Fig. 2
From: An ensemble code in medial prefrontal cortex links prior events to outcomes during learning
Outcome-selective recall of an ensemble activity pattern is learning-related. a Example similarity matrices for neural population activity during inter-trial intervals (same examples as Fig. 1b). b Example comparison of ensemble activity between inter-trial intervals. “Recall” R(t,u) is the similarity between the core population’s similarity matrices in intervals t and u. For the matrices in panel a, the recall is lower following errors than following correct choices. c Example of consistent recall after reinforcement in one session. Left: the recall matrix R for the session, each entry the recall value R(t,u) for inter-trials intervals t and u. The recall matrix is ordered by the outcome of the preceding trial. Below we plot the probability density functions for the distribution of recall values, for the post-error pairs of intervals (bottom-left block diagonal in the recall matrix) and for the post-correct pairs of intervals (top right block diagonal in the recall matrix). Right: the residual recall matrix for the same session, after correction for the effects of interval duration. d The average recall values for post-error and post-correct intervals of the two matrices in c. The distribution of recall in the post-correct intervals was higher than in the post-error intervals (K–S test; recall: P < 0.005; residual recall: P < 0.005; N(correct) = 24×24 = 576; N(error) = 17×17 = 238.) e The difference in average residual recall between the post-correct and post-error intervals, sorted by session type. Each dot is one session. Filled circles indicate a positive difference at P < 0.05 between the distributions of recall values in the post-error and post-correct intervals (Kolmogorov–Smirnov test). Within each session type, two-sided sign tests that the median difference is not zero gave: learning, P = 0.002 (N = 10); rule change, P = 0.72 (N = 8); other sessions, P = 0.37 (N = 32)
