Fig. 8: Weight fluctuations amplify errors in categorization task.

a, b Stimulus space for categorization task. Each point represents a combination of inferred drift rates for a given trial in the pure DDM (with no learning) that was fit to identification task (see Methods for more details). Solid oval lines represent the Mahalanobis distance of 1 in relation to the population average for each of the eight stimuli. Solid black line depicts the ideal classifying process: above it implies a right-side decision, below it a left. Color code for each point and line follows the same logic as Fig. 2a, b. The larger overlap of each set in the identification task (a) explains the performance degradation, as most points are located around the origin. For the categorization task (b), the lack of overlap between stimuli clarifies the higher performance seen in Fig. 4. c, d Mean drift rates for the most difficult left-decision choice for Bayes-DDM model (A > B, with correct choices always being left choices). In the case of the identification task this is the 10⁻⁴/0 stimuli (c); for the categorization task the stimulus is the 56%/44% mixture (d). Blue signal the correct classified choices and red the incorrect. Considering the ideal separation bound, we show the projected histograms for the difference 〈μ₁〉−〈μ₂〉. e, f Same as c, d but now with fluctuating weights depicted as the slope of the category bound. Gray area indicates the weight fluctuations that are equivalent to 1 standard deviation for both tasks. Blue indicates trials that were originally incorrect in c, d but became correct, and red indicates trials that became incorrect but originally correct. Light gray dots indicate answers that remained unchanged. Histograms quantify the four populations of dots.