Error discounting in probabilistic category learning

Abstract

The assumption in some current theories of probabilistic categorization is that people gradually attenuate their learning in response to unavoidable error. However, existing evidence for this error discounting is sparse and open to alternative interpretations. We report 2 probabilistic-categorization experiments in which we investigated error discounting by shifting feedback probabilities to new values after different amounts of training. In both experiments, responding gradually became less responsive to errors, and learning was slowed for some time after the feedback shift. Both results were indicative of error discounting. Quantitative modeling of the data revealed that adding a mechanism for error discounting significantly improved the fits of an exemplar-based and a rule-based associative learning model, as well as of a recency-based model of categorization. We conclude that error discounting is an important component of probabilistic learning.

Figure 1

Probability of responding with category A for each of the training items across training blocks for early, mid, and late conditions in Experiment 1. The legend identifies each item’s reinforcement probabilities, P(A|j), before and after the shift.

Figure 2

Observed slope through response probabilities across all 4 training items in Experiment 1 (solid lines) and objective slopes (dotted lines). Error bars indicate 95% confidence intervals. The three panels show the early (top), mid (middle), and late (bottom) condition, respectively.

Figure 3

Decisional-recency analysis for the mid and late conditions in Experiment 1. The probability of shifting the response from one category to the other on trial n is shown as a function of whether the response on trial n − 1 was an error or correct. The top panel is for two identical stimuli in succession and the bottom panel is for two maximally dissimilar stimuli following each other. The smooth lines are based on locally-weighted regression (lowess). See text for details of alignment of conditions.

Figure 4

Observed slope through response probabilities across all 4 training items in Experiment 2 (solid lines) and objective slopes (dotted lines). Error bars indicate 95% confidence intervals. The two panels show the early (top) and late (bottom) condition, respectively.

Figure 5

Decisional-recency analysis for all conditions in Experiment 2. The probability of shifting the response from one category to the other on trial n is shown as a function of whether the response on trial n − 1 was an error or correct. The top panel is for two identical stimuli in succession and the bottom panel is for two maximally dissimilar stimuli following each other. The smooth lines are based on locally-weighted regression (lowess). See text for details of alignment of conditions.