Who says models can only do what you tell them? Unsupervised category learning data, fits, and predictions

Abstract

Who says models can only do what you tell them? Unsupervised category learning data, fits, and predictions Todd M. Gureckis (gureckis@love.psy.utexas.edu) Department of Psychology; The University of Texas at Austin Austin, TX 78712 USA Bradley C. Love (love@psy.utexas.edu) Department of Psychology; The University of Texas at Austin Austin, TX 78712 USA Abstract How do people learn and organize examples in the absence of a teacher? This paper explores this ques- tion through a examination of human data and com- putational modeling results. The SUSTAIN (Super- vised and Unsupervised STratified Incremental Net- work) model successfully fits human learning data drawn from two published studies. The first study examines how correlations between features can fa- cilitate unsupervised learning. The second set of studies examines the role that similarity and at- tention play in unsupervised category construction (i.e., sorting) tasks. Importantly, SUSTAIN sug- gests two novel behavioral predictions that are con- firmed. Introduction The study of human category learning has focused on supervised learning. Researchers typically utilize a experimental procedure in which the participant must learn to classify a set of stimuli while receiving corrective feedback on every trial. Certainly, there are many other ways to learn about the world. Our environment does not always provide us with explicit feedback and thus, some learning is better charac- terized as unsupervised. For example, we routinely categorize incoming email as “junk mail” in the ab- sence of a teacher. A great deal of human learning may be unsupervised. The goal of this paper is to expand our understanding of how humans learn from examples without supervision. To achieve this goal, we fit the SUSTAIN model of category learning to Billman and Knutson’s (1996) studies concerning how humans learn correlations through observation and to Medin, Wattenmaker, and Hampson’s (1987) data on unsupervised cate- gory construction (i.e., sorting) behavior. SUSTAIN successfully accounts for human performance in both of these studies with one set of parameters. Impor- tantly, SUSTAIN’s account of these studies suggests novel predictions which are subsequently tested (and confirmed) with human subjects. The Modeling Approach SUSTAIN has been successfully applied to an array of challenging human data sets spanning a variety of category learning paradigms including supervised classification (Love & Medin, 1998), inference learn- ing (Love, Markman, & Yamauchi, 2000), and un- supervised learning (Gureckis & Love, 2002). One primary goal of our modeling approach is to address multiple forms of category learning (both supervised and unsupervised) with one consistent set of princi- ples. After a brief introduction to the operation of SUSTAIN, these core principles will be discussed. Introduction to SUSTAIN SUSTAIN is a clustering model of human category learning. The internal representation of the model consists of a set of clusters. Category representations consist of one or more associated clusters. At the start of learning, the network has a single cluster that is centered in this representational space upon the first input pattern. When a new stimulus item is presented, SUSTAIN attempts to assign the item to the most similar ex- isting cluster. This assignment is unsupervised since it is based only on the similarity between item and cluster. If a surprising event occurs, such as a mis- prediction in supervised learning or a stimulus is en- countered in unsupervised learning that is not simi- lar to any existing cluster, SUSTAIN creates a new cluster to encode the current stimulus. This new cluster is centered in the representational space on the misclassified item. When a stimulus is not surprising, the item is as- signed to the most similar existing cluster and this cluster updates its internal representation to become more similar to the current item (a process some- what analogous to prototype formation). Classifica- tion decisions are based on the cluster to which a stimulus instance is assigned. Like other models of category learning (e.g., Kruschke, 1992), SUSTAIN’s selective attention mechanism learns to selectively weight stimulus feature dimensions that are most useful for categorization. The Principles of SUSTAIN With this general understanding of the operation of the model, we now examine the six key principles that underly SUSTAIN.