Brain, Vol. 122, No. 12, 2413-2416,
December 1999
© 1999 Oxford University Press
Book reviews |
FUNDAMENTALS OF NEURAL NETWORK MODELING: NEUROPSYCHOLOGY AND COGNITIVE NEUROSCIENCE.
.
Physiology Unit, School of Biosciences, University of Wales, Cardiff, UK
Attempts to model human mental activity have sometimes involved abstract information-processing schema. Alternatively, the structural features of brain regions believed to be associated with neuropsychological test performance have been considered without an appreciation of the contributions of multiple brain areas. This book emphasizes the shift towards the development of models of cognitive processes which are better-founded in an understanding of neural circuitry. The volume is the twelfth in a series in Computational Neuroscience, which began with Methods in Neuronal Modeling: From Synapses to Networks (1989), edited by Christof Koch and Idan Segev. Computer simulations of neurons and neural networks are now properly regarded as complementary to traditional techniques in neuroscience. Although some of the first applications were to industrial problems in the area of pattern recognition or signal processing, their potential for improving understanding of brain and cognitive processes soon became clear. In the 1990s, substantial improvements in computer architecture, software design and computing speed occurred, coupled with an increased interest in applying the methodology in neuropsychology and in studies of mental dysfunction. The feasibility of making connections between models and clinical data has also increased. The editors are persuasive on the need to provide a textbook that brings together much recent work in this area. This book considers models of many widely used neuropsychological tests and tasks, including the Wisconsin Card Sorting, Stroop, verbal fluency, Tower of Hanoi and Line Cancellation tests. It also covers a range of syndromes, such as Alzheimer's and Parkinson's disease, schizophrenia, epilepsy, alcoholism, stroke, attention deficit/hyperactivity disorder and frontal lobe disorders.
The book has 17 chapters from 36 contributors. It is divided into four sections: Introduction to neural networks, Behavioural states, Neuropsychological tests and clinical syndromes, and Applications in dementia. Each chapter is followed by a full list of references and an index is provided at the end of the book.
Parks, Levine and Long provide an opening chapter that gives a satisfyingly broad and erudite introduction to neural network modelling. The basic architecture of neural networks is described and some of the assumptions underlying this approach are discussed. Parks, Levine and Long also contribute as co-authors to three other chapters in the book.
The term `neural network' refers to a large class of models sharing certain architectural and processing features. Such models contain many simple processing units operating in parallel with connections whose strength or weight can be varied; they learn by adjusting their weights as they gain experience. Thus, these are connectionist or parallel-distributed processing models. An important characteristic is that activation of each unit changes constantly in response to the activity of the other units to which it is connected. The massively parallel nature of neural networks corresponds to an important aspect of brain organization and processing. Although an individual neuron may be a slow processing device, large numbers of neurons operating in parallel can accomplish a task very quickly. These models discover a set of connection weights that capture the internal structure of a domain. Connection weights are typically found by means of error-driven adjustments in the connections between units. Once a neural network has discovered a set of weights that results in appropriate behaviour, the model can repeat a pattern of activity at some later time or produce this pattern when given some portion as a cue. It can show spontaneous generalization. A novel item can be classified in the network's existing representational scheme when its features activate representations that show some similar features. These systems can `programme' themselves to perform extraordinarily complex tasks. The characteristics of neural networks make them an attractive alternative to traditional symbol-processing models. Distributed representations are appropriate for modelling events underlying the mental rotation of objects, pattern recognition, motor control and associative learning. The models described in various chapters are not dominated by one `school of modelling', but include a variety of modelling paradigms such as backpropagation and adaptive resonance, and a systems approach to interacting brain regions implicated in cognitive tasks. However, neural network modellers still face substantial problems in developing models that are biologically plausible. A network simulation of any significant size can discover a solution that produces the desired output from the specified input, but it is likely that this solution is one of an infinitely large number of solutions to the problem. Does the particular solution have theoretical implications for the biological system modelled? Given the difficulty in understanding how a complex network accomplishes a particular task, it has been argued by McClosky that neural networks should not be viewed as simulations of cognitive theories, but as tools for theory development. Networks might be thought of as analogous to animal models of cognitive activities. The animal system is an object of study rather than a simulation of a cognitive theory. The goal is to develop from the animal system a theory of functioning that can be extended to man, albeit after considerable modification. The two systems may share enough crucial features such that systematic analysis of the animal system aids in the development of a theory of the human system. Similarly, neural networks can be studied to develop a theory of their structure and functioning, which can be applied to the human brain. Clearly, a neural network can be more easily manipulated in ways that a human, or indeed an animal, system cannot in order to examine the effects of different types of damage.
The chapter by Ashford, Coburn and Fuster in Part I provides a general account of the evolution of function in the brain and considers the value of primate models of cognitive processes. The readier exploration of the structural substrate of information processing possible in the non-human primate, using electrophysiological recording and brain imaging, is emphasized. Posner and Badgaiyan discuss the attentional system of the brain in relation to brain networks. They describe recent work involving neuroimaging and recording of event-related potentials. Banquet et al. contribute a long and well-argued essay on `A neural network model of memory, amnesia and cortico-hippocampal interactions'. They review the evidence for long-term, declarative memory consolidation by the hippocampal system (hippocampus plus entorhinal cortex, etc.) attributed to long-term potentiation. A functional model of hippocampal–cortical relations is proposed to handle the spectrum of phases of memory covering an increasing time span. These range from short-term memory (STM), intermediate transient memory (ITM) to long-term memory (LTM). ITM is considered to be an automatic support mechanism for working memory (WM). The chapter might usefully have included a list of these and the many other abbreviations employed. The authors use arguments from neuropsychology and brain imaging to support the view that the hippocampal system is involved in every phase of information processing from the short-term to the transient long-term. They have designed a neural network model for temporal order and probability coding with the capacity to learn. The neocortex they argue, is endowed with dual short-term and long-term capacities relying on different functional aspects of the same neural networks. Fast-transient learning might plausibly occur at the synaptic interface between hippocampus and cortex. The hippocampus has a full array of continuously varying memory ranges, coupled with loop systems either within the hippocampus or between hippocampus and cortex. Intrahippocampal loops may be devoted to the iteration of single-event representations. WM has the processing capacity critical to the selection, elaboration and organization of information flow in order to decide what should be permanently stored in LTM. Transient LTM is viewed as the intermediate step between selecting information and permanent store through long-term potentiation. However, both processes are likely to be based on reactivation or re-enactment of cortical patterns by synchronous activation in the hippocampus of recently co-activated neuronal assemblies
The chapters in Part II describe neural network models of behavioural states, such as alcohol dependence, learned helplessness/depression, and waking and sleeping. Ownby presents a computational model of alcohol dependence, which attempts to simulate genetic differences in alcohol preference in certain strains of rats. Increasing inherent dopaminergic activity in the network, for instance, could alter predicted alcohol preference. It is acknowledged that the usefulness of such models depends upon their ability to generate testable hypotheses about modes of treatment. The chapter on a computer simulation of learned helplessness and depression is perhaps less persuasive about the predictive value of network schemes. Sutton and Hobson consider how waking and sleeping states can be addressed through three types of model. A first category of model, the chronological, does not incorporate anatomical or neurophysiological detail, but deals with circadian and other biological rhythms, capturing features such as the coupled oscillations between endocrine, sleep–wake and seasonal cycles. A second category of model deals with circadian control and modulation of detailed neurobiological systems, focusing on individual neurons and small networks. These cellular network models are helpful in understanding brainstem and thalamocortical interactions. The third category uses large-scale network models, which emphasize co-operative effects among many neuron-like units. Such models can examine the effects of sleep on cognitive processes. In Sutton and Hobson's own distributed network model, temporal sequences of cognitive information are generated and aminergic and cholinergic modulation is simulated. The change from the waking state to REM sleep is believed to be associated with reduced activity in 5-HT and noradrenergic systems, but increased activity in cholinergic systems and phasic bursting in brainstem neurons which generate ponto-geniculo-occipital waves. Results support the notion that different neurophysiological states mediate different types of sequencing and learning operations.
The five chapters of Part III consider neuropsychological tests and clinical syndromes. Servan-Schreiber and (Jonathan) Cohen explore the ability of connectionist models to explain aspects of schizophrenia, relating them to a single functional deficit. This is modelled as a reduction of dopaminergic activity in the prefrontal cortex and a subsequent reduction in responsiveness of neurons to other synaptic inputs (a reduction in the gain of the system). This type of model suggests actions for neuroleptic drugs, commonly used to treat schizophrenia, which depend less on blockade of dopamine D2 receptors and more on other pharmacological actions. Cardoso and Parks consider neural network models of executive functioning and a putative role for the frontal lobes in tackling the Tower of Hanoi neuropsychological test. They emphasize the value of approaches which integrate brain imaging data with neural network modelling. Dehaene, (Laurent) Cohen and Changeux discuss the utility of several types of network model when considering number processing and calculation. They are careful to relate possible anatomical aspects of their models to the observed selective deficits in number processing in cases of focal brain lesions. They remind us that `circuits of circuits' (neuronal assemblies) may be regarded as undergoing a kind of internal evolution, during which the most appropriate representations of aspects of the external world are gradually retained at the expense of others, based upon stabilized synapses and changes in circuitry. From this viewpoint, development and learning are given a `mental Darwinist' perspective. Tranel, Damasio and Damasio consider retrieval of words in relation to damage to higher-order cortices outside the classic language areas. They present evidence for the separation of neural systems supporting retrieval of concepts and words for different categories of object (e.g. particular people, animals, fruits/vegetables and tools/utensils) and a parallel regionalization of the underlying anatomical systems.
Part IV contains five chapters on applications of modelling in dementia, four of these referring to Alzheimer's disease. The chapter by Hasselmo, Wyble and Stern takes the reader back to the role of the hippocampal formation in memory. The authors explore the properties of models simulating its cellular physiology and combining hypotheses about individual hippocampal subregions. Their network is capable of learning and recalling episodic memories. They describe simulations that address the type of function required for paired-associate learning and free recall in tests of human memory. The importance of a cellular physiology-based model is emphasized, since it allows the effects of scopolamine (hyoscine) and diazepam to be realistically simulated. Mahurin discusses the basal ganglia and their involvement in closed loops with certain cortical structures. Separate circuits have been broadly designated motor including oculomotor, cognitive or limbic. The basal ganglia play critical roles in a range of functions, including motor control and planning, attention and emotion. These circuits are under powerful modulation by dopamine-releasing pathways. Artificial networks have proved to be valuable tools for modelling interactions between cortical and subcortical structures vital to the production of controlled motor behaviour. An appropriate backpropagation network can model elements of a self-correcting motor system, such as simple ballistic movements and repetitive movement at a single joint. It can be manipulated to mimic the depletion of dopamine that occurs in Parkinson's disease. Such networks can also model self-organizing motor maps and display the ability to learn. Schizophrenia and depression are discussed as examples of neuropsychiatric disorders associated with basal ganglia dysfunction. Aspects of these and other disorders, such as Huntington's disease, Tourette's syndrome, hyperactivity disorder and obsessive-compulsive disorder can also be explored with these network systems. Further consideration of the way deficits associated with Alzheimer's disease can be represented by network models is presented by Parks and Levine, modelling the effects on the Wisconsin Card Sorting test and on verbal fluency tests, and by Chen, Salmon and Butters, who review semantic abnormalities. In the final chapter, Tippett and Farah demonstrate that parallel distributed processing networks can make useful predictions about the kind of damage which causes impaired neuropsychological performance. Importantly, they show that in an interactive system, a semantic impairment alone can account for patterns of performance which, at first glance, appear to indicate a variety of cognitive deficits. Sensitivity to visual factors in tasks or the production of visual errors, for instance, does not necessarily imply damage to visual processing areas in highly interactive systems.
The book will be of considerable interest to neuropsychologists and those in related disciplines for its excellent coverage of a number of areas. Prerequisites of a technical nature are minimized. Deliberately, little emphasis is placed on the mathematics involved in the network models. In most cases, the theoretical basis of model structure can be discerned from the diagrams found within the text. The intention is to encourage clinicians to begin their own exploration of computational neuropsychology, perhaps as part of a team consisting of a mathematician or computer scientist, fluent in programming techniques, and a neuropsychologist. The book will prove a valuable tool for anyone interested in neural network modelling of complex neuropsychological tests. It should greatly assist anyone seeking to integrate quantitative test data, neuroimaging data and databases of neuropathological conditions within a theoretical framework. An integrated approach is required to ensure that as close a correspondence as possible is achieved between network models, brain structure and neuropsychological processes. This will increase the `realism' of network models and improve their predictive capability.
Notes
By Randolph W. Parks, Daniel S. Levine and Debra L. Long. 1999. Pp. 428. Cambridge, MA: The MIT Press.Price £41.95. ISBN 0-262-16175-3.
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||