|
Devlin, J.T., Gonnerman, L.M., Anderson, E.S., & Seidenberg, M.S. (1998). Category-specific semantic deficits in focal and widespread brain damage: A computational account. Journal of Cognitive Neuroscience, 10, 77-94. Purpose Category-specific semantic deficits have been well documented in various types of neurologically damaged subjects including those with Alzheimer’s disease (AD), cerebrovascular accidents, and herpes simplex encephalitis. Research initially indicated that artifacts and natural kinds could be preferentially affected by brain damage and were dissociable. The distinction between artifacts and natural kinds, though, has been recast as one between functional and perceptual features. The emergence of category-specific impairments may actually be secondary to selective damage to different types of features. Both functional and perceptual features are relevant to artifacts and natural kinds, but perceptual features are more important than functional ones for natural kinds and are about equally important for artifacts. In a model created by Farah and McClelland (1991), when damage was applied to perceptual features, a deficit in natural kinds arose, but when damage was applied to functional features a deficit in artifacts arose. The model, although not organized semantically by category, produced category-specific deficits. This model relies on the basic assumption that information is topographically organized such that focal brain damage can preferentially affect perceptual or functional features. CVA and HSE are both examples of focal brain damage that the model mirrors well. AD, on the other hand, leads to widespread, patchy damage. Recent evidence suggests that AD subjects display the same kinds of category-specific deficits, but it is unclear how similar deficits could result from different types of damage. The purpose of the article was to provide a connectionist computational account of how similar patterns of category-specific semantic impairment could arise both from localized/focal and widespread damage. The authors created a connectionist model based on two assumptions from Farah and McClelland. These two assumptions are that perceptual and functional features are topographically distinct and the ratios of perceptual and functional features differ in natural kinds and artifacts. They also rely on two additional assumptions about the properties of semantic representation. First, features differ in the degree to which they help distinguish among concepts. For example, “has fur” or “has claws” is less helpful than “has stripes” when distinguishing between lions and tigers. Second, intercorrelations exist among features and there are differences in the distribution of the intercorrelated features across natural kinds and artifacts. It has been found that there are more intercorrelations for natural kinds than artifacts. For example, if an animal has fur, it is also likely to have claws, whiskers, and a tail. The authors predict that category-specific impairments can arise in AD subjects due to random damage in connections between semantic units. When random damage affects the semantic system, categories that contain many informative features and few intercorrelated ones (artifacts) lose individual category members. On the other hand, as the amount of damage increases, categories with many intercorrelated features (natural kinds) will lose clusters of category members simultaneously affected by the loss of shared sets of features (e.g., lemons, limes, oranges). Hence, initially, with less dabmage, the authors predict a deficit in artifacts will occur, but as damage increases, a deficit in natural kind will become evident. Experimental Work In the present set of experiments, a connectionist model was trained on a set of 60 words, half of which were artifacts and half of which were natural kinds. The model’s semantic representations were based on feature norms collected from 50 undergraduates. The undergraduates were given a word and asked to list its perceptual and functional features. If a feature described a visual, auditory, or tactile property, it was considered perceptual. A feature was considered functional if it described what an item does or what it is used for. Experiment 1 The goal of Experiment 1 was to simulate the behavioral effects of widespread, patchy brain damage (as seen in AD) using a single mechanism, damage to the connections between semantic units. Testing was equivalent to a picture-naming task in that semantic features were activated, and these then activated corresponding phonological outputs. Because AD is progressive, the model was “lesioned” cumulatively with initially small increments of damage that gradually grew larger as damage increased. The general pattern of results indicated that low levels of damage led to more naming errors on artifacts than natural kinds, but with increasing damage, a deficit appeared for natural kinds over artifacts. These results parallel those observed in AD subjects, but the simulations did vary considerably. Some simulations resulted in more difficulty with natural kinds that persisted throughout the simulation. One simuluation had an initial deficit in natural kinds which progressed into an artifact deficit. This variability, though, also mirrors variability in AD subjects and actually helps to explain some inconsistencies found in the literature. Experiment 2 Experiment 2 tested a different type of damage to the model. Instead of a progressive loss of connections within the semantic system, a progressive loss of semantic units was applied. This was done to test whether the different types of damage would yield different results. The resulting data provided evidence that the damage applied to the model in Experiment 1, a progressive loss of connections, more closely resembled subject data than the damage applied in Experiment 2. For low levels of damage, a small natural kinds deficit emerged in Experiment 2, but a global anomia appeared as the model became more severely impaired. The range of emerging patterns did not match any observed patterns in AD subjects. This indicated that the analogy of a loss of connections in the model is closer to the real neurobiology of AD patients than a loss of semantic units. Experiment 3 The purpose of Experiment 3 was to replicate previous research regarding focal brain damage and category-specific impairments with the newly developed model. Instead of progressive damage to the model, damage was “one-shot” and corresponded to differing degrees of unit loss to either perceptual or functional units. When damage was applied to perceptual units, a selective deficit for natural kinds emerged, while a selective deficit for artifacts emerged when damage was to functional units, indicating a successful replication. Conclusions The model demonstrates how the same behavioral deficit (category-specific impairments) can arise from different types of brain damage (focal vs. widespread). It also helps to explain the variability and dissociations observed from neurologically impaired subjects. The present findings also indicate that the following four properties of the semantic system are relevant to category-specific deficits:
The intercorrelation may explain the pattern and time course of deficits in AD. Points for Discussion Consider how the semantic representations of the model were constructed. Has-an-engine was not listed for bus, but that is certainly a part of our knowledge of a bus. Has-a-mouth was only listed for hippo, but other items certainly have mouths. Is this damaging to the model or do you think we actually only think of the salient semantic features when identifying an object? In other words, do we even care that buses have engines or that dogs have mouths when we identify them? The authors mention the importance of how their model addresses both the variability in individual subjects and the group profile. Do you think one is more important than the other is? How careful should models be in accounting for patient data? Should a given model be responsible for accounting for any observed anomaly, for some, or maybe none? If our semantic system does not store objects based on category, how do we all end up with them? Are categories all just learned?
|
|
This page was last updated: |