Experimental approach.

a, For each rat, a Neuropixels recording probe was implanted using a custom designed assembly that allowed use of a strain-relieving attachment for the tether. b, The probe was positioned to record simultaneously from both FOF and ADS, with the schematic showing the coronal plane. The right side shows traces from example sites recorded across both regions simultaneously. c, Task schematic.

Behavioral performance.

a, Combined hit rates as a function of change magnitude. Hit rate was calculated excluding false alarm trials (i.e. only including trials in which the rat was presented with a change). b, Reaction times for hits measured as the time of center port withdrawal following a change. c, Psychophysical reverse correlation (PRC). Calculated as the average local click rate preceding false alarms. For all plots, error bars and error shading show SE.

Retrospective temporal decoding from stimulus onset.

a, Example decoding “confusion matrix” derived from ADS units. Color bar indicates number of trials (out of 169) classified into each bin comparing predicted to true time from stimulus onset. b, Decoding confusion matrix from FOF units in the same session. c, Average root-mean squared error (RMSE) of temporal decoding for each session as a function of the total number of units in that session, separated by region. Darker points correspond to the example session shown in earlier panels. N = 61 sessions. d, ADS temporal decoding RMSE as a function of time for each analyzed session (gray traces). Magenta trace shows example session from earlier panels. Red trace shows average RMSE across sessions. e, FOF temporal decoding RMSE as a function of time for each analyzed session (gray traces). Magenta trace shows example session from earlier panels. Blue trace shows average RMSE across sessions. f, Session by session comparison of RMSE for FOF versus ADS. Magenta point shows example session from earlier panels. Cyan point shows mean and SEM across sessions. Only sessions that meet criteria for balance between ADS and FOF units are included in panels d-f. N = 16 sessions.

Prospective temporal decoding to movement initiation.

a, Example decoding “confusion matrix” derived from ADS units. Color bar indicates number of trials (out of 214) classified into each bin comparing predicted to true time from stimulus onset. b, Decoding confusion matrix from FOF units in the same session. c, Average root-mean squared error (RMSE) of temporal decoding for each session as a function of the total number of units in that session separated by region. Darker points correspond to the example session shown in earlier panels. N = 61 sessions. d, ADS temporal decoding RMSE as a function of time for each analyzed session (gray traces). Magenta trace shows example session from earlier panels. Red trace shows average RMSE across sessions. e, FOF temporal decoding RMSE as a function of time for each analyzed session (gray traces). Magenta trace shows example session from earlier panels. Blue trace shows average RMSE across sessions. f, Session by session comparison of RMSE for FOF versus ADS. Magenta point shows example session from earlier panels. Cyan point shows mean and SEM across sessions. Only sessions that meet criteria for balance between ADS and FOF units are included in panels d-f. N = 16 sessions.

Retrospective temporal encoding from stimulus onset in ADS.

a, Example session showing each unit plotted in the space of the first two principal components of the decoder weights. Each point corresponds to one unit. b, z-scored firing rate of all units from this session aligned to stimulus start. The vertical line corresponds to the end of the decoding epoch for all panels showing neural dynamics. c, z-scored firing rate of all units from all sessions aligned to stimulus start. d, Same as panel a, but with units sorted into quintiles along the first principal component. e, z-scored firing rate of all units from this session sorted into quintiles along the first principal component and aligned to stimulus start. f, z-scored firing rate of all units from all sessions sorted into quintiles along the first principal component and aligned to stimulus start. g, Same as panel a, but with units sorted into quartiles along the second principal component. h, z-scored firing rate of all units from this session sorted into quartiles along the second principal component and aligned to stimulus start. i, z-scored firing rate of all units from all sessions sorted into quartiles along the second principal component and aligned to stimulus start. j, Same as panel a, but with units sorted into six groups along both of the first two principal components. k, z-scored firing rate of all units from this session sorted into six groups along both of the first two principal components and aligned to stimulus start. l, z-scored firing rate of all units from all sessions sorted into six groups along both of the first two principal components and aligned to stimulus start. N = 16 sessions.

Retrospective temporal encoding from stimulus onset in FOF.

a, Example session showing each unit plotted in the space of the first two principal components of the decoder weights. Each point corresponds to one unit. b, z-scored firing rate of all units from this session aligned to stimulus start. The vertical line corresponds to the end of the decoding epoch for all panels showing neural dynamics. c, z-scored firing rate of all units from all sessions aligned to stimulus start. d, Same as panel a, but with units sorted into quintiles along the first principal component. e, z-scored firing rate of all units from this session sorted into quintiles along the first principal component and aligned to stimulus start. f, z-scored firing rate of all units from all sessions sorted into quintiles along the first principal component and aligned to stimulus start. g, Same as panel a, but with units sorted into quartiles along the second principal component. h, z-scored firing rate of all units from this session sorted into quartiles along the second principal component and aligned to stimulus start. i, z-scored firing rate of all units from all sessions sorted into quartiles along the second principal component and aligned to stimulus start. j, Same as panel a, but with units sorted into six groups along both of the first two principal components. k, z-scored firing rate of all units from this session sorted into six groups along both of the first two principal components and aligned to stimulus start. l, z-scored firing rate of all units from all sessions sorted into six groups along both of the first two principal components and aligned to stimulus start. N = 16 sessions.

Prospective temporal encoding to movement initiation in ADS.

a, Example session showing each unit plotted in the space of the first two principal components of the decoder weights. Each point corresponds to one unit. b, z-scored firing rate of all units from this session aligned to movement onset. The vertical line corresponds to the end of the decoding epoch for all panels showing neural dynamics. c, z-scored firing rate of all units from all sessions aligned to movement onset. d, Same as panel a, but with units sorted into quintiles along the first principal component. e, z-scored firing rate of all units from this session sorted into quintiles along the first principal component and aligned to movement onset. f, z-scored firing rate of all units from all sessions sorted into quintiles along the first principal component and aligned to movement onset. g, Same as panel a, but with units sorted into quartiles along the second principal component. h, z-scored firing rate of all units from this session sorted into quartiles along the second principal component and aligned to movement onset. i, z-scored firing rate of all units from all sessions sorted into quartiles along the second principal component and aligned to movement onset. j, Same as panel a, but with units sorted into six groups along both of the first two principal components. k, z-scored firing rate of all units from this session sorted into six groups along both of the first two principal components and aligned to movement onset. l, z-scored firing rate of all units from all sessions sorted into six groups along both of the first two principal components and aligned to movement onset. N = 16 sessions.

Prospective temporal encoding to movement initiation in FOF.

a, Example session showing each unit plotted in the space of the first two principal components of the decoder weights. Each point corresponds to one unit. b, z-scored firing rate of all units from this session aligned to movement onset. The vertical line corresponds to the end of the decoding epoch for all panels showing neural dynamics. c, z-scored firing rate of all units from all sessions aligned to movement onset. d, Same as panel a, but with units sorted into quintiles along the first principal component. e, z-scored firing rate of all units from this session sorted into quintiles along the first principal component and aligned to movement onset. f, z-scored firing rate of all units from all sessions sorted into quintiles along the first principal component and aligned to movement onset. g, Same as panel a, but with units sorted into quartiles along the second principal component. h, z-scored firing rate of all units from this session sorted into quartiles along the second principal component and aligned to movement onset. i, z-scored firing rate of all units from all sessions sorted into quartiles along the second principal component and aligned to movement onset. j, Same as panel a, but with units sorted into six groups along both of the first two principal components. k, z-scored firing rate of all units from this session sorted into six groups along both of the first two principal components and aligned to movement onset. l, z-scored firing rate of all units from all sessions sorted into six groups along both of the first two principal components and aligned to movement onset. N = 16 sessions.

Temporal trajectory of PCA loadings from decoder weights.

a, PCA loadings as a function of time in ADS for each session. Color indicates time from stimulus start. b, PCA loadings as a function of time in FOF for each session. Color indicates time from stimulus start. c, PCA loadings as a function of time in ADS for each session. Color indicates time from movement onset. d, PCA loadings as a function of time in FOF for each session. Color indicates time from movement onset. In all plots, darker points correspond to the loadings for the example session from earlier figures. N = 16 sessions.

Regional comparisons of mean PCA loading trajectories across time.

a, Comparison of loading on PC1 for FOF versus ADS aligned to stimulus start. b, Comparison of loading on PC2 for FOF versus ADS aligned to stimulus start. For both panels a and b, color of each point corresponds to time from stimulus start. c, Comparison of loading on PC1 for FOF versus ADS aligned to movement onset. d, Comparison of loading on PC2 for FOF versus ADS aligned to movement onset. For both panels c and d, color of each point corresponds to time from movement onset. For all plots, error bars show SEM. N = 16 sessions.

Dimension of neural population activity across trial.

a, Schematic of locally-linear approach to characterizing dimension. PCA is used to find subspaces of high variance at different points in the task. The dimension is then calculated using the participation ratio, which measures the effective number of modes needed to capture variance. Schematic captures activity becoming lower-dimensional over the course of trial. b, Dimension of activity in late evidence evaluation against dimension in early evaluation, capturing stability of subspaces. c, Dimension of activity during decision commitment movement initiation against dimension during late evidence evaluation, capturing decrease of dimension. Sizes of points show numbers of units in each recording session.

Geometry of neural population activity across trial.

a, Left: Schematic of locally-linear approach to characterizing geometry via the alignment of subspaces of high variance at different points in the trial. Schematic shows the subspace of high variance rotating over time. Right: An alignment index is calculated by comparing how well principal component vectors (PCs) from one epoch capture population activity in another epoch when compared to PCs from that epoch. The alignment index is the area of the gap between the cumulative fraction of variance explained by each set of PCs (shared area) normalized by the total area under the higher curve (see main text and Methods for more details). b, Left: Cumulative fraction of variance of ADS activity during early evidence evaluation explained by early evidence PCs (top curve) and by late evidence evaluation PCs (bottom curve) for an example session. Inset shows similar plot but for activity during late evidence evaluation. Right: As in the left plot but showing movement vs. late evidence evaluation. c, As in (b) but for FOF activity. d, Scatter plot of alignment index of ADS activity between late evidence evaluation and movement against alignment index between early evidence evaluation and late evidence evaluation, showing significantly larger change upon decision commitment movement. e, As in (d) but for FOF activity. f, Scatter plot of alignment between early evidence evaluation and late evidence evaluation in FOF against corresponding alignment for ADS, showing that subspaces are similarly stable across evidence evaluation. g, Scatter plot of alignment between late evidence evaluation and movement in FOF against corresponding alignment for ADS, showing that change in FOF is larger. Sizes of points show numbers of units in each recording session.