The 1980's saw two important developments in the sciences of the mind: The development of neural network models in cognitive psychology, and the rise of cognitive neuroscience. In the 1990's, these two separate approaches converged, and one of the results was a new field that we call "Computational Neuropsychology." In contrast to traditional cognitive neuropsychology, computational neuropsychology uses the concepts and methods of computational modeling to infer the normal cognitive architecture from the behabvior of brain-damaged patients. In contrast to traditional neural network modeling in psychology, computational neuropsychology derives constraints on network architectures and dynamics from functional neuroanatomy and neurophysiology.

We are thus about ten years into a new paradigm, with a growing literature, a maturing set of methods and models, and an emerging consensus on the theoretical issues that differentiate various models. However, work in computational neuropsychology has had relatively little contact with the Neural Information Processing Systems (NIPS) community. As a result, the NIPS community has not exploited the theoretical ideas developed in computational neuropsychology or the interesting data from neuropsychology. Conversely, computational neuropsychology is firmly rooted in the "parallel distributed processing" paradigm of the 1980's (Rumelhart & McClelland, 1987), and neuropsychological models seldom exploit the many exciting developments that have taken place in the NIPS community.

We thus propose a workshop to expose the NIPS community to the amazing patient cases in neuropsychology and the sorts of inferences that can be drawn from these patients based on computational models, and to expose researchers in computational neuropsychology to some of the more sophisticated modeling techniques and concepts that have emerged in recent years.

The workshop will not assume any specific background in neuropsychology. All speakers will present neuropsychological data in a manner comprehensible to researchers outside the field. We will attempt to bring in experimentalists as well as modelers, but because the most likely participants are primarily modelers, the focus will probably be on models of specific phenomena.

We are interested in speakers from all aspects of neuropsychology, including:

• attention (neglect)
• visual and auditory perception (agnosia)
• reading (acquired dyslexia)
• face recognition (prosopagnosia)
• memory (Alzheimer's, amnesia, category-specific deficits)
• language (aphasia)
• executive function (schizophrenia, frontal deficits).

If you are interested in speaking in the workshop, please send a brief description of your work or topic to Mike Mozer (mozer@cs.colorado.edu).

Click here for more information on the NIPS*2001 Workshops.

Hippocampal amnesia reveals that the hippocampus and the spared neocortex make different contributions to memory. Ken Norman and I have been applying neural network models of the neocortex and hippocampus to a variety of recognition memory phenomena to more precisely and mechanistically characterize these different contributions. Specifically, the models show how the neocortical and hippocampal contributions differ in their sensitivity to manipulations of item similarity and interference, making predictions that have been confirmed in neuropsychological and behavioral tests. Furthermore, they address some counterintuitive effects of partial damage, and provide measures of statistical correlations between familiarity and recall memory signals, which are often assumed to be independent.
 
 

 Traditionally, semantic memory is understood as general world knowledge which is not referenced to specific episodes. It is specific regarding its content rather than regarding its context, or source. As a consequence, it may have very few details and modalities involved. In contrast, an episodic memory is traditionally viewed as context-specific, rich in details, multimodal memory trace of a specific experience, which is referenced to one particular episode. Consistent with this understanding of the episodic-semantic distinction is the traditional theory of memory consolidation (Tayler & Discenna, 1985; Squire, 1992; McClelland, McNaughton & O'Reilly, 1995). An alternative episodic-semantic distinction is based on the notion of agency (e.g., Wheeler, Stuss and Tulving, 1997). The key assumption of the present work is that hippocampal activity patterns serve as representations of particular instances of the agent (the self). These representations are assumed to be "allocentric", while their counterpart "egocentric" representations are assumed to exist in the prefrontal cortex. A connectionist model based on these assumptions accounts for the main properties of the hippocampal amnesia, including the 3 cases reported by Vargha-Khadem et al. (1997) and several cases studied by Moscovitch. In addition, the model may help to clarify some mysterious details of schizophrenia and of the multiple personality disorder.
 
 

Many studies have shown that Parkinson's patients and amnesics perform poorly on probabilistic categorization tasks, including one known as "the weather prediction task." We have applied an adaptation of the Gluck & Myers cortico-hippocampal model to this weather task data in order to examine the nature of these patients' learning deficit.

This new model demonstrates that a change in learning rate can account for much of the weather task data. A high-level model of learning processes informed by neural circuitry, it consists of three interacting neural networks. The model's units cooperate to build representations of the stimuli and create the appropriate output.
 
 

There is a broad consensus that the hippocampal system (HS) consisting of the hippocampal formation (HF) and neighboring cortical areas in the ventromedial temporal lobe play a critical role in the encoding and retrieval of episodic memory. But how the HS subserves this mnemonic function is not fully understood.

SMRITI is a computational model of episodic memory that explicates the role of the HS in the acquisition, maintenance and retrieval of episodic memory, and proposes a detailed circuit-level explanation of how the HS realizes this function in concert with cortical representations. The model predicts the functional roles of different components of the HS, the properties of relational schemas underlying episodic memories, the sorts of memories that must persist in the HS for the long-term, the nature of memory consolidation, and memory deficits resulting from cell loss in the HS and high-level cortical circuits.

Computer simulations and quantitative analyses of the model using neurally plausible parameter values lead to several specific predictions about the nature of memory deficits that would result from insults to various cortical and hippocampal regions. These results have the potential for explaining the nature of memory deficits resulting from aging, and pathological conditions such as Alzheimer's disease and semantic dementia. The talk will provide an overview of the SMRITI model and discuss its predictions.
 
 

The implications of a neural-level model of the medial temporal and parietal  roles in retrieval of the spatial context of an event (Becker & Burgess, NIPS 13) are considered.  The model combines the idea that retrieval of the rich context of real-life events is a central characteristic of episodic memory, and the idea that medial temporal allocentric representations are used in long-term storage while parietal egocentric representations are used to imagine, manipulate and re-experience the products of retrieval. It is consistent with single unit recordings in both regions and provides
an explanation for the involvement of Papez's circuit in both the representation of heading direction and in the recollection of episodic information. I discuss two functional neuroimaging studies supporting the model's functional localization and three neuropsychological studies suggesting that the hippocampus stores an allocentric representation of
spatial locations, and is involved  in topographical and epsodic memory, with some lateralisation of function.
 
 

We present a model of the role of the prefrontal cortex in controlled memory storage and retrieval. The model self-organizes its own internal representations, which act as mnemonic codes, using internally generated performance measures. These mnemonic codes serve as retrieval cues by biasing retrieval in medial temporal lobe memory systems. We simulate strategic learning and retrieval effects in the free recall of categorized lists of words, in both normal and frontally lesioned models. Our model captures a number of aspects of the performance of intact and frontally lesioned individuals, but also has a number of shortcomings, which point to extensions of the model that would enable more sophisticated forms of strategic control.
 
 

Neuropsychological studies of patients with acquired semantic impairments have yielded two distinct and contrasting patterns of performance in a spoken-word/picture matching task (Warrington & Cipolotti, 1996, Brain, 119, 611-25). Patients labeled access/refractory are strongly influenced by presentation rate, semantic relatedness of distractors, and repetition, yet they seem relatively unaffected by lexical frequency. Degraded-store patients, on the other hand, are strongly affected by lexical frequency but less affected by presentation rate, semantic relatedness, or repetition. Our account of these patterns of performance is based on the distinction between two different types of neurological damage: 1) damage to neuromodulatory systems that function to amplify neural signals while suppressing normal refractory-like effects, and 2) damage to connections between groups of neurons that encode semantic information and are sensitive to frequency/familiarity. We present a connectionist model that learns to map spoken-word input to semantic representations and that incorporates a particular form of neural refractoriness referred to as synaptic depression, as well as a simple form of neuromodulation. We show that the model is capable of accounting for the contrasting pattern of semantic impairment under these two different forms of damage, and further demonstrate how it is capable of handling several documented cases that are exceptions to the basic patterns of impairment.
 
 

We examine the influence of the pre-existing organization of the semantic memory on forming new episodic associations between words. First, testing human subjects' performance we found that a semantic relationship between words facilitates forming episodic associations, increasing linearly as a function of the number of co-occurrences. Constrained by these empirical findings we developed a computational model based on spreading activation over laterally connected self-organizing maps. The model replicates the empirical results and predicts (among other effects) that a less focused activation wave, which may be the cause of schizophrenic thought disorder, should decrease the advantage in learning rate of related over unrelated pairs.
 
 

We have developed a neural network model of the brain systems involved in mediating the effects of nicotine on the performance of attentional tasks.  Our simulations have reproduced results showing smoking abstinence to impair performance on a rapid information processing (RIP) task involving responses to seeing particular patterns of even or odd digits.

Our model treats the presence of nicotine as an enhancement of modulatory acetylcholine signals from the midbrain, which are widely regarded as playing a role in focusing attention in working memory areas of the cerebral cortex toward stimuli relevant to the current task.  The model includes analogs of the nucleus basalis, the midbrain cholinergic nucleus; the dorsolateral prefrontal cortex, which is believed to be an important area for working memory; the entorhinal cortex, a likely gateway for cortical inputs to the nucleus basalis; and signals (presumed to be somewhere in rhinal cortex) representing recency of stimuli.   Digits are represented at the entorhinal, nucleus basalis, and dorsolateral levels, along with cortical representations of the "even" and "odd" categories.
 
 

We introduce a novel approach that addresses one of the fundamental questions in Computational Neuropsychology: that of the relation between structure and function. We develop an algorithm termed PPA (Performance Prediction Algorithm) that quantitatively measures the contributions of elements of a neural system to the tasks it performs. The algorithm identifies the neurons or areas that participate in a cognitive or behavioral task, given data about performance decrease in a small set of lesions, and serves as a powerful tool for predicting multi-lesion damage. The effectiveness of the PPA algorithm is studied in the analysis of Artificial and Evolved Neural Networks, and in reversible deactivation data from spatial attention studies in the mammalian brain. Our analysis enables a rigorous quantification of local versus distributed computation in the brain, and can serve in the analysis of functional imaging data.

[Joint work Ranit Aharonov, Lior Segev, Isaac Meilijson and Claus Hilgetag].