by Michael J. Carter

The phrase "reverse engineering" conjures up images of a team of technical wizards huddled over a competitor's product, seeking an understanding of its key design features that have made its performance so exceptional that it has become the market leader. The essence of successful reverse engineering is to carefully examine the external behavior of the product, probe its innards with appropriate tools (e.g., oscilloscope, logic analyzer), and posit models for its operation that agree with the available observations. Electrical and computer engineers are particularly well-prepared for this sort of "skunk works" role with their broad training in control and communication system engineering, computer architectures, analog and digital electronics, and the basic sciences and mathematics. Yet if I were to suggest that an EE or CE might undertake the reverse engineering of the most complex control and communication system known to us - the human brain - you would probably guffaw at the very notion! My purpose here is to illustrate why an engineer's skills are particularly appropriate to the understanding of brain disorders, and to encourage other EE/CEs to consider study in this fascinating interdisciplinary area.

The neuroscience research community has, of course, been actively engaged in reverse engineering of the brain for more than a century. Remarkable advances in computed tomography, especially PET (positron emission tomography) and functional MRI (magnetic resonance imaging), have yielded new insight into the roles of various brain structures in the performance of tasks by awake human subjects. The majority of brain research has naturally been concerned with the understanding of normal function, but as engineers know well, there is often more to be learned from studying the modes of failure of a complex system. Indeed, the earliest written records of brain research, attributed to an Egyptian physician from 3000 BC, describe the physical symptoms associated with various brain and spinal cord traumas. An important classical technique for revealing the pathways of information flow in the brains of experimental animals is the "lesion study", whereby specific regions of brain tissue are deliberately injured and the corresponding effects on behavior and task performance are noted. Although one cannot ethically perform such studies on human subjects, the misfortunes of nature have provided the basis for a rich body of literature on the behavioral correlates of many neuropathologies. Only very recently has the power of computational modeling of neural circuitry been brought to bear on the understanding of dysfunction in the brain.

Engineers are well acquainted with the use of computational models for the analysis of complex electronic circuit and system behavior, with the most familiar example being the SPICE family of circuit simulators. The appeal of computational models of neural circuitry to the neuroscience community is simple: laboratory neuroscience is slow and painstakingly detailed work, and studies of human subjects are constrained by ethics. With an accurate computational model one can rapidly perform experiments that might take weeks or months using animal studies, and one can investigate the effects of highly localized, specific changes in neural connectivity and neuronal dynamics that would be impossible to replicate in vivo. The crucial issue in the computational neuroscience approach is, of course, accuracy in the representation of real neural circuit behavior by the simulation model.

Accuracy of representation in neural circuits, as in the study of electronic circuits, is innately tied to the scale on which one views the operation of the circuit. On the microscopic end of the modeling spectrum, many neuroscientists make use of simulation tools (e.g. NEURON, developed at Yale, and GENESIS, developed at CalTech) that incorporate the detailed electrical and chemical dynamics of small networks of neurons (perhaps 100-1000 neurons). These simulators generate neural spike train waveforms that can be compared directly with those measured in laboratory experiments (in vitro or in vivo). Yet the complexity of the computations involved in these simulations argues against increasing the network size to the scale of modular structures seen in the brain (100,000 to 1,000,000 neurons). At the other end of the modeling spectrum one finds the artificial neural network models popularized by Rumelhart and McClelland (feedforward networks) and by Hopfield (recurrent networks). These classes of models do not accurately represent the function of individual real neurons, but instead attempt to replicate some of the gross functions of large neuronal networks (e.g., associative memory or motor control).

In the nascent field of neural modeling of brain disorders, there is both a fundamental scientific goal and a clinical goal. The scientific goal is simply to understand normal and abnormal function of the brain, whereas the clinical goal is to use knowledge of the neurological basis of human behavior to facilitate treatment of cognitive and memory disorders. With an improved understanding of modular sub-networks within the brain and their functional roles in determining behavior, and of complex neurotransmitter and neuromodulatory systems, there are exciting prospects for the design of custom drug therapies that subtly modify brain function in confined regions! Having just returned from the first Workshop on Neural Modeling of Cognitive and Brain Disorders held at the University of Maryland, I can attest to the excitement that rippled through the audience comprised mostly of psychiatrists and neuroscientists. Their enthusiasm stemmed from the many encouraging papers that reported on the development of neural models of the brain sub-structures involved in Alzheimer's dementia, epilepsy, aphasia and acquired dyslexia, schizophrenia, and Parkinson's disease. The computational approach to the study of these disorders is invariably the same: one first develops a neural network model that mimics normal task performance (whether in animals or humans), and one then subjects the network to various faults (changes in connectivity or model parameters) that have some plausible basis in experimental data. This approach is one that is near and dear to my heart since I began studying the fault tolerance properties of artificial neural networks upon my arrival at UNH in 1987!

Of the vast array of neurological and psychiatric disorders, those that impair memory are often the most troubling. Many of us have experienced an occasion when memory embarrassingly fails us (e.g., when recalling the birthdate of one's spouse or significant other!), and have laughingly chalked up the unforgivable lapse to "early Alzheimer's" or "senility". Yet for millions of people worldwide who suffer from such disorders, these lapses of memory and more severe forms of impairment are not laughing matters. In addition to the immense emotional burden that such disorders impose upon the patient and their loved ones, there is a very real financial burden that accompanies the affliction. One study of the costs associated with the care of Alzheimer's disease patients, published in the August, 1994 issue of the American Journal of Public Health, estimated that caring for a single patient for the typical four year span between diagnosis and death costs society more than $213,000 -- and that figure excludes directmedical expenses! Nearly two-thirds of the total burden is borne directly by the patient and his or her family because it is not covered by any existing insurance plan. As physicians begin to diagnose Alzheimer's disease much earlier (8-10 years before death), the potential cost to society becomes staggering. On the basis of the estimates made in that study, the cost of caring for Alzheimer's patients ranks it as the third most expensive medical problem in the United States - after heart disease and cancer.

While Alzheimer's disease may be the most familiar of the traumatic brain disorders (nearly 4 million people in the US alone are afflicted with it), there are other disorders that are equally devastating and yet garner much less public attention. Schizophrenia is estimated to afflict 1-2% of the world's population, and is characterized (in part) by hallucinations and delusions that are thought to arise from impaired memory function. With these sobering facts as motivation, and with the blessing of a sabbatical leave during the Spring, 1995 semester, I decided to change the course of my research to investigate neural models of the memory impairments seen in Alzheimer's dementia, schizophrenia, and stroke patients. I had already spent much of the previous two years reading neuropsychology and computational neuroscience books in order to gain a better appreciation of biological neural networks (a process which I have discovered is never-ending!). While my own efforts are too recent to report any significant results, I have embarked on a study of a neural model of the pathogenesis of so-called schizophrenic "positive" symptoms. The adjective "positive" belies the very negative aspect of such symptoms as hallucinations and delusions, but this is the term that the medical community has settled upon! The approach of this research is to model human memory retrieval using an attractor network architecture of associative memory (of which the Hopfield network is the most familiar example), and then subject the model's internal parameters (synaptic weights) and external stimuli (memory cues, or inputs) to perturbations which nominally represent some of the abnormal neuroanatomical or neurophysiological features seen in the brains of schizophrenic patients. The memory retrieval performance under "faulty" conditions is compared to that exhibited under "normal" conditions. Prior studies of this kind by other investigators have shown tantalizing similarity of behavior by the network model and human subjects. In particular, the perturbed network displays biased recall of a few selected memories, regardless of the input memory cue, and may also exhibit spurious attractors which do not coincide with any previously stored memory. These phenomena are said to be reminiscent of fixated hallucinations and delusions symptomatic of schizophrenic patients. In a joint study this summer with recent ECE graduate John Canfield and current ECE senior Amy Bonsall, I'm investigating an alternative model for the pathogenesis of schizophrenic positive symptoms that builds upon recent work by others, but uses a modified attractor network with more biologically plausible features. We hope to show that the memory impairments reported by earlier investigators can be generated by a much different neurological mechanism.

As encouraging as the initial results might be from the computational modeling of cognitive and brain disorders, there is still a significant gap to be bridged between the microscale models (e.g. GENESIS) and the macroscale models (artificial neural networks). This gap suggests an opportunity for motivated EE/CEs in this fascinating branch of neuroscience. Our profession has long struggled with circuit simulation as the complexity of circuits grew by orders of magnitude every few years. As the physical size of the electronic devices consequently shrank, it became necessary to incorporate subtle nonlinear operational effects in order to accurately predict overall circuit behavior. I contend that the EE community has already faced many of the issues with which the neuroscience modeling community is now confronted. The gap between microscale simulation and macroscale simulation in electronic circuits, which I refer to as mesoscale simulation, is now being bridged by simulators that take advantage of so-called "behavioral models" of small sub-circuits. These models replicate the dynamical behavior of the sub-circuit as observed at its external terminals, but a behavioral model doesn't specifically simulate the operation of each individual circuit element within the sub-circuit. The field of large-scale analog or mixed mode circuit simulation is constantly in search of low-order behavioral models that accurately depict the operation of fairly complex nonlinear circuits (e.g., a phase-locked loop). I believe that a similar trend will develop in the neural modeling of brain disorders. As more accurate microscale simulations of small networks of neurons are developed, there will be a concurrent effort to find reduced complexity models that exhibit nearly the same "terminal" behavior. This will enable macroscale modelers to improve the biological realism of their large-scale networks that reflect the interaction of many widely separated brain regions.

My final observation is that reverse engineering of brain disorders is something that should appeal to many EE/CEs, not just the author! As one becomes better acquainted with the literature of computational neuroscience, one truly appreciates the fundamental courses in differential equations, circuit and system analysis, control and communication systems, and yes, EVEN EMAG (!), that make up the core of one's undergraduate education. With the appropriate decoding of the medical and neuroscience jargon, it's possible to apply one's engineering intuition and some pretty heavy duty mathematics to neural modeling. Reverse engineering of brain disorders is not altogether different from the reverse engineering of a competitor's product that often takes place in industry. One observes the operation of the product, perhaps dissecting it and examining its internal constituents, and if possible, one stimulates internal nodes or varies control parameters while watching the product perform. With a little knowledge of the constraints that present day technology imposes on one's competitors, one can usually succeed in figuring out just how the product operates. While one is ethically constrained from opening up the skull of one's office mate (even though one might be sorely tempted on some occasions!) to explore his or her brain function, the reverse engineering approach to understanding cognitive and memory disorders is beginning to show real promise. I hope to share with you some of the fruits of this new line of research in a future issue of Signals and Noise.

(For the interested reader, here are a few highly recommended books that are accessible to the non-neuroscientist with an interest in brain research:

Neuronal Man : The Biology of Mind by Jean-Pierre Changeux, New York: Oxford University Press, 1985.

The Cerebral Symphony: Seashore Reflections on the Structure of Consciousness by William H. Calvin, New York: Bantam Books, 1990.

Conversations with Neil's Brain: The Neural Nature of Thought and Language by William H. Calvin and George A. Ojemann, Reading, MA: William Patrick/Addison-Wesley, 1994.

For the truly voracious reader, I suggest the following excellent reference text (which may also serve admirably as a boat anchor!):

The Cognitive Neurosciences (Michael S. Gazzaniga, editor), Cambridge, MA:

Bradford Books/MIT Press, 1995.