Fig. 4: Biases in representation lead to biases in policy.

a After successful learning, behavior is forced back to chance. b Difference in firing rates on A and B trials (“splitness”) of two example neurons over 4000 trials of learning. Representation re-merges since neither C′ nor C′ results in positive reward above expectation. c Behavior, plotted as the fraction of correct turns on A trials (red) and B trials (black) over 4000 trials of learning. Left/right turn behavior is fixed to chance at trial 2000, chance behavior indicated by gray boxed area. d After successful learning, B-type splitters are replaced by generic place fields. e, f Same as in (b, c), but for the case of forced representational bias (d) leading to behavioral bias (f). g Population activity after general (left) or specific (right) ablations of splitters. Top row shows network activity following presentation of an A cue, while bottom row shows network activity following presentation of a B cue. h Behavior, plotted as the fraction of correct turns on A trials (red) and B trials (black) over 4000 trials of learning. Top panel, general ablation is performed at trial 2000, and turn performance is impacted across both trial types. Bottom panel, B-type splitters are specifically ablated at trial 2000, and turn performance is only impacted in B-type trials. For all plots, curves of activity and behavior are plotted as a running mean over 100 trials.