The influence of prior experience on some forms of behavior and cognition is drastically affected by damage to the hippocampal system. However, if the hippocampal system is left intact both during the experience and for a period of time thereafter, subsequent damage can have much less or even no effect. Such findings suggest that memory traces change over time in a way that makes them less dependent on the hippocampal system. This process of change has often been called consolidation. Consolidation is a very gradual process; in humans, it appears to span up to 15 years. This article asks what consolidation is and why it occurs. We take as our point of departure the view that the initial memory trace that results from a relevant experience consists of changes to the strengths of the connections among neurons in the hippocampal system. Bidirectional connections between the neocortex and the hippocampus allow these initial traces to mediate the reinstatement of representations of events o...

Two connectionist frameworks, GRAIN (McClelland, 1993) and BSB (Anderson, 1991), and the diffusion model (Ratcliff, 1978) were evaluated using data from a signal detection task. Subjects were asked to choose one of two possible responses to a stimulus and were provided feedback about whether the choice was correct. The dependent variables included response probabilities, reaction times for correct and error responses, and reaction time distributions, and the independent variables were stimulus value, stimulus probability, and lag from an abrupt switch in stimulus probability. The diffusion model accounted for all aspects of the asymptotic data, including error reaction times, which had previously been a problem. The connectionist models accounted for many aspects of the data adequately, but each failed to a greater or lesser degree in important ways except for one model very similar to the diffusion model. The connectionist learning mechanisms were unable to account for initial learning or abrupt changes in stimulus probability. The results provide an advance in the development of the diffusion model and show that the long tradition of reaction time research and theory is a fertile domain for development and testing of connectionist assumptions about how decisions are generated over time.

It is commonly assumed that innate linguistic constraints are necessary to learn a natural language, based on the apparent lack of explicit negative evidence provided to children and on Gold's proof that, under assumptions of virtually arbitrary positive presentation, most interesting classes of languages are not learnable. However, Gold's results do not apply under the rather common assumption that language presentation may be modeled as a stochastic process. Indeed, Elman (Elman, J.L., 1993. Learning and development in neural networks: the importance of starting small. Cognition 48, 71--99) demonstrated that a simple recurrent connectionist network could learn an artificial grammar with some of the complexities of English, including embedded clauses, based on performing a word prediction task within a stochastic environment. However, the network was successful only when either embedded sentences were initially withheld and only later introduced gradually, or when the network itself was given initially limited memory which only gradually improved. This finding has been taken as support for Newport's `less is more' proposal, that child language acquisition may be aided rather than hindered by limited cognitive resources. The current article reports on connectionist simulations which indicate, to the contrary, that starting with simplified inputs or limited memory is not necessary in training recurrent networks to learn pseudonatural languages; in fact, such restrictions hinder acquisition as the languages are made more English-like by the introduction of semantic as well as syntactic constraints. We suggest that, under a statistical model of the language environment, Gold's theorem and the possible lack of explicit negative evidence do not implicate i...

We present a theoretical framework for understanding the roles of the hippocampus and neocortex in learning and memory. This framework incorporates a theme found in many theories of hippocampal function, that the hippocampus is responsible for developing conjunctive representations binding together stimulus elements into a unitary rep- resentation that can later be recalled from partial input cues. This idea appears problematic, however, because it is contradicted by the fact that hippocampally lesioned rats can learn nonlinear discrimination problems that require conjunctive representations. Our framework accommodates this finding by establishing a principled division of labor between the cortex and hippocampus, where the cortex is responsible for slow learning that integrates over multiple experiences to extract generalities, while the hippocampus performs rapid learning of the arbitrary contents of individual experiences. This framework shows that nonlinear discrimination problems are not good tests of hippocampal function, and suggests that tasks involving rapid, incidental conjunctive learning are better. We implement this framework in a computational neural network model, and show that it can account for a wide range of data in animal learning, thus validating our theoretical ideas, and providing a number of insights and predictions about these learning phenomena.

The development of reading skill and bases of developmental dyslexia were explored using connectionist models. Four issues were examined: the acquisition of phonological knowledge prior to reading, how this knowledge facilitates learning to read, phonological and non phonological bases of dyslexia, and effects of literacy on phonological representation. Compared with simple feedforward networks, representing phonological knowledge in an attractor network yielded improved learning and generalization. Phonological and surface forms of developmental dyslexia, which are usually attributed to impairments in distinct lexical and nonlexical processing “routes,” were derived from different types of damage to the network. The results provide a computationally explicit account of many aspects of reading acquisition using connectionist principles.

We present a computational neural network model of recognition memory based on the biological structures of the hippocampus and medial temporal lobe cortex (MTLC), which perform complementary learning functions. The hippocampal component of the model contributes to recognition by recalling specific studied details. MTLC can not support recall, but it is possible to extract a scalar familiarity signal from MTLC that tracks how well the test item matches studied items. We present simulations that establish key qualitative differences in the operating characteristics of the hippocampal recall and MTLC familiarity signals, and we identify several manipulations (e.g., target-lure similarity, interference) that differentially affect the two signals. We also use the model to address the stochastic relationship between recall and familiarity (i.e., are they independent), and the effects of partial vs. complete hippocampal

To investigate the relations between structure and function in both artificial and natural neural networks, we present a series of simulations and analyses with modular neural networks. We suggest a number of design principles in the form of explicit ways in which neural modules can cooperate in recognition tasks. These results may supplement recent accounts of the relation between structure and function in the brain. The networks used consist out of several modules, standard subnetworks that serve as higher-order units with a distinct structure and function. The simulations rely on a particular network module called CALM (Murre, Phaf, and Wolters, 1989, 1992). This module, developed mainly for unsupervised categorization and learning, is able to adjust its local learning dynamics. The way in which modules are interconnected is an important determinant of the learning and categorization behaviour of the network as a whole. Based on arguments derived from neuroscience, psychology, compu...

Humans routinely generalize universal relationships to unfamiliar instances. If we are told ‘‘if glork then frum,’ ’ and ‘‘glork,’ ’ we can infer ‘‘frum’’; any name that serves as the subject of a sentence can appear as the object of a sentence. These universals are pervasive in language and reasoning. One account of how they are generalized holds that humans possess mechanisms that manipulate symbols and variables; an alternative account holds that symbol-manipulation can be eliminated from scientific theories in favor of descriptions couched in terms of networks of interconnected nodes. Can these ‘‘eliminative’ ’ connectionist models offer a genuine alternative? This article shows that eliminative connectionist models cannot account for how we extend universals to arbitrary items. The argument runs as follows. First, if these models, as currently conceived, were to extend universals to arbitrary instances, they would have to generalize outside the space of training examples. Next, it is shown that the class of eliminative connectionist models that is currently popular cannot learn to extend universals outside the training space. This limitation might be avoided through the use of an architecture that implements symbol manipulation.

Previously, we have introduced the idea of neural network transfer, where learning on a target problem is sped up by using the weights obtained from a network trained for a related source task. Here, we present a new algorithm, called Discriminability-Based Transfer (DBT), which uses an information measure to estimate the utility of hyperplanes defined by source weights in the target network, and rescales transferred weight magnitudes accordingly. Several experiments demonstrate that target networks initialized via DBT learn significantly faster than networks initialized randomly. 1 INTRODUCTION Neural networks are usually trained from scratch, relying only on the training data for guidance. However, as more and more networks are trained for various tasks, it becomes reasonable to seek out methods that avoid "reinventing the wheel", and instead are able to build on previously trained networks' results. For example, consider a speech recognition network that was only trained on Ameri...

Adaptive network and exemplar-similarity models were compared on their ability to predict category learning and transfer data. An exemplar-based network (Kruschke, 1990a, 1990b, 1992) that combines key aspects of both modeling approaches was also tested. The exemplar-based network incorporates an exemplar-based category representation in which exemplars become associated to categories through the same error-driven, interactive learning rules that are assumed in standard adaptive networks. Experiment 1, which partially replicated and extended the probabilistic classification learning paradigm of Gluck and Bower (1988a), demonstrated the importance of an error-driven learning rule. Experiment 2, which extended the classification learning paradigm of Medin and Schaffer (1978) that discriminated between exemplar and prototype models, demonstrated the importance of an exemplar-based category representation. Only the exemplar-based network accounted for all the major qualitative phenomena; it also achieved good quantitative predictions of the learning and transfer data in both experiments. One of the major current models for explaining performance in arbitrary category learning paradigms is the context model proposed by Medin and Schaffer (1978) and elaborated by Estes (1986a) and Nosofsky (1984, 1986). According to the context model, people represent categories by storing individual exemplars in memory and make classification decisions on the basis of similarity comparisons with the stored exemplars. The context model has proved to be successful at predicting quantitative details of classification performance in a wide variety of experimental settings and has compared favorably with a variety of alternative models, including prototype, independent-feature, and certain logical-rule based models (see Medin & Florian, in press, and Nosofsky, in press-a, in press-b, for reviews). However, some shortcomings of the context model were recently demonstrated in series of probabilistic classification learning experiments conducted by Gluck and Bower (1988a)