Search Results

This chapter aims to describe the connectionist perspective on repetition priming. The first section provides an overview of the connectionist framework, and more specifically, connectionist models of word perception. The section that follows reviews empirical findings concerning several different issues related to repetition priming — the purpose of this section is both to show how the network approach has served to generate new hypotheses and to demonstrate how this approach allows for new conceptualizations of some old ideas. The chapter closes with a discussion of the relationship between the connectionist approach and accounts which have developed within the activation and memory traditions.

This chapter focuses on case studies of connectionist models of perseveration for understanding infants' reaching behavior and children's rule-guided behavior. These models demonstrate how general connectionist principles (1) provide a unified framework for understanding perseveration across ages, task paradigms, and response modalities, (2) capture important details of children's performance, and (3) lead to novel and testable predictions.

This chapter examines two approaches to resolving the question of how prior knowledge and current knowledge interact in category learning. The first relies on mathematical models of statistical inference. The second is an implemented connectionist computational model. To illustrate the usefulness of these latter models, the chapter develops a possible connectionist model of how prior knowledge and on-line learning integrate during early concept learning.

This chapter focuses on a connectionist model called the fractal learning neural network (FLNN). The FLNN model's account is appealing because it (a) uses a simple learning mechanism and thus corresponds to a theory that links development and learning; (b) predicts stagewise behavior with progressive quantitative changes occurring within the stages, a phenomenon also observed in development; (c) derives phase transitions from the learning, thus predicting, rather than stipulating, the stages; (d) shows highly structured, domain-specific behavior arising from a mechanism which is capable of a wide variety of behaviors, not all of which are highly structured (thus, in some sense, deriving the structure where no structure was apparent); and (e) generalizes beyond its input in a way that is, at least partially, consistent with human generalization in the comprehension and production of recursive sentence structures.

Computational models can be outlined into two distinct kinds: symbolic models and connectionist models. One particular focus of computational modeling research has been the development of object interactions during infancy. In this chapter, the authors describe a number of computational approaches to understand the developing object concept, focusing particularly on the contribution of such models to our understanding of the development of representations governing infant–object interactions across the first two years of life. The authors argue that connectionist models have an important role to play in directing theoretical and empirical research in this area, because they best enable researchers to understand the causal factors at play in the ontogeny of knowledge about objects and cognition in general. A brief review of connectionist modeling principles is also presented.

This chapter asks how multisensory perception can develop. That is, what are the mechanisms by which separate modalities become differentiated and integrated during the first years of life? This chapter argues that computer models provide a powerful tool for answering these questions. The chapter begins by describing the role of computational modelling in understanding causal mechanisms of developmental change, going on to characterise connectionist neural network models as a family of modelling approaches that is particularly well suited for studying learning and development. In the next three sections the chapter illustrates how neural network models have been used to grapple with how pair-wise multisensory (and likewise, sensorimotor) integration develops beginning with auditory-visual, then auditory-motor, and then finally visual-motor coupling. Finally, the chapter reviews our findings and point to some challenges for future research.

This chapter begins with a discussion of the arguments and concepts of mathematician turned theoretical neurophysiologist, David Marr. It then covers the rejection of the Marrian approach; modules, isolable subsystems, and functional specialisation; the anatomical bases of modularity; information-processing models; the COGENT formalism; connectionist models; and more advanced symbolic architectures.

This chapter discusses the methodological criteria of implicit knowledge, the meaning of implicit/explicit knowledge, the three types of explicitness, analysis of a connectionist model of artificial grammar learning, and the differences of connectionist models and implicit knowledge. It focuses on artificial grammar learning as a paradigm example of the attainment of knowledge that appears to be implicit by some criterion but not by others. Typical artificial grammar learning tasks ask the subjects to first memorize strings of letters that are generated by a finite state grammar and then they are informed that the strings were generated by a complex set of rules. The subjects were then ask to classify new strings as either obeying the rules or not. Subject knowledge are assessed as being implicit by using four methodological criteria. This chapter also discusses a possible connectionist model of artificial grammar learning and also considers the way connectionist models illuminate the way in which subjects knowledge can be regarded as implicit.

This chapter introduces formal models of cognition and explains how they are similar to verbal theories but use computer programs and mathematics to avoid the many limitations of human reasoning, thereby adding precision and rigor to their explanations. The chapter discusses Marr’s (1982) levels of analyses and how information-processing systems can be understood and described in terms of the task being performed, the representations and algorithms used to perform the task, and how the latter are implemented by physical systems. This then motivates discussion of three common approaches to modeling human cognition and behavior: process models, production-system models, and connectionist models. Each of these approaches is critiqued, with discussion of its merits and limitations. The three modeling approaches are then further illustrated by showing how each might be used to explain the finding that words can be identified more efficiently if they occur in predictable sentence contexts. The chapter closes with a discussion of how cognitive models are evaluated using their simplicity, theoretical scope, compatibility (e.g., with biology), and their capacity to generate novel predictions for guiding research.

This chapter defends that the thesis that, if a certain family of connectionist hypotheses turn out to be right, they will surely count as revolutionary. The chapter is organized as follows. Section 2 gives a brief account of what eliminativism claims, and sketches a pair of premises that eliminativist arguments typically require. Section 3 discusses how we conceive of common sense psychology and the propositional attitudes that it posits. It also illustrates one sort of psychological model that exploits and builds upon the posits of folk psychology. Section 4 is devoted to connectionism. Models that have been called “connectionist” form a fuzzy and heterogeneous set whose members often share little more than a vague family resemblance. However, the present argument linking connectionism to eliminativism will work only for a restricted domain of connectionist models, interpreted in a particular way; the main job of Section 4 explains what that domain is and how the models in the domain are to be interpreted. Section 5 illustrates what a connectionist model of belief that comports with our strictures might look like, and goes on to argue that if models of this sort are correct, then things look bad for common sense psychology. Section 6 assembles some objections and replies.

This chapter examines the structural characteristics of a cognit, that is, the morphological feature of the cognitive networks of the cerebral cortex. Knowledge about oneself and the environment is the subject matter of cognitive science and the matter of all cognitive functions: perception, attention, memory, intelligence, and language. For more than a century, neurologists and experimental psychologists have endeavored to assign one or another aspect of cognition, beginning historically with language, to one or another sector of the cerebral cortex. Largely because it derives from neural inferences, connectionism has become the most plausible model of the organization of knowledge in the cerebral cortex. All connectionist models and neural network models assume the distribution of knowledge in assemblies of units, neurons, or nodes that constitute and represent the component elements of knowledge. This chapter also looks at different categories of knowledge, cortical modularity, cortical hierarchy of perceptual networks and executive networks, and hierarchical representation in association cortex.

This chapter empirically investigates the issue of systematicity and connectionism under more realistic conditions than was the case in previous studies. A connectionist and a symbolic model of sentence processing are compared on their ability to perform systematically. Both models are trained on over 700,000 sentences, and tested on 361 sentences, from naturally occurring texts. Although the symbolic model does display slightly stronger systematicity, there is a striking similarity between the two models’ performance. It is argued that real-world tasks pose such strong demands and constraints that performance cannot differ much across models. Consequently, the issue of systematicity loses much of its relevance.

This chapter examines what might be some of the causes of the language impairments observed in children with specific language impairments (SLI). These are children who appear to function in the normal range on a broad range of cognitive tasks, but who have very poor language skills. One alternate hypothesis holds that such children have subtle impairments in the detection or processing of very low-level auditory features. The chapter argues that these low-level deficits, coupled with the fact that language acquisition is a cumulative developmental process, are the real causes of these language impairments. Neural network models can be used to illustrate how early low-level auditory processing deficits can lead to specific syntactic deficits in later development. From the neuroconstructivist perspective, the important conclusion here is that high-level cognitive abilities are constrained by the low-level input constraints of the body.

This chapter reviews origins of the modern scientific inquiry into cognition by examining the contributions to the field of behavioral neurology spanning the rich period from 1860 to 1950. Modularity of function, concepts of hemispheric asymmetry, and the early modeling of cognitive function based on connectionstic principles are reviewed. Specifically the contributions of Broca, Wernicke, Lichtheim, Liepmann, and Lissauer are discussed while showcasing the lasting insights these investigators made to the understanding of aphasia, apraxia, and visual agnosia. In this time before modern imaging, advances were often derived from the study of neuropathological correlations. Case studies were informed by principles of brain organization and led to the generation of new hypotheses and models of neuropsychological functions that remain relevant today.

This chapter presents an upgrade of the neural network by implementing the reward prediction error. It then compares the final product with the actor-critic model and discusses the similarities and differences. Reinforcement learning algorithms, more specifically actor-critic models, are currently very successful in the field of decision-making. They are notably related to properties of dopaminergic neurons which have not yet been addressed in previous chapters. It has been demonstrated that dopaminergic neurons respond when the subject receives a reward or when the subject associates a conditional stimulus with the reward, and that this response to the stimulus is proportional to the utility function of the reward. In fact, dopaminergic neurons behave exactly like a process that computes temporal difference. The amplitude of their response when the reward is administered is proportional to the difference between the expected utility at time and the reward actually obtained at the moment, i.e. the temporal difference. This chapter then assesses whether the telencephalic loop is an actor-critic system.