Semantic Memory: Some Insights from Feature-Based Connectionist Attractor Networks

Publisher Summary

This chapter focuses on people's knowledge of concrete noun concepts, that is, lexical concepts corresponding to living and nonliving things such as chair and robin. The meaning of these types of words consists of the confluence of multiple knowledge types, including visual knowledge of various sorts (—for example, parts, shape, size, color, characteristic motion), knowledge associated with the other senses (the sounds things produce and how they smell, taste, and feel), typical behaviors of creatures, and multiple types of situation knowledge, such as knowledge about how, where, when, by whom, and for what purpose things tend to be used. It presents the evidence that people learn and use both feature correlations and relations. The discussion focused on the careful consideration of the nature of the knowledge, how it is learned, and the type of task used to tap into that knowledge, with particular emphasis on matching task and knowledge to maximize the probability of observing an influence of both feature correlations and relations. In conclusion, a full understanding of semantic memory can benefit maximally by striving to combine research on language use, concepts and categorization, object recognition, patient research, imaging research, and implemented computational models.

Introduction

Making our way around the world during our daily lives depends on a great deal of knowledge about events and the entities and things that are part of those events. This knowledge includes information regarding how entities behave on their own and how we use things to perform the functions that are necessary for daily living, like driving a car, putting on our clothes, preparing food, eating our fruits and vegetables, listening to music, and understanding the behaviors of the creatures that cohabit the earth with us. This knowledge builds across the developmental life span, seems to be computed naturally and effortlessly during adulthood, and can, unfortunately, break down due to neural impairments of various sorts.

Central to our ability to deal with all of these aspects of our daily lives is the knowledge that is subsumed under the umbrella of semantic memory. I use the term semantic memory to refer to people's memory for word meaning, where word meaning is construed broadly. Thus, semantic memory includes the various types of conceptual information that are tied to specific words. This subset of people's knowledge is central to accomplishing tasks such as recognizing and naming objects or pictures, computing the meaning of spoken and written words, and reasoning about possible identities, functions, and behaviors of objects or entities when presented with partial information. In this chapter, I focus on people's knowledge of concrete noun concepts, that is, lexical concepts corresponding to living and nonliving things such as chair and robin.¹.The meaning of these types of words consists of the confluence of multiple knowledge types, including visual knowledge of various sorts (e.g., parts, shape, size, color, characteristic motion), knowledge associated with the other senses (the sounds things produce and how they smell, taste, and feel), typical behaviors of creatures, and multiple types of situation knowledge, such as knowledge about how, where, when, by whom, and for what purpose things tend to be used. Semantic memory, of course, also includes verb (event) concepts such as telling and running, abstract concepts such as love and justice, and concepts corresponding to adjectives and adverbs. However, a great deal can be and has been learned from the study of concrete nouns.

For quite a while now, among researchers who study semantic memory, spreading activation networks (Collins 1975, Collins 1969) have been the bases for the majority of theorizing and empirical investigation. In fact, arguably, in terms of overall popularity, they may still be the approach on which most researchers base their work. Without question, spreading activation networks have played a huge role in propelling the field forward.

The game is changing, however, in large part due to two developments. One is the rise to prominence of connectionist models of semantic memory, particularly in the form of attractor networks (which can be viewed as computational updates and extensions of feature list models, such as those espoused by E. E. Smith, Shoben, & Rips, 1974). This includes important work by Farah 1991, Hinton 1991, Masson 1995, Plaut 1995, Plaut 1993). Attractor networks offer principles and metaphors that are extremely useful for understanding various phenomena in this domain. The second development is the excitement generated by the conjunction of research on patients with semantic impairments (Forde 1999, Martin 2003) and the imaging of semantic memory (Martin & Chao, 2001). These two developments are central to the story presented herein.

In this chapter, I describe some research that has been conducted in our lab over the past 8 years or so that focuses on predictions and insights derived from a connectionist feature-based approach to studying semantic computations. The methodological basis of our research has been semantic feature production norms that provide an empirically derived representation of people's semantic knowledge. Features are verbal proxies for packets of knowledge, such as <has a handle> or <swims>. Almost all models of semantic memory and concepts and categorization are based at least in part on the notion of semantic features, although they may be instantiated in various architectures (Collins 1975, Kruschke 1992, Love 2004, Medin 1978, Rehder 2003, Sloman 1998).

Connectionist attractor networks serve as the theoretical basis for our semantic memory research and the testing ground of our theories. There are two key aspects of attractor networks on which I focus in this chapter. The first is the fact that these models naturally encode and use the distributional statistics of patterns to which they are exposed. This serves as a straightforward prediction that humans do as well. The second aspect concerns the temporal dynamics of computations in these networks. Conducting simulations using models that gradually compute representations over time leads to intriguing insights and predictions that would not be possible otherwise (as compared to, for example, feed-forward back-propagation networks or static computations of similarity). These two aspects of attractor networks are interdependent because the manner in which concepts are computed over time depends on the distributional statistics that are stored in a network's weights. Predictions derived from the principles underlying these models and concrete simulations of human experiments can be used to test the validity of this approach. Thus, the goal of this chapter is to present evidence that a view inspired by feature-based attractor networks provides insight into behavioral phenomena regarding significant aspects of semantic knowledge and computations.

The outline of this chapter is as follows. Section II briefly describes our large set of feature norms and outlines how I view them, including some of their strengths and limitations. Section III presents some arguments concerning the reasons why attractor networks are a useful tool for studying semantic memory and conceptual computations. Section IV focuses on people's knowledge of implicit statistically based feature correlations and explicit theory-based feature relations. I emphasize a careful consideration of the character of each of these types of knowledge with respect to analyses of various tasks that might be used to test for their influence. In short, I illustrate that the influence of both types of knowledge is apparent in appropriate tasks. Section V focuses on the dynamics of concept similarity, in particular how the computational dynamics of attractor networks cohere seemingly inconsistent results regarding priming between basic-level concepts (truck–van) versus priming from superordinate to basic-level exemplar concepts (vehicle–truck). Section VI presents insights into the organization of semantic memory in the mind and brain that were inspired by connectionist principles and neural imaging. This investigation uses data regarding category-specific semantic deficits as the target phenomena to be explained. I conclude in Section VII.