Connectionist modeling is applied to issues in cognitive rehabilitation, concerning the degree and speed of recovery through retraining, the extent of generalization to untreated items, and how treated items are selected to maximize this generalization. A network previously used to model impairments in mapping orthography to semantics is retrained after damage. The degree of relearning and generalization varies considerably for different lesion locations, and has interesting implications for understanding the nature and variability of recovery in patients. In a second simulation, retraining on words whose semantics are atypical of their category yields more generalization than retraining on more prototypical words, suggesting a surprising strategy for selecting items in patient therapy to maximize recovery.

A central challenge In cognitive neuroscience is to explain how disorders of brain function give rise to disorders of cognition. In this regard.connectionist modeling provides a useful computational formalism for relating cognitive processes to their underlying neurological Implementation. In the domain of visual comprehension of words and objects. I will show how two peculiar patterns of impairment observed after brain damage. deep dyslexia and optic aphasia, also arise in simulations embodying a set of general computational principles: (1) visual and semantic Information is represented as distributed patterns of activity over separate groups of units such that the patterns exhibit the appropriate similarities within and between these domains: (2) the knowledge of the relationships between representations is encoded as weights on connections between units: and (3) the mapping between representations is accomplished by Interactivity among umts, forming "attractors" for familiar patterns of activity. Further assumptions are that short-term correlational information is useful in object recognition but not In word recognition, and that there is less structure in the mapping from visual to semantic representations for words than for objects. In a simulation of word reading, damage leads to the peculiar Interactions of visual and semantic similarities in errors found in deep dyslexia. In a simulation of object naming, very few purely visual errors occur after damage but now semantic similarity Interacts with perseverative effects from previous trials, as in optic aphasia. The replication of complex empirical phenomena concerning Impaired visual comprehension of both words and objects provides evidence that the general principles underlying the simulations also apply to the semantic processing of visual information and its breakdown following brain damage In humans.

A distributed attractor network is trained on an abstract version of the task of deriving the meanings of written words. When processing a word, the network starts from the final activity pattern of the previous word. Two words are semantically related if they overlap in their semantic features, whereas they are associatively related if one word follows the other frequently during training. After training, the network exhibits two empirical effects that have posed problems for distributed network theories: much stronger associative priming than semantic priming, and significant associative priming across an intervenmg unrelated item. It also reproduces the empirical findings of greater priming for low-frequency targets, degraded targets, and high-dominance category exemplars.

Deep dyslexics are patients with neurological damage who exhibit a variety of symptoms in oral reading, including semantic, visual and morphological effects in their errors, a part-of-speech effect, and better performance on concrete than abstract words. Extending work by Hinton & Shallice (1991), we develop a recurrent connectionist network that pronounces both concrete and abstract words via their semantics, defined so that abstract words have fewer semantic features. The behavior of this network under a variety of "lesions" reproduces the main effects of abstractness on deep dyslexic reading: better correct performance for concrete words, a tendency for error responses to be more concrete than stimuli, and a higher proportion of visual errors in response to abstract words. Surprisingly, severe damage within the semantic system yields better performance on abstract words, reminiscent of CAV, the single, enigmatic patient with "concrete word dyslexia."

Optic aphasic patients are selectively impaired at naming visually presented objects but demonstrate relative intact comprehension of those objects (e.g., by gesturing or categorization) and are able to name them when presented in other modalities (e.g., via tactile input). This and other modality-specific naming deficits have been taken as evidence that semantics is organized into distinct modality-specific subsystems. We adopt an alternative view in which semantics is a set of learned, internal representations within a parallel distributed processing system that maps between multiple input and output modalities. We account for the critical aspects of optic aphasia in terms of the effects of damage to such a system, despite its lack of modality-specific specialization. We show that the robustness of a task in such a system depends critically on its systematicity, and that modality-specific naming deficits can arise because naming is an unsystematic task.

Networks that leam to make familiar activity patterns into sta- ble attractors have proven useful in accounting for many aspects of normal and impaired cognition. However, their ability to generalize is questionable, particularly in quasiregular tasks that involve both regularities and exceptions, such as word reading. W e trained an attractor network to pronounce virtually all of a large corpus of monosyllabic words, including both regular and exception words. W h e n tested on the lists of pronounceable nonwords used in several empirical studies, its accuracy was closely comparable to that of human subjects. The network gen- eralizes because the attractors it developed for regular words are componential—they have substructure that reflects common sub- lexical correspondences between orthography and phonology. This componentiality is faciliated by the use of orthographic and phonological representations that make explicit the structured relationship between written and spoken words. Furthermore, the componential attractors for regular words coexist with much less componential attractors for exception words. These results demonstrate that attractors can support effective generalization, challenging "dual-route" assumptions that multiple, independent mechanisms are required for quasiregular tasks.

Connectionist models usually have a single weight on each connection. Some interesting newproperties emerge if each connection has two weights: A slowly changing, plastic weight which stores long-term knowledge and a fast-changing, elastic weight which stores temporary knowledge and spontaneously decays towards zero. If a network learns a set of associations and then these associationsare "blurred" by subsequent learning, all the original associations can be "deblurred" by rehearsing on just a few of them. The rehearsal allows the fast weights to take on values that temporarily cancel outthe changes in the slow weights caused by the subsequent learning.