Figure 2.

Choice bias during category learning reflects a dynamic preference that mice rely on when drawing category boundaries. A, The PsyTrack model extracts dynamic weights for each category as well as a dynamic choice bias term. Error bars represent 95% CIs. Right: The original traces for the category-conditioned accuracies throughout learning can be estimated from the extracted weights by taking the logistic transformation of the weights. B, The session-level behavioral choice data at three different points throughout learning. C, Visualization of psychometric curve parameters. The animal behavior is modeled as a weighted average of attending behavior (as captured by a sigmoid) and lapsing behavior, which is stimulus-independent. D, Left: The drift over the final day of training is correlated with across-mouse variability in the location of the psychometric threshold during the first 5 testing days. Center: Two examples of isolated GLM choice bias traces. Top panel shows a trace with high variability, bottom panel shows a trace with low variability. Right: Psychometric curves from the testing stage for the same mice, demonstrating that a mouse with higher (lower) variability in dynamic weight has higher (lower) variability in threshold location. Thin lines are individual sessions, thick line is generated from average parameters across the first five sessions. E, The drift over the final day of training is not significantly correlated with the across-mouse variability in the psychometric slope. F, The drift over the final day of training is not significantly correlated with the across-mouse variability in the psychometric guess rate. G, The drift over the final day of training is not significantly correlated with the across-mouse variability in the psychometric lapse rate. In all plots: n.s., not significant; *p<0.05; **p<0.01; ***p<0.001; ****p<0.0001.