Fig. 4: A reward-prediction signal in supervised training only.

a, Raster plot of simultaneously recorded neurons, over a period of approximately 2.5 min, with behavioural annotations. Rastermap was used to sort neurons across the y axis. b, Zoomed-in view of a vertical segment in Rastermap that corresponds to a group of neurons active in the leaf1 corridor. c, Spatial distribution of selected neurons. d, Population average response for the neurons selected in panel b across trials of circle1 and leaf1, sorted by sound cue position. e, A discrimination index (\({d}_{{\rm{late\; vs\; early}}}^{{\prime} }\)) was constructed to compare trial-averaged neuron responses between trials with early versus late rewards. EC, early cue; LC, late cue. f, Distribution of reward-prediction neurons (\({d}_{{\rm{late\; vs\; early}}}^{{\prime} }\ge 0.3\)) in task and unsupervised mice, before and after learning. g, Percentage of reward-prediction neurons in the anterior region. h, Responses of a single selective neuron (top) or the entire population (bottom) across trials in the test 1 session. i, Reward-prediction neuron activity and lick rates in the leaf1 corridor, aligned to the sound cue in the test 1 session, averaged across mice (n = 5). Dashed line indicates 0. j, Same as panel i but aligned to the first lick in the corridor for trials in which the first lick was after 2 m (n = 4 mice). k, Average activity of reward-prediction neurons inside the leaf2 corridor on trials with or without licks in the test 1 session (n = 4 mice). l, Same as panel k but for the leaf1-selective neurons in the medial regions (defined as in Fig. 1h). m,n, Same as panel h (bottom) for test 2 (m) and test 3 (n) sessions. For panels d, h, m and n, the dashed line indicates end of the corridor. All data are mean ± s.e.m. *P < 0.05, **P < 0.01 and not significant (NS) from paired, two-sided t-tests.