SUSTAIN (Supervised and Unsupervised STratified Adaptive Incremental Network) is a model of how humans learn categories from examples. SUS-TAIN initially assumes a simple category structure. If simple solutions prove inadequate and SUSTAIN is confronted with a surprising event (e.g., it is told that a bat is a mammal instead of a bird), SUSTAIN recruits an additional cluster to represent the surprising event. Newly recruited clusters are available to explain future events and can themselves evolve into

The nature of categories and their acquisition is one of the central open questions in Cognitive Science. We suggest that categories are represented via structured descriptions and formed by a process of progressive abstraction, through successive comparison with incoming exemplars. This paper describes how SEQL (Skorstad, Gentner, & Medin, 1988), a computer model for category learning, which is based on SME (Falkenhainer et al 1986, 1989; Forbus et al 1994) can be used to simulate a recent categorization experiment (Ramscar & Pain, 1996), using a new algorithm, Generalization and Exemplar Learning (GEL). We demonstrate that SEQL produces behavior consistent with human subjects. Introduction Similarity is often viewed as central to categorization. For instance, prototype theories of categorization posit that categorization decisions are made on the basis of the similarity of an entity to the prototypical member of that category (Rosch 1975). However, similarity-based accou...

The rational model of categorization (RMC; Anderson, 1990) assumes that categories are learned by clustering similar stimuli together using Bayesian inference. As computing the posterior distribution over all assignments of stimuli to clusters is intractable, an approximation algorithm is used. The original algorithm used in the RMC was an incremental procedure that had no guarantees for the quality of the resulting approximation. Drawing on connections between the RMC and models used in nonparametric Bayesian density estimation, we present two alternative approximation algorithms that are asymptotically correct. Using these algorithms allows the effects of the assumptions of the RMC and the particular inference algorithm to be explored

Memory evolved to supply useful, timely information to the organism’s decision-making systems. Therefore, decision rules, multiple memory systems, and the search engines that link them should have coevolved to mesh in a coadapted, functionally interlocking way. This adaptationist perspective suggested the scope hypothesis: When a generalization is retrieved from semantic memory, episodic memories that are inconsistent with it should be retrieved in tandem to place boundary conditions on the scope of the generalization. Using a priming paradigm and a decision task involving person memory, the authors tested and confirmed this hypothesis. The results support the view that priming is an evolved adaptation. They further show that dissociations between memory systems are not—and should not be—absolute: Independence exists for some tasks but not others. Memory is a gift of nature, the ability of living organisms to retain and to utilize acquired information or knowledge.... Owners of biological memory systems are capable of behaving more appropriately at a later time because of their experiences at an earlier time, a feat not possible for organisms without memory. (Tulving, 1995a, p. 751) If there is one proposition on which all psychologists seem to

Categorization research typically assumes that the cognitive system has access to a (more or less noisy) representation of the absolute magnitudes of the properties of stimuli and that this information is used in reaching a categorization decision. However, research on identification of simple perceptual stimuli suggests that people have very poor representations of absolute magnitude information and that judgments about absolute magnitude are strongly influenced by preceding material. The experiments presented here investigate such sequence effects in categorization tasks. Strong sequence effects were found. Classification of a borderline stimulus was more accurate when preceded by a distant member of

An important element of learning from examples is the extraction of patterns and regularities from data. This paper investigates the structure of patterns in data defined over discrete features, i.e. features with two or more qualitatively distinct values. Any such pattern can be algebraically decomposed into a spectrum of component patterns, each of which is a simpler or more atomic ‘‘regularity.’ ’ Each component regularity involves a certain number of features, referred to as its degree. Regularities of lower degree represent simpler or more coarse patterns in the original pattern, while regularities of higher degree represent finer or more idiosyncratic patterns. The full spectral breakdown of a pattern into component regularities of minimal degree, referred to as its power series, expresses the original pattern in terms of the regular rules or patterns it obeys, amounting to a kind of ‘‘theory’ ’ of the pattern. The number of regularities at various degrees necessary to represent the pattern is tabulated in its power spectrum, which expresses how much of a pattern’s structure can be explained by regularities of various levels of complexity. A weighted mean of the pattern’s spectral power gives a useful numeric summary of its overall complexity, called its algebraic complexity. The basic theory of algebraic decomposition is extended in several ways, including algebraic accounts of the typicality of individual objects within concepts, and estimation of the power series from noisy data. Finally some relations between these algebraic quantities and empirical data are discussed.

Rational models of cognition typically consider the abstract computational problems posed by the environment, assuming that people are capable of optimally solving those problems. This differs from more traditional formal models of cognition, which focus on the psychological processes responsible for behavior. A basic challenge for rational models is thus explaining how optimal solutions can be approximated by psychological processes. We outline a general strategy for answering this question, namely to explore the psychological plausibility of approximation algorithms developed in computer science and statistics. In particular, we argue that Monte Carlo methods provide a source of “rational process models” that connect optimal solutions to psychological processes. We support this argument through a detailed example, applying this approach to Anderson’s (1990, 1991) Rational Model of Categorization (RMC), which involves a particularly challenging computational problem. Drawing on a connection between the RMC and ideas from nonparametric Bayesian statistics, we propose two alternative algorithms for approximate inference in this model. The algorithms we consider include Gibbs sampling, a procedure

We live in a world consisting of concrete experiences, yet we appear to form abstractions that transcend the details of our experiences. In this contribution, we argue that the abstract nature of our thought is overstated and that our representations are inherently bound to the examples we experience during learning. We present three lines of related research to support this general point. The first line of research suggests that there are no separate learning systems for acquiring mental rules and storing exceptions to these rules. Instead, both items types share a common representational substrate that is grounded in experienced training examples. The second line of research suggests that representations of abstract concepts, such as same and diVerent that can range over an unbounded set of stimulus properties, are rooted in experienced examples coupled with analogical processes. Finally, we consider how people perform in dynamic decision tasks in which short ‐ and long‐term rewards are in opposition. Rather than invoking explicit reasoning processes and planning, people’s performance is best explained by reinforcement learning procedures that update estimates of action values in a reactive, trial‐by‐trial fashion. All three lines of research implicate mechanisms of thought that are capable of broad generalization, yet inherently local in terms of the procedures

Numerous proposals have been put forward concerning the nature of human category representations, ranging from rules to exemplars to prototypes. However, it is unlikely that a single, fixed form of representation is sufficient to account for the flexibility of human categories. In this chapter, we describe an alternative to these fixed-representation accounts based on the principle of adaptive clustering. The specific model we consider, SUSTAIN, represents categories in terms of feature bundles called clusters which are adaptively recruited in response to task demands.

We investigated the mechanisms by which concepts are learned from examples by manipulating the presentation order in which the examples were presented to subjects. We introduce the idea of a rule-based presentation order, which is a sequence that respects the internal organization of the examples within a category. We find that such an order substantially facilitates learning, as compared with previously known beneficial orders, such as a similarity-based order. We discuss this result in light of the central distinction between rule-based and similaritybased learning models. A number of studies have investigated whether category learning is influenced by the order in which examples are presented. Elio and Anderson (1981) found that categories are learned faster when training is blocked into groups of mutually similar examples (see also Elio & Anderson, 1984). More recently, Medin and Bettger (1994) demonstrated a strong learning advantage when training objects were presented in an order that tended to maximize similarity between successive examples. Other