Associative change in Pavlovian conditioning: A reappraisal.

Abstract

Robert A. Rescorla changed how Pavlovian conditioning was studied and interpreted. His empirical contributions were fundamental and theoretically driven. One involved testing a central tenet of the model that he developed with Allan R. Wagner. The Rescorla-Wagner learning rule uses a pooled error term to determine changes in a directional association between the representations of the conditioned stimulus (CS) and unconditioned stimulus (US). This learning rule predicts that 2 equally salient CSs (A and B) will undergo equivalent associative change when they are conditioned in compound (i.e., AB!US). Rescorla's results suggested that this was not the case (e.g., Rescorla, 2000). Here, we show that these results can be reconciled with a model that uses a learning rule with a pooled error term once that rule is applied equivalently to all of the stimuli presented on a given trial, and the resulting reciprocal associations (directly and indirectly) contribute to performance. This model, called HeiDI, integrates several features of Rescorla's research and theorizing while addressing an issue that he recognized required further analysis: how learning is translated into performance.

Developments in associative theory: A tribute to the contributions of Robert A. Rescorla.

“…This finding points to the importance of an "individual" prediction error term that might operate in concert with a global error term (Rescorla, 2000a(Rescorla, , 2001c(Rescorla, , 2002a(Rescorla, , 2006c(Rescorla, , 2008. Whereas Rescorla's compound test procedure has generated a wealth of new data relevant to the basic problem of understanding what factors affect learning, his interpretation of the data continues to enjoy lively debate in the literature (e.g., Chan et al, 2021;Holmes et al, 2019;Honey et al, 2022;Jones et al, 2021;Spicer et al, 2022;Uengoer et al, 2020).…”

“…His own research, of course, revealed differences in the amount of associative gain or loss between stimuli conditioned in compound as a function of their individual prediction discrepancies. Whether or not this would require a wholesale overhaul of learning and computational models (e.g., Holmes et al, 2019; see in this issue Honey et al, 2022) remains an important topic.…”