Extended Data Fig. 8: Definition of coding dimensions.

(a) Mean animal speed during ITI shuttles used for the definition of motion dimensions (n = 1568 ITI shuttles). (b) To define motion dimensions, we computed the average neural activity in the joint subspace over all ITI shuttles and performed PCA. (c) Variance is explained by the first five PCs evaluated for the ITI shuttles used in dimensionality reduction (DR data) and for the ITI shuttles used in decoding analyses (val. data). Mean and 95% CIs for 80 repetitions. (d) PC1 projections for ITI shuttles of individual subjects. While there are differences in the magnitude, the temporal evolution of the projections is highly similar to the one of PC1 obtained using the pooled data (red line in b, Pearson correlation coefficient 0.983 ± 0.010, mean ± s.d. over 12 subjects). (e) Correlation coefficients for pairwise comparisons of time-step decoder weights (mean over 80 repetitions) for avoid versus error decoders from Fig. 3c (right, black line). Especially before action onset, decoder weights show high correlations, indicating a stable representation of avoidance-predictive activity. Based on this finding, we trained a single time-independent decoder for all time steps, whose weights we then used to define the Avoid dimension. (f) Accuracy of avoid versus error decoding per time step for a set of time-dependent decoders trained individually per time step, and one single time-independent decoder trained using data from all time steps (mean over 80 repetitions). The time-independent decoder was separately evaluated with data from individual time steps. The accuracies before action onset are matched between the two settings (time-dependent and time-independent), suggesting that avoidance-predictive activity can be captured using a single decoder. We therefore used the weights of this single time-independent decoder to define avoid dimensions. (g) Decoding accuracy of time-independent decoders for the progressive removal of avoid dimensions (mean and 95% CIs over 80 repetitions). Avoid dimensions are defined by the decoder’s weights vector and are iteratively removed via nullspace projections. Comparison to randomly removed dimensions in blue. (h) Analysis of variability over subjects. To assess decoding accuracies per subject we trained a single decoder using data from all subjects but evaluated it separately using data from individual subjects. The resulting accuracies show differences in magnitude but all follow the same temporal dynamics as time-independent decoder trained on pooled data (the green line in e, Pearson correlation coefficient 0.984 ± 0.012, mean ± s.d. over 12 subjects). This suggests that the effects captured with our joint analysis of all subjects are representative of effects on the single subject level. (i–l) Analogous to e–h for tone decoding. Time-step tone decoders show representational stability (i), performance of time-step decoders can be matched using a single time-independent decoder (j) and the temporal evolution of decoding accuracies is consistent between pooled data and individual subjects (l, Pearson correlation coefficient 0.987 ± 0.015, mean ± s.d. over 12 subjects).