Journal of Experimental Psychology:Learning, Memory, and Cognition2003, Vol. 29, No. 4, 626-640Copyright 2003 by the American Psychological Association, Inc.0278-7393/03/$12.00 DOI: 10.1037/0278-7393.29.4.626Learning Myopia: An Adaptive Recency Effect in Category LearningMatt JonesUniversity of MichiganWinston R. SieckOhio State UniversityRecency effects (REs) have been well established in memory and probability learning paradigms buthave received little attention in category learning research. Extant categorization models predict REs tobe unaffected by learning, whereas a functional interpretation of REs, suggested by results in otherdomains, predicts that people are able to learn sequential dependencies and incorporate this informationinto their responses. These contrasting predictions were tested in 2 experiments involving a classificationtask in which outcome sequences were autocorrelated. Experiment 1 showed that reliance on recentoutcomes adapts to the structure of the task, in contrast to models’ predictions. Experiment 2 providedconstraints on how sequential information is learned and suggested possible extensions to current modelsto account for this learning.Recency effects (REs) are a robust phenomenon in cognitivepsychology. REs are said to occur whenever more recent experi-ences are better remembered or are more influential in judgmentsabout present situations. For example, in research on verbal work-ing memory, REs are arguably among the most fundamental es-tablished phenomena, most commonly seen as increased perfor-mance on the final positions in free- or serial-recall tasks (e.g.,Crowder, 1972; Murdock, 1962). Similar results have since beenobserved in visuo-spatial working memory (Broadbent & Broad-bent, 1981), as well as in animals (Thompson & Herman, 1977;Wright, Santiago, Sands, Kendrick, & Cook, 1985). REs in work-ing memory have often been attributed to spontaneous decay ofstored information (Baddeley, 1986; Burgess & Hitch, 1999);however, this simple interpretation has been called into question byrecent results showing that the rate of information loss can change,adaptively, in response to temporal statistics of the task (R. B.Anderson, Tweney, Rivardo, & Duncan, 1997). This flexibility ismore consistent with a functional account of working memory(J. R. Anderson & Schooler, 1991, 2000; Schacter, 1999) andsuggests that there is more underlying the phenomenon than sim-ple architectural constraints.Another area in which REs commonly arise is animal condi-tioning experiments. Common learning phenomena that dependon trial order, such as extinction, counterconditioning, anddiscrimination-reversal learning, all fall into the category of REsbecause they are characterized by behavior at the conclusion oflearning being based primarily on the most recent (second) phaseof training, rather than an average of both phases. However, theexistence and magnitude of such trial order effects depend cru-cially on the relationship of physical and temporal contexts amongMatt Jones, Department of Psychology, University of Michigan; Win-ston R. Sieck, Department of Psychology, Ohio State University.This work was supported in part by National Science Foundation GrantSES-9911301.Correspondence concerning this article should be addressed to MattJones, who is now at the Department of Psychology, University of Texasat Austin, 1 University Station A8000, Austin, Texas 78712. Email:[email protected] phases of training and testing, in a manner that is easily arguedto be rational and adaptive (see Bouton, 1993, for a review).Similar REs are commonly observed in human experimentationwith probability learning, for instance as a tendency to respondwith whatever option was given as correct on the previous trial,independent of the reinforcement pattern over the prior history ofthe task (e.g., Jarvik, 1951; Nicks, 1959). This effect is easilypredicted both by associative-learning models and by models thatassume decay of memory for past events. However, such theorieshave trouble with further results indicating that REs in these tasksare subject to learning effects in response to sequential dependen-cies in the target sequence. In particular, when outcomes of suc-cessive trials are made to depend on each other, rather than beingindependently sampled, observed REs adapt accordingly (N. H.Anderson, 1960).Some evidence for order effects has also been found in multiple-cue category learning (MCCL). For instance, Busemeyer andMyung (1988) found that humans’ subjective prototypes forgroups of random dot patterns showed a temporal bias involvingboth primacy and REs. Still, a detailed investigation of sequentialeffects in MCCL has yet to be performed. This situation is re-flected in current models of the task, which either do not addressREs or treat them as static by-products of other processes. Such astance is at variance with the results from other domains, whichsuggest instead that REs may be adaptive and sensitive to statis-tical characteristics of the task. In particular, when viewed as anextension of probability learning, MCCL might be expected toproduce effects similar to those mentioned above wherein REsadapt to sequential dependencies among outcomes. On the otherhand, the presence of variable cues gives two reasons that thepattern might be different. First, there is evidence that REs inMCCL are more complex in that they depend on the similaritybetween present and previous stimuli (Sieck, 2000). Second,whereas uncued learners have little to attend to but the targetsequence, subjects in MCCL tasks have more tangible stimuliabout which they are explicitly instructed to learn. This attentionshift, combined with the memory demands associated with using626ADAPTIVE RECENCY627multiple cues, could conceivably eliminate or greatly attenuate anylearning of sequential dependencies.In the remainder of this article we describe a focused explora-tion of REs in category learning intended to address these issues.Our primary aims are to investigate (a) the presence and nature ofREs in cued category learning, (b) the role played by cues and bystimulus similarity, and (c) potential learning effects in response tosequential dependencies in the target sequence, along with themechanisms underlying such effects. We begin with a more de-tailed account of results from the probability learning literature, inwhich REs have been well documented and extensively investi-gated. Next we discuss some of the more
View Full Document
Unlocking...