Toward a common representational framework for adaptation.

We develop a computational model-the adaptive representation model (ARM)-for relating 2 classic theories of learning dynamics: instance and strength theory. Within the model, we show how the principles of instance and strength theories can be instantiated, so that the validity of their assumptions can be tested against experimental data. We show how under some conditions, models embodying instance representations can be considered a special case of a strength-based representation. We discuss a number of mechanisms for producing adaptive behaviors in dynamic environments, and detail how they may be instantiated within ARM. To evaluate the relative strengths of the proposed mechanisms, we construct a suite of 10 model variants, and fit them to single-trial choice response time data from three experiments. The first experiment involves dynamic shifts in the frequency of category exposure, the second experiment involves shifts in the means of the category distributions, and the third experiment involves shifts in both the mean and variance of the category distributions. We evaluate model performance by assessing model fit, penalized for complexity, at both the individual and aggregate levels. We show that the mechanisms of prediction error and lateral inhibition are strong contributors to the successes of the model variants considered here. Our results suggest that the joint distribution of choice and response time can be thought of as an emergent property of an evolving representation mapping stimulus attributes to their appropriate response assignment. (PsycINFO Database Record (c) 2019 APA, all rights reserved).