Computations of Contingency Guide How Experience is Summarized in Memory

Abstract

The human mind spontaneously summarizes experience into predictive models by encoding regularities across individual observations. However, there is a variety of possible ways it may do so, and diverse models of this process have been proposed in different learning paradigms. Here, we bridge the tradition of statistical learning (SL) with that of classical conditioning to show that memory formation, not just anticipatory behavior, follows the principle of contingency or ΔP. We showed that participants' memory for pairwise predictive regularities reflects contingency rather than conditional probability, as often assumed. By this principle, the relationship between stimulus pairs A and B is a function not only of how often A is followed by B, but also how often B occurs without A. We showed that ΔP captures what participants remember about event pairs after exposure, revealing that this principle indeed guides how experience is summarized into models in memory during incidental encoding. We additionally showed that this effect can arise either from strong competing predictors or an overall high base rate of the outcome. An adapted version of the Rescorla Wagner learning model captured these findings, indicating a continuity between SL and many other forms of learning.