Copyright © 1999 Massachusetts Institute of Technology
Patterns of cognitive breakdown after brain damage in humans can often be interpreted in terms of damage to particular components of theories of normal cognition developed within cognitive science. Along with the new methods of functional neuroimaging, neurological impairments of cognition provide us with prime evidence about the organization of cognitive systems in the human brain. Yet neuro-psychologists have long been aware that the relation between a behaviorally manifest cognitive deficit and an underlying cognitive lesion may be complex. As early as the nineteenth century, authors such as John Hughlings-Jackson (1873) cautioned that the brain is a distributed and highly interactive system, such that local damage to one part can unleash new modes of functioning in the remaining parts of the system. As a result, one cannot assume that a patient's behavior following brain damage is the direct result of a simple subtraction of one or more components of the mind, with those that remain functioning normally. More likely, it results from a combination of the subtraction of some components, and changes in the functioning of other components that had previously been influenced by the missing components. At stake in deciding between these two types of account is not only our understanding of cognition in neurological patients but also the inferences we draw from such patients about the organization of the normal cognitive system.
Computational modeling provides a conceptual framework, and concrete tools, for reasoning about the effects of local lesions in distributed, interactive systems such as the brain (Farah 1994). It has proved helpful in understanding a number of different neuropsychological disorders. In the second part of this article, three examples will be presented of computational models that provide alternative interpretations of a neuropsychological disorder, with correspondingly different implications for theories of normal cogni-tion.
Many of the computational models used in neuropsychology are parallel distributed processing (PDP) models (see COGNITIVE MODELING, CONNECTIONIST and NEURAL NETWORKS), which share certain features with what is known of brain function. These brain-like features include the use of distributed representations, the large number of inputs to and outputs from each unit, the modifiable connections between units, the existence of both inhibitory and excitatory connections, summation rules, bounded activations, and thresholds. Of course, there are many important differences between the computation of PDP models and real brains; for example, even the biggest PDP networks are tiny compared to the brain, PDP models have just one kind of "unit," compared to a variety of types of neurons, and just one kind of activation (which can act excitatorily or inhibitorily) rather than a multitude of different neurotransmitters, and so on. Computational architectures other than PDP, which have fewer patent correspondences to real neural computation, have also been used to mediate inferences between the behavioral impairments of brain-damaged patients and theories of normal cognition. The final example to be summarized here is a production system model (see also PRODUCTION SYSTEMS), which sacrifices some explicit resemblances to brain function in the service of making explicit other key aspects of the theory used to explain patient behavior.
Computational models in neuropsychology, like all models in science, are simplifications of reality, with some theory-relevant features and some theory-irrelevant ones. Our models allow us to find out what aspects of behavior, normal and pathological, can be explained by the theory-relevant attributes, that is, those that are shared with real brain function. Of course, some behavior may be explainable only with the incorporation of other features of neuroanatomy and neurophysiology not used in current computational models. But this is not a problem for models that already account well for patient data. In such cases, the only worry is that the model's success might depend on some theory-irrelevant simplification. We must be on the lookout for such cases, but also recognize that it is unlikely that the success of most models will happen to depend critically on their unrealistic features.
In closing, I provide pointers to three concrete examples of computational modeling in neuropsychology. Only the barest outlines can be given here of the questions to which the models are addressed, and the mechanisms by which the models provide answers.
Deep Dyslexia: Interpreting Error Types
Patients with a READING disorder known as "deep DYSLEXIA" make two very different types of reading errors, which have been interpreted as indicating that two functionally distinct lesions are needed to account for the reading errors of these patients. Deep dyslexic patients make semantic errors, that is, errors that bear a semantic similarity to the correct word, such as reading cat as "dog." They also make visual errors, that is, errors that bear a visual (graphemic) similarity to the correct word, such as reading cat as "cot." The fact that both semantic and visual errors are common in deep dyslexia has been taken to imply that deep dyslexic patients have multiple lesions, with one affecting the visual system and another affecting semantic knowledge. However, Hinton and Shallice (1991) showed that a single lesion (removal of units) in an attractor network that has been trained to associate visual patterns with semantic patterns is sufficient to account for these patients' errors. Indeed, they showed that mixtures of error types will be the rule, rather than the exception, when a system normally functions to transform the stimulus representation from one form that has one set of similarity relations (e.g., visual, in which cot and cat are similar) to another form with different similarity relations (e.g., semantic, in which cot and bed are similar).
Covert Face Recognition: Dissociation Without Separate Systems
Prosopagnosia is an impairment of FACE RECOGNITION that can occur relatively independently of impairments in object recognition (Farah, Klein, and Levinson 1995; see OBJECT RECOGNITION, HUMAN NEUROPSYCHOLOGY). Recently it has been observed that some prosopagnosic patients retain a high degree of face recognition ability when tested in certain ways ("covert recognition"), while performing poorly on more conventional tasks ("overt recognition") and professing no conscious awareness of face recognition. This has been taken to imply that recognition and awareness depend on dissociable and distinct brain systems (De Haan, Bauer, and Greve 1992). My colleagues and I were able to account for covert recognition with a network consisting of units representing facial appearance, general information about people, and names, but without any part of the network dedicated to awareness (Farah, O'Reilly, and Vecera 1993). The dissociations between overt and covert recognition observed in three different tasks were simulated by lesioning the visual face representations of the network. Our conclusion was that it is unnecessary to hypothesize separate cognitive components for recognition and awareness of recognition; covert recognition tasks are simply those that can tap the residual knowledge of a damaged visual system.
Frontal Lobe Impairments: Loss of an Executive System, or Working Memory?
Studies of frontal lobe
function in nonhuman primates have overwhelmingly focused on WORKING MEMORY, the capacity to hold information "on-line" for
an interval of seconds or minutes. By contrast, studies of frontal
lobe function in humans have documented a broad array of abilities,
including PLANNING, PROBLEM SOLVING, sequencing,
and inhibiting impulsive responses (Kimberg, D'Esposito,
and Farah 1997). The diversity of abilities affected, and their "high-level" nature, has
led many to infer that the cognitive system contains a supervisory "executive," residing
in the frontal lobes.
With the animal literature in mind, Dan Kimberg and I wondered whether damage to working memory might produce the varied and apparently high-level behavioral impairments associated with frontal lobe damage (Kimberg and Farah 1993). We used a production system architecture because it makes very explicit the process of weighing different sources of information to select an action. We found that damaging working memory resulted in the system failing a variety of frontal-sensitive tasks, and indeed committing the same types of errors as frontal-damaged patients. This could be understood in terms of the decreased influence of working memory on action selection, and the consequently greater contribution of other influences, including priming of recently executed actions and habit. We concluded that the behavior of frontal-damaged patients does not imply the existence of an executive.
De Haan, E. H., R. M. Bauer, and K. W. Greve. (1992). Behavioral and physiological evidence for covert face recognition in a prosopagnosic patient. Cortex 28:77-95.
Farah, M. J. (1994). Neuropsychological inference with an interactive brain: A critique of the "locality assumption." Behavioral and Brain Sciences 17:43-61.
Farah, M. J., K. L. Klein, and K. L. Levinson. (1995). Face perception and within-category discrimination in prosopagnosia. Neuropsychologia 33:661-674.
Farah, M. J., R. C. O'Reilly, and S. P. Vecera. (1993). Dissociated overt and covert recognition as an emergent property of a lesioned neural network. Pychological Review 100:571-588.
Hinton, G. E., and T. Shallice, (1991). Lesioning an attractor network: Investigations of acquired dyslexia. Psychological Review 98:74-95.
Hughlings-Jackson, J. (1873). On the anatomical and physiological localization of movements in the brain. Lancet 1:84-85, 162 - 164, 232 - 234.
Kimberg, D.Y., M. D'Esposito, and M. J. Farah. (1997). Frontal lobes: Cognitive neuropsychological aspects. In T. E. Feinberg and M. J. Farah, Eds., Behavioral Neurology and Neuropsychology. New York: McGraw-Hill.
Kimberg, D. Y., and M. J. Farah. (1993). A unified account of cognitive impairments following frontal lobe damage: The role of working memory in complex, organized behavior. Journal of Experimental Psychology: General 112:411-428.
Copyright © 1999 Massachusetts Institute of Technology