though in some cases they may converge in a local minimum. Some connectionist systems, that include MLP , use local objective (goal) function to optimize the structure during the learning process. In this case when a data pair (x, y) arrives, the system optimizes its functioning always in a local vicinity of x from the input space X, and in the local vicinity of y from the output space Y. In MLP, the activation function of formal neurons is in the shape of sigmoid function , in the so-called RBF networks the neurons have a radial basis activation function, most frequently a Gaussian.

4.4.2 MLP for Brain and Gene Data Classification MLP can be trained on labeled data and can be used to identify the class a new sample belongs to. Such data can be EEG channel, data measuring different states of the brain labeled by categories (class labels) . MLP can be trained on this data and then used to identify for a new data sample, which brain state (class label) it belongs to. Gene expression data of several genes , identifying different categories (class labels) can be used to train a MLP. The trained system can be used to classify a new gene expression vector to one of the pre-defined categories. In (Khan et al. 1998) a multilayer perceptron ANN was used to achieve a classification of 93 % of Ewings sarcomas, 96 % of rhabdomyosarcomas and 100 % of neuroblastomas. From within a set of 6567 genes, 96 genes were used as variables in the classification system. Whether these results would be different using different classification methods needs further exploration.

Example The CNS cancer gene expression data from (Pomeroy et al. 2002) is used here to build a classifier based on MLP. First, 12 features (genes) are se-

100

4 Artificial Neural Networks (ANN)

lected from the whole data set using t-test, rather than using SNR as it was in Fig.1.6. The selected features are shown in Fig. 4.6. Figure 4.7 shows the projection of the CNS data and the selected 12 features into a PCA space that shows that the features can possibly be successfully used for building a classifier. After the features are selected and evaluates, a MLP classifier is built (Fig. 4.8). A MLP that has 12 inputs (12 gene expression variables), 1 output (the class - control or cancer) and 5 hidden nodes is trained on all 60 samples of the gene expression data from the CNS cancer case study data. The error decreases with the number of iterations applied (Fig. 4.8). (The experiments are performed in a software environment NeuCom (www.theneucom.com). For a full validation of the classification accuracy of the method used (here it is MLP) and the features selected (here they are 12 genes) a crossvalidation experiment needs to be done, where a model is trained on part of the data and tested for generalization on the other part (Baldi and Brunak 2001, Kasabov 2006).

..... '_0--. ICNS__ " 3 " - " - 1'-"":-:oA:7"..-,...,.-----------,3 J:

5

I "'-

10.....

I.....,... I"""",odn__

•

3 3

5 3

5 2

5 1

5 I ...

GIl52 G327 G3'8 Gl\oI. G2196 GElS G3;A5 G33lllG2'96 G2996 G3185 G1Q5.l

"''''Iblt,

Fig. 4.6. Selecting the top 12 genes from the case study CNS cancer data (Pomeroy et al. 2002) as it was shown in Fig.1.6, but here using the t-test method. The selected genes compare well with the selected in Fig.1.6 genes. (A proprietary software system SIFTWARE (www.peblnz.com) was used for the purpose). See Color Plate 2

4.4 Supervised Learning

-

...

101

1r:·_=::_=r;I"";:;:I;:;:::::======::;r~======~;;;::::;:====;:==i OJ ... "

II

~

j

t

."

J.

," .." , '

",

J e ••~". '"

fJ .: a.a·. .'.

..-".

,.

Gll52

'" 1 eo ~ eo

f

•

j

Cl

~ 3l

-_.

1 :2 J 4 5

s

D' 0 _ _ - - - . .-

789 11J11 121314

~C~l

....

0-

Fig. 4.7. The discriminative power of the selected genes 12 genes in Fig. 4.6 is evaluated through Principal Component Analysis (PCA) method. It is seen that the first PC has a significant importance in keeping the samples still distant after the PCA transformation. (A proprietary software system SIFTWARE (www.peblnz.com) was used for the purpose of the analysis). See Color Plate 2

1".. ....... 0 .....

lorsl..... T.....L-...~'..

31

es-

CD

...........

13

II

",--------------==='l ~P.­

N......,..hddIooo_~

~ • •--O ~

to

r-;;r-

~v... ~ ~ OwIcur~,--~ ..-..-,-~

-~

Fig. 4.8. MLP that has 12 inputs (12 gene expression variables from Fig. 4.6), 1 output (the class of survivals vs not responding to treatment) and 5 hidden nodes is trained on all 60 samples of the gene expression data from the CNS cancer case study data. The error decreases with the number of iterations applied (altogether 500). (The experiments are performed in a software environment NeuCom (www.theneucom.com). See Color Plate 3

102

4 Artificial Neural Networks (ANN)

4.5 Spiking Neural Networks (SNN) SNN models are more biologically plausible to brain principles than any of the above ANN methods. Spiking model of a neuron - element of the spiking neural network (SNN) communicates with other neurons in the network by means of spikes (Maass and Bishop 1999, Gerstner and Kistler 2002). Neuron i receives input spikes from presynaptic neurons jEf';, where T, is a pool of all neurons presynaptic neurons to neuron i (Fig. 4.9). This is a more biologically realistic model of neuron that is currently used to model various brain functions, for instance the pattern recognition in the visual system (Delorme et al. 1999, Delorme and Thorpe 2001). We will describe the Spike Response Model (SRM) as a representative of spiking neuron models that are all variations of the same theme (Gerstner and Kistler 2002). In SRM, the state of a neuron i is described by the state variable u;(t) that can be interpreted as a total somatic postsynaptic potential (PSP). The value of the state variable u;(t) is the weighted sum of all excitatory and inhibitory synaptic PSPs, Gif (t - t j - ~if)' such that: u;(t) = LLJij&/I-lj-~ij)

(4.8)

jEfj liEF]

where: I', is the pool of neurons presynaptic to neuron i, F; is the set of times f.j < t when presynaptic spikes occurred, and l1ij is an axonal delay between neurons i and j, which increases with Euclidean distance between neurons in the network. The weight of synaptic connection from neuron j to neuron i is denoted by Jij. It takes positive (negative) values for excitatory (inhibitory) connections, respectively. When u;(t) reaches the firing threshold 9;(/) from below, neuron i fires, i.e. emits a spike (see Fig. 4.10). The moment of 9;(t) crossing defines the firing time I; of an output spike.

Fig. 4.9. Spiking model of a neuron sends and receives spikes to and from other neurons in the network, exactly like biological neurons do

4.5 Spiking Neural Networks (SNN)

103

Immediately after firing the output spike at t;, neuron's firing threshold 9;(1) increases k-times and then returns to its initial value .90 in an exponential fashion. In such a way, absolute and relative refractory periods are modeled: 9;(t-tJ = k x90

t-t J exp( ----:;:

(4.9)

where '() is the time constant of the threshold decay. Synaptic PSP evoked on neuron i when a presynaptic neuron) from the pool 1; fires at time tj , is expressed by the positive kernel Bij(t -t j - t:.ij)= Bij(S) such that Bij(s)=A

(exp[__ s ]-exp(-~J] t'decay

(4.10)

T rIse

where t's are time constants of the decay and rise of the double exponential, respectively, and A is the amplitude of PSP. To make the model more biologically realistic, each synapse, be it excitatory or inhibitory one, can have a fast and slow component of its PSP, such that Btype(S) = Atype(exp[_ _ s lj type T decay

]_exp(__ type J]

(4.11)

S

T rise

where type denotes one of the following:fast_excitation,fast_inhibition, slow_excitation, and slow_inhibition, respectively. These types ofPSPs are based on neurobiological data (Destexhe 1998, Deisz 1999, Kleppe and Robinson 1999, White et al. 2000). Thus, in each excitatory and inhibitory synapse, there can be a fast and slow component of PSP, based on different types of postsynaptic receptors (listed in Table 4.1). Table 4.1 represents a relationship between the activity of a neuron and its molecular (protein, gene) basis. This is in the core ofthe CNGM developed later in Chap. 8. As an example of network comprised of spiking neurons we present the architecture that can be used for modeling the generation of cortical local field potential (LFP) (see Fig. 4.11). Model neural network has two layers. The input layer is supposed to represent the thalamus (the main subcortical sensory relay in the brain) and the output layer represents cerebral cortex. Individual model neuron can be based upon the classical Spike Response Model (SRM) (Gerstner and Kistler 2002). The weight of synaptic connection from neuron} to neuron i is denoted by Jij. It takes positive (negative) values for excitatory (inhibitory) connections, respectively. Lateral and input connections have weights that decrease in value with distance from the

l04

4 Artificial Neural Networks (ANN)

center neuron i according to a Gaussian formula while the connections themselves can be established at random (for instance with p = 0.5). For instance, the asynchronous thalamic activity in the awake state of the brain can be simulated by series of random input spikes generated in the input layer neurons. For the state of vigilance, a tonic, low-frequency, nonperiodic and non-bursting firing of thalamocortical input is typical (Beierlein et al. 2002). For simulation of the sleep state we can employ regular oscillatory activity coming out of the input layer, etc. LFP can be defined as an average of all instantaneous membrane potentials, i.e.: I

<1>(t) =

(4.12)

LUi (t) N N

-

H

a

b

u;(I)

9;(1 - 1/)

---

90

I,'

t/

time (ms)

"")~ C

90

time (ms)

Iti

"")r 1~ <,>"

)

time (rn s )

Fig. 4.10. (a) Suprathresholdsummation ofPSPs in the spiking neuron model. Af-

ter each generation of postsynaptic spike there is a rise in the firing threshold that decays back to the resting value between the spikes. (b) Subthreshold summation of PSPs that does not lead to the generation of postsynaptic spike, but still can contribute to the generation of LFP/EEG. (c) PSP is generated after some delay taken by the presynaptic spike to travel from neuronj to neuron i

Spiking neurons can be interconnected into neural networks of arbitrary architecture similarly like the traditional formal neurons. At the same time it has been shown that SNN have the same computational power as traditional ANNs (Maass and Bishop 1999). With spiking neurons, however, new types of computation can be modeled, like coincidence detection, synchronization phenomena, etc. (Konig et al. 1996). Spiking neurons are more easily implemented in the hardware than traditional neurons (Tikovic et al. 2001) and integrated with neuromorphic systems (Smith and Hamilton 1998).

4.6 Summary

105

Table 4.1. Neuronal parameters and their related protein s in our model of SNN. This is used in the CNGM in Chap. 8. Neuron's parameter Pj , i.e. amplitude and time constants of: Fast excitation PSP Slow excitation PSP Fast inhibition PSP Slow inhibition PSP Firing threshold

Relevant protein(s) Pt

*

AMPAR NMDAR GABRA GABRB SCN, KCN, CLCN

*Abbreviations: PSP = postsynaptic potential, AMPAR = (amino- methylisoxazole- propionic acid) AMPA receptor, NMDAR = (N-methyl-D-aspartate acid) NMDA receptor, GABRA = (gamma-aminobutyric acid) GABA A receptor, GABRB = GABA B receptor , SCN = Sodium voltage-gated channel, KCN = kalium (potassium) voltage-gated channel, CLCN = chloride channel. b

Fig. 4.11. (a) ANN model of the thalamocortical (TC) system. (b) The SNN represents the cortex and the input layer the thalamus. About 10-20% of N neurons are inhibitory (filled circles) . The model does not have a feedback from the cortex to the thalamus

4.6 Summary This chapter presented some principles and basic models of ANN and also demonstrated the use of these models on brain, or on gene expression data, but not on brain-and-gene data together. Artificial neural networks are very sophisticated modeling techniques capable of modeling extremely complex functions. Traditional ANNs like MLP suffer from frequent convergence to local minima of the error function. They are also difficult to train and for practical applications require a lot of experimentation with the number of hidden neurons and other parameters. In spite of that, ANNs are applicable in virtually every situation in which a relationship between the predictor variables (independent variables, inputs) and predicted variables

106

4 Artificial Neural Networks (ANN)

(dependent variables, outputs) exists, even when that relationship is very complex and not easy to articulate in the usual terms of correlations or differences between classes. ANNs also keep in check the curse of dimensionality problem that bedevils attempts to model nonlinear functions with large numbers of variables (Bishop 1995). ANNs learn by example. The neural network user gathers representative data, and then invokes training algorithms to automatically learn the structure of the data. All these advantages are carried over to a new class of ANNs called spiking neural networks, or pulse-coupled networks, since they are comprised of neuron models emitting and receiving spikes like the biological neurons. These latter models are nowadays used to model neural functions. However, in spite of growing knowledge about genetic influence upon neural functions, computational models of brain and neural functions lack this important component. Classical ANNs with sigmoid neurons are no longer considered to be faithful models of the brain or neural system in general. Instead, spiking neuron models are currently preferred in this respect. SRM that we described in the previous section is a highly simplified spiking model of a neuron and neglects many aspects of neuronal dynamics. In particular, all postsynaptic potentials are assumed to have the same shape. The form of postsynaptic potentials depends on the location of the synapse on the dendritic tree. Synapses that are located at the distal end of the dendrite are expected to evoke a smaller postsynaptic response at the soma than a synapse that is located directly on the soma. If several inputs occur on the same dendritic branch within a few milliseconds, the first input will cause local changes of the membrane potential that influence the amplitude of the response to the input spikes that arrive slightly later. Such nonlinear interactions between different presynaptic spikes are neglected in the SRM. A purely linear dendrite, on the other hand, can be incorporated into the model. For more detailed models of biological neurons other models may be more suitable, which one can find in General Neural Simulation System (GENESIS) (Bower and Beeman 1998) and NEURON (Carnevale and Hines 2006). However, these more detailed models are computationally very expensive and have many parameters to fit. In spite of that they are widely used to model neural functions especially at a one cell level or at the level of relatively small networks. One of the problems of ANNs used to model brain functions is that they do not contain genes, in spite it is known genes playa crucial role in determining the functioning of brain neural networks in norm and disease (Chin and Moldin 2001). In Chap. 8, we will take modeling of brain functions one step further by incorporating genes and gene interactions as causal forces for neural dynamics.

5 Evolving Connectionist Systems (ECOS)

This chapter extends Chap. 4 and presents another type of ANNs that evolve their structure and functionality over time from incoming data and learn rules in an adaptive mode. They are called ECOS (Kasabov 2002b, Kasabov 2006). ECOS learn local models allocated to clusters of data that can be modified and created in an adaptive mode, incrementally. Several ECOS models are presented along with examples of their use to model brain and gene data.

5.1 Local Learning in ECOS Evolving connectionist systems (ECOS) are modular connectionist-based systems that evolve their structure and functionality in a continuous, selforganized, on-line, adaptive, interactive way from incoming information; they can process both data and knowledge in a supervised and/or unsupervised way (Kasabov 2002b, Kasabov 2006). ECOS learn local models from data through clustering of the data and associating a local output function for each cluster. Clusters of data are created based on similarity between data samples either in the input space (this is the case in some of the ECOS models, e.g. the dynamic neurofuzzy inference system DENFIS (Kasabov and Song 2002), or in both the input space and the output space (this is the case in the EFuNN models (Kasabov 2001)). Samples that have a distance to an existing cluster center (rule node) N ofless than a threshold Rmax (for the EFuNN models it is also needed that the output vectors of these samples are different from the output value of this cluster center in not more than an error tolerance E) are allocated to the same cluster Nc. Samples that do not fit into existing clusters, form new clusters as they arrive in time. Cluster centers are continuously adjusted according to new data samples, and new clusters are created incrementally. The similarity between a sample S = (x,y) and an existing rule node N = (WI, W2) can be measured in different ways, the most popular of them being the normalized Euclidean distance:

108

5 Evolving Connectionist Systems (ECOS)

L(X; _~(i))2 d(S,N) =.:...(i=.:...I,.:... .../7..:...)

(5.1)

_

n where n is the number of the input variables. ECOS learn from data and automatically create a local output function for each cluster, the function being represented in the W2 connection weights, thus creating local models. Each model is represented as a local rule with an antecedent - the cluster area, and a consequent - the output function applied to data in this cluster, e.g.:

IF (data is in cluster Nc)

(5.2)

THEN (the output is calculated with a function Fc) Implementations of the ECOS framework require connectionist models that support these principles. Such model is the Evolving Fuzzy Neural Network (EFuNN).

5.2 Evolving Fuzzy Neural Networks EFuNN In EFuNN the nodes representing membership functions (MF) can be modified during learning. Each input variable is represented here by a group of spatially arranged neurons to represent a fuzzy quantization of this variable. For example, three neurons can be used to represent "small", "medium" and "large" fuzzy values of the variable. Different MFs can be attached to these neurons. New neurons can evolve in this layer if, for a given input vector, the corresponding variable value does not belong to any of the existing MFs to a degree greater than a set threshold. A new fuzzy input neuron, or an input neuron, can be created during the adaptation phase of an EFuNN. An optional short-term memory layer can be used through feedback connections from the rule node layer (see Fig. 5.1). The layer of feedback connections could be used if temporal relationships between input data are to be memorized structurally. The third layer contains rule nodes that evolve through supervised/unsupervised learning. The rule nodes represent prototypes of inputoutput data associations, graphically represented as an association of hyper-spheres from the fuzzy input and fuzzy output spaces. Each rule node r is defined by two vectors of connection weights - WIer) and W2(r), the latter being adjusted through supervised learning based on the output error, and the former being adjusted through unsupervised learning based on a

5.2 Evolving FuzzyNeural Networks EFuNN

109

similarity measure within a local area of the problem space. The fourth layer of neurons represents fuzzy quantization for the output variables, similar to the input fuzzy neurons representation. The fifth layer represents the real values for the output variables.

Fuzzy outputs Wz

Rule (case) layer

~r..r-=f--~~::::--.-........ Fuzzy input

layer

Inputs

XI. XZ, ••• , Xn

Fig. 5.1. An EFuNN architecture with a short term memory and feedback connections (adapted from (Kasabov 2001))

The evolving process can be based on two assumptions, that either no rule nodes exist prior to learning and all of them are created during the evolving process, or there is an initial set of rule nodes that are not connected to the input and output nodes and become connected through the learning process. Each rule node (e.g., r,) represents an association between a hypersphere from the fuzzy input space and a hyper-sphere from the fuzzy output space (see Fig. 5.2), the WI(rj) connection weights representing the coordinates of the centre of the sphere in the fuzzy input space, and the W2(rj) - the coordinates in the fuzzy output space. The radius of an input hypersphere of a rule node is defined as (1 - Sthr), where Sthr is the sensitivity threshold parameter defining the minimum activation of a rule node (e.g., rl) to an input vector (e.g., (Xd 2 , Yd2 ) ) in order for the new input vector to be associated to this rule node. Two pairs of fuzzy input-output data vectors d, = (XdI, Yd l) and d 2 = (Xd 2, Yd2) will be allocated to the first rule node ri if they fall into the rl

110

5 EvolvingConnectionist Systems(ECOS)

input sphere and in the ri output sphere, i.e. the local normalized fuzzy difference between XdI and Xd-, are correspondingly smaller than the radius r and the local normalized fuzzy difference between Yd I and Ydz is smaller than an error threshold (Errthr) .

...........

~

i

I I - Sthr

Fig. 5.2. Each rule created during the evolving process associates a hyper-sphere from the fuzzy input space to a hyper-sphere from the fuzzy output space. Through accommodating new nodes the centre of the rule node moves slightly (adapted from (Kasabov2001))

The local normalized fuzzy difference between two fuzzy membership vectors dIf and d Zj that represent the membership degrees to which two real values di and d: data belong to the pre-defined MF are calculated as D(d Ij , dzj) = I IdIf- dz~ /I (dIf + d zj). For example, if d Ij = [0, 0, 1,0,0,0] and dzj = [0, 1,0,0,0,0], then D(dIj, dzj) = (l + 1) / 2 = 1 which is the maximum value for the local normalized fuzzy difference. If data example d, = (XdI, YdI) where Xd I and YdI are correspondingly the input and the output fuzzy membership degree vectors, and the data example is associated with a rule node rI with a centre rI], then a new data point d: = (Xdz, Ydz), that is within the shaded area as shown in Fig. 5.2, will be associated with this rule node too. Through the process of associating (learning) new data points to a rule node, the centre of this node hyper-sphere is adjusted in the fuzzy input space depending on a learning rate lr, and in the fuzzy output space de-

5.2 Evolving Fuzzy Neural Networks EFuNN

III

pending on a learning rate Ir2' as it is shown in Fig. 5.2 on two data points. The adjustment of the centre rl I to its new position rl 2 can be represented mathematically by the change in the connection weights of the rule node rl l) l) from WI(rl and W2(rl to W I(rI 2) and W2(r j2) as it is presented in the following vector operations:

Wr (1[2) = Wr (ri) + lr; * Ds(XdpXdJ

(5.3)

(5.4) where: Erri Yd», Yd2) = Ds(YdI, Yd2) = Yd, - Yd2 is the signed value rather than the absolute value of difference vector; A I(rl I) is the activation of the rule node rl I for the input vector Xd2. The idea of dynamic creation of new rule nodes over time for a time series data is graphically illustrated in Fig. 5.3. While the connection weights from WI and W2 capture spatial characteristics of the learned data (centers of hyper-spheres), the temporal layer of connection weights W3 from Fig. 5.1 captures temporal dependences between consecutive data examples. If the winning rule node at the moment (t - 1) (to which the input data vector at the moment (t - 1) was associated) was rl = inda, (t - 1), and the winning node at the moment t is ri = indal(t), then a line between the two nodes is established as follows: ( 2)(t-l) +r I 3 * Arl[ ( )(t-l) W31[,r ( 2) (1) -- W31[,r

* Arr ( 2)(1)

(5.5)

t

where: AI(ri ) denotes the activation of a rule node r at a time moment (t); lr, defines the degree to which the EFuNN associates links between rules (clusters, prototypes) that include consecutive data examples (if Ir3 = 0, no temporal associations are learned in an EFuNN). The learned temporal associations can be used to support the activation of rule nodes based on temporal, pattern similarity. Here, temporal dependences are learned through establishing structural links. These dependences can be further investigated and enhanced through synaptic analysis (at the synaptic memory level) rather than through neuronal activation analysis (at the behavioral level). The ratio (spatial - similarity) / (temporal - correlation) can be balanced for different applications through two parameters Ss and Tc such that the activation of a rule node r for a new data example dnew is defined as the following vector operations: (5.6) where: f is the activation function of the rule node r, D(r, dnew) is the normalized fuzzy difference value and r(t-l) is the winning neuron at time moment (t - 1).

112

5 Evolving Connectionist Systems (ECOS)

Several parameters were introduced so far for the purpose of controlling the functioning of an EFuNN. Some more parameters will be introduced later, that will bring the EFuNN parameters to a comparatively large number. In order to achieve a better control of the functioning of an EFuNN structure, the three-level functional hierarchy is used, namely: genetic level, long-term synaptic level, and short-term activation level. output

Input data overtime

Fig. 5.3. The rule nodes in an EFuNN evolve in time depending on the similarity in the inputdata

At the genetic level, all the EFuNN parameters are defined as genes in a chromosome, these are: 1. Structural parameters, e.g., number of inputs, number of MF for each of the inputs, initial type of rule nodes, maximum number of rule nodes, number of MF for the output variables, number of outputs. 2. Functional parameters, e.g., activation functions of the rule nodes and the fuzzy output nodes; mode of rule node activation (,one-of-n', or 'many-of-a') depending on how many activation values of rule nodes are propagated to the next level); learning rates lr., Ir2 and Ir3; sensitivity threshold (Sthr) for the rule layer; error threshold (Errthr) for the output layer; forgetting rate; various pruning strategies and parameters, as explained in the EFuNN algorithm below.

5.3 The Basic EFuNN Algorithm In an EFuNN, a new rule node r; is connected and its input and output connection weights are set. The EFuNN algorithm, to evolve EFuNNs

5.3 The Basic EFuNN Algorithm

113

from incoming examples, is given below as a procedure of consecutive steps (Kasabov 200 I). Vector and matrix operation expressions are used to simplicity of presentation. 1. Initialize an EFuNN structure with maximum number of neurons and no (or zero-value) connections. Initial connections may be set through inserting fuzzy rules in the structure. If initially there are no rule nodes connected to the fuzzy input and fuzzy output neurons, then create the first node r, = I to represent the first example d, and set its input Wj(rn) and output Wz(rn ) connection weight vectors as follows:

Wj(rn )

=

EX; Wz(rn )

=

(5.7)

TE

where TE is the fuzzy output vector for the current fuzzy input vector

EX 2. WHILE DO Enter the current example (Xd; Ydi ) , EX denoting its fuzzy input vector. If new variables appear in this example, which are absent in the previous examples, create new input and / or output nodes with their corresponding membership functions. 3. Find the normalized fuzzy local distance between the fuzzy input vector EX and the already stored patterns (prototypes, exemplars) in the rule (case) nodes, ri, rj = rl, ri. ... , rn , (5.8)

4. Find the activation zt.Ijj) of the rule (case) rj, rj = rj, rz. ... , r.; Here, radial basis, radbas, activation, or a saturated linear one, satlin, can be use, i.e. Al (r j

)

= radbas(D(EX,rj »,or

Al (r j ) = satlin(l - D(EX,rj »

(5.9)

The former may be appropriate for function approximation tasks, while the latter may be preferred for classification tasks. In case of the feedback variant of an EFuNN, the activation is calculated as explained above: Aj(r) = radbas(Ss *D(EX,r)- ti- WJ),or Al(r i ) = satlin(l- Ss* D(EX,rj ) + Tc * WJ )

(5.10)

114

5 Evolving Connectionist Systems (ECOS)

5. Update the pruning parameter value for the rule nodes, e.g. age, average activation as pre-defined in the EFuNN chromosome. 6. Find all case nodes rj with an activation value A 1(rj) above a sensitivity threshold Sthr. 7. If there is no such case node, then using the procedure from step 1. ELSE 8. Find the rule node inda, that has the maximum activation value (e.g., maxa.i. 9. There are two modes: 'one-of-n' and 'many-ofn'. (a) In case of 'one-ofn' EFuNNs, propagate the activation maxa, of the rule node inda, to the fuzzy output neurons. (5.1 1)

(b)

In case of 'many-of-n' mode, the activation values of all rule nodes that are above an activation threshold of Athr are propagated to the next neuronal layer. Find the winning fuzzy output neuron inda- and its activation maxa-.

10. Find the desired winning fuzzy output neuron indt, and its value maxt-. 11. Calculate the fuzzy output error vector: Err = A 2 - TE. 12. IF (inda, is different from indt2) or (D(A 2, TE) > Errthr), ELSE Update: (a) the input, (b) the output, and (c) the temporal connection vectors (if such exist) of the rule node k = inda, as follow: (a)

(b)

(c)

Ds(EX, WI(k)) = EX - WI (k); WI(k) = WI(k) + lr, * (5.12) Ds(EX, WI(k)), where lr, is the learning rate for the first layer; Ds(EX, WI(k)) = EX - WI (k); WI(k) = WI(k) + lr, * (5.13 Ds(EX, WICk)), where tr, is the learning rate for the first layer; Wll, k) = W3(l, k) + lrs * AI(k) * AI(lit- 1J, here I is the (5.14 winning rule neuron at the previous time moment (t-l),

5.3 The Basic EFuNN Algorithm

115

and A I (l)(t-I) is its activation value kept in the short term memory. 13. Prune rule nodes} and their connections that satisfy the following fuzzy pruning rule to a pre-defined level: 14. IF (a rule node rj is OLD) AND (average activation A1airj) is LOW) AND (the density of the neighboring area of neurons is HIGH or MODERATE (i.e. there are other prototypical nodes that overlap with) in the input-output space; this condition apply only for some strategies of inserting rule nodes as explained in a sub-section below) THEN the probability of pruning node (rj) is HIGH. The above pruning rule is fuzzy and it requires that the fuzzy concepts of OLD, HIGH, etc., are defined in advance (as part of the EFuNN's chromosome). As a partial case, a fixed value can be used, e.g. a node is OLD if it has existed during the evolving of an EFuNN from more than 1000 examples. The use of a pruning strategy and the way the values for the pruning parameters are defined depends on the application tasks.

15. Aggregate rule nodes, if necessary, into a smaller number of nodes. 16. END of the while loop and the algorithm. 17. Repeat steps 2 to step 17 for a second presentation of the same input data. With good dynamic characteristics, the EFuNN model is a novel efficient model especially for on-line tasks. The EFuNN model has the following major strong points: • • • • •

Incremental Fast learning (possibly 'one pass') On-line adaptation 'Open'structure Allowing for time and space representation based on biological plausibility • Rule extraction and rule insertion

116

5 Evolving Connectionist Systems(ECOS)

5.4 DENFIS The Dynamic Evolving Neural-Fuzzy Inference Systems (DENFIS) is also based on the ECOS principle and motivated by EFuNNs. DENFIS has an approach similar to EFuNNs especially similar to EFuNNs' m-of-n mode. DENFIS is a kind of dynamic Takagi-Sugeno type fuzzy inference systems (Takagi and Sugeno 1985). An evolving clustering method (ECM) is used in DENFIS models to partition the input space for creating the fuzzy rules. DENFIS evolve through incremental, hybrid (supervised/unsupervised), learning and accommodate new input data, including new features, new classes, etc. through local element tuning. New fuzzy rules are created and updated during the operation of the system. At each time moment the output of DENFIS is calculated through a fuzzy inference system based on mmost activated fuzzy rules which are dynamically selected from the existing fuzzy rule set. As the knowledge, fuzzy rules can be inserted into DENFIS before, or during its learning process and, they can also be extracted during the learning process or after it. The fuzzy rules used in DENFIS are indicated as follows:

R,: if XI is F Il and X2 is F 12 and ... and Xp is F IP,

(5.15)

then y, = b» + blIxI + b12X2 +... + btpxp where "Xi is F,", 1= 1,2, ... m;j = 1,2, ... P, are M x P fuzzy propositions that form m antecedents for m fuzzy rules respectively; xj>j = 1,2, ... , P, are antecedent variables defined over universes of discourse ~,j = 1, 2, ... , P, and Fij, 1= 1,2, ... M;j = 1,2, ... , P are fuzzy sets defined by their fuzzy membership functions !!Fij: X, ~ [0, 1], I = 1, 2, M; j = 1, 2, ... , P. In the consequent parts of fuzzy rules, y" 1= 1,2, m, are the consequent variables defined by linear functions. In DENFIS, Fij are defined by the following Gaussian type membership function GaussianMF = a exp [-

(x _m)2]

(5.16)

20'2

When the model is given an input-output pair (Xi, di ) , it calculates the following output value:

Color Plate 1

...

-

-- - -

· 1-

II

.. .. .. .. ... .. ••

II

"

••

•

...

-

..

I .:J

.

~._-

.---- _....

~ c

~ ~

Fig. 1.6. 12 genes selected as top discriminating genes from the Central Nervous System (CNS) cancer data that discriminates two classes - survivals and not responding to treatment (Pomeroy et al. 2002). The NeuCom software system is used for the analysis (www.theneucom.com) and the method is called "Signal-toNoise ratio"

Fig. 4.5. SOM output derived from 60 samples and 12 input gene expression variables from the CNS data (Pomeroy et al. 2002) (See Fig. 1.6) - the top left map; in the top right map the class labels are mapped (class survival in the left blob, and class fatal - on right side); the bottom three maps show the contribution of gene G1, G3 and G4 respectively - none of them on its own can discriminate correctly the samples and a good discrimination is achieved through their interaction and pattern formation. The software system ViscoverySOMine was used for this experiment (http://www.somine.info/)

Color Plate 2

..

\ . - - . I _ .... \r..-........

.

-

Fig. 4.6. Selecting the top 12 genes from the case study CNS cancer data (Pomeroy et al. 2002) as it was shown in Fig.1.6, but here using the t-test method. The selected genes compare well with the selected in Fig.1.6 genes. (A proprietary software system SIFTWARE (www.peblnz.com) was used for the purpose)

.[rt

• -

_o-

.....

".

..e : ._

x

....

lW, --

...;..;;;.....-

--

Mol

-

-,

"JI rr.!L..J

.!J

,. ............ ,.. v

I

.!.L...J

~

-"-'

.

Fig. 4.7. T e iscnmmatrve power 0 t e se ecte genes genes mig. 4.6 is evaluated through Principal Component Analysis (PCA) method. It is seen that the first PC has a significant importance in keeping the samples still distant after the PCA transformation. (A proprietary software system SIFTWARE (www.peblnz.com) was used for the purpose of the analysis)

Color Plate 3

, ........- - JCiAI

'~ '.

•

II

..

:::J

J Fig. 4.8. MLP that has 12 inputs (12 gene expression variables from Fig. 4.6), 1 output (the class of survivals vs not responding to treatment) and 5 hidden nodes is trained on all 60 samples of the gene expression data from the CNS cancer case study data. The error decreases with the number of iterations applied (altogether 500). (The experiments are performed in a software environment NeuCom (www.theneucom.com) , _ ..=::.

Fig. 5.8. The rule nodes of an evolved ECOS model from data of a person A using 37 EEG channels as input variables, plotted in a 3D PCA space. The circles represent rule nodes allocated for class 1 (auditory stimulus), asterisks - class 2 (visual stimulus), squares - class 3 (AV- auditory and visual stimulus combined) and triangles - class 4 (no stimulus). It can be seen that rule nodes allocated to one stimulus are close in the space, which means that their input vectors are similar

Color Plate 4

_ • -

r-rr-rr-r-

Fig. 5.10. A leave-one-cross validation method is applied to validate an ECF ECOS model on the 60 CNS cancer samples (Pomeroy et al. 2002), where 60 models are created - each one on 59 samples, after one example is taken out, and then the model is validated to classify the taken out example. The average accuracy over all 60 examples is 82%, where 49 samples are classified accurately and 11 incorrectly. Class 1 is the non-responding group (21 samples) and class 2 is the group of survivals (39 samples)

-- -_

or-

-

r--r-

-

--

..,

Fig. 5.11. An ECOS classifier is evolved on the 12 CNS cancer genes from Fig. 4.6. Aggregated (across all clusters) general profiles for each of the two classes are shown. The profiles, that capture the interaction between genes, show that genes 1, 5 and 11 are differently expressed across samples of each class, gene 6 is highly expressed in both classes and the other genes - lowly. This suggests an interesting interaction between some genes that possibly define the outcome of cancer of the CNS. The analysis is performed with the use of a proprietary software system SIFTWARE (www.peblnz.com)

Color Plate 5

a

'0 11

•

a...._

.!l ...J

;: . ,-... - ....----"I

b