Fundamentals of Neural Network Modeling: Neuropsychology and Cognitive Neuroscience

When the first report on neural network modeling of psychiatric symptoms, authored by Ralph Hoffmann and published in Archives of General Psychiatry, appeared on the scientific map in 1987 (1), people must have wondered what this type of reasoning was all about. After all, there are roughly 20 billion neurons in the human cortex, and any attempt to really understand their function and link it to psychopathology must have appeared deeply misguided. The ensuing decade, however, has witnessed nothing short of a revolution in brain science and an ever-increasing stream of research on brain function. While we learn more than ever before about neurons at the cellular level, thinking in psychiatry has undergone a major change toward so-called descriptive psychopathology. The concepts have been refined and sharpened and at the same time cleansed of useless theoretical baggage. The net result is that within this framework of clearly defined diagnostic criteria and biological brain science, research in psychiatry has flourished. The problem, however, of how to relate brain and mind—a psychiatric symptom, for example, on the one hand and a dopamine receptor subtype on the other—remained and became increasingly bothersome. This is where neural networks enter the picture.

According to Parks, Levine, and Long, as well as many of the contributors to this edited volume, neural networks are needed to bridge the gap between neurobiology and psychology, including psychopathology. The book starts with a general introduction to network modeling, but subsequent chapters take the term “neural network” to denote just any connected brain areas. So we learn—in well-written chapters by well-known authors—about neuroanatomy (chapter 2), attention networks (chapter 3), lexical retrieval (chapter 12), and semantic abnormalities in patients with Alzheimer’s disease (chapter 16) without learning anything about neural network models. But there are also role-model chapters on simulations and real-world data, which show what this type of research can do for the understanding of such diverse phenomena as mathematics and acalculia (chapter 10), cognition and schizophrenia (chapter 8), and hippocampal function and memory (chapter 13). Within these and other chapters, the reader can find valuable ideas and discussions regarding neural network modeling. Such models, for example, rarely prove any hypotheses, but they are great in generating them. They are the “theoretical concepts that are intermediate between the details of neuroscientific observations and the box-and-arrow diagrams of traditional information processing or neuropsychological theories,” as Servan-Schreiber and Cohen (p. 191) point out, and they may in some cases “be viewed as an existence proof that such a learning mechanism is feasible” (Dehaene et al., p. 246). The limitations of some models are explicitly discussed in some chapters (for example, p. 267), while other chapters explore their applications to psychology.

A minor criticism applies to the way the essays are arranged in sections. Why are alcoholism and depression discussed in the section titled Behavioral States, whereas mathematics, attention, and lexical retrieval are to be found under Neuropsychological Tests and Clinical Syndromes? Parkinson’s disease is not dementia, and, in general, why discuss dementia separately in several chapters and leave out autism, obsessive-compulsive disorder, and attention deficit disorder as well as hallucinations, delusions, and thought disorder? Why are we advised to learn back-propagation and/or adaptive resonance models (p. 25), but self-organizing feature maps (Kohonen networks) are hardly mentioned, although they provide the basic means to understand the formation and change of cortical maps?

Network models are a theoretical tool for the investigation of the behavior of idealized neurons. If we need these models (and the editors argue we do), then we need textbooks to educate the psychiatrist about modeling (and the editors claim they have provided us with one) and to help us see the link between models and the real world of psychopathology. There must be consequences of modeling psychopathology for clinical practice. In addressing these issues, what does the book tell us about the field of neural network research in psychiatry, and what are the advances made in the field within the past decade?

Sadly enough, the first thing to notice is that there really is no field. Neural network research in psychiatry has never taken off the ground to become a mainstream enterprise. Although there remains hardly a symptom, syndrome, or psychiatric disorder for which there is no computational model, network modeling has remained a sacred trade performed by a handful (or two) of devoted psychiatrists. Just seven out of the 37 contributing authors (i.e., less than 20%) in this book work within a psychiatric department; add a few more and the world’s psychiatric modeling community is assembled.

Does the book provide a means to change just that? The authors have made a brave attempt, but more needs to be done. The book comprises 17 chapters of different scope, quality, and level of sophistication, ranging from neuropsychology and cognitive neuroscience to some aspects of modeling. The very term “neural network” is used technically as well as colloquially (denoting any connected brain areas), and there is little, if any, common theme that links the chapters. Although the editors provide an introductory chapter to network modeling, they fail to provide the common thread that would make this book a whole that is more than the sum of its parts. So the book is likely to leave the clinician with questions about the consequences for the real world and has nothing to say about how network models relate to the diagnosis and treatment of mental disorders. More work in this regard is needed, and the book may help to encourage the reader to get involved.