Fig. 2. Decoding performance (multinomial logistic regression) is improved during running relative to stationary periods, independent of net neural tuning or noise correlations.

a Average fraction of correctly classified visual stimuli during running and stationary periods (average over ten 50:50 train/test splits). Each data point is an individual experiment. Colors indicate brain region recorded. b Data from (a) displayed separated into visual regions in dataset, only including experiments in which the difference between running and stationary periods was significant (in either direction). Each dot is an individual experiment. c, d Same as (a), (b) but excluding neurons that increase their activity during running. e, f Same as a. but trial-shuffled to remove noise correlations. g, h Same as a. but excluding neurons that increase their activity during running and trial-shuffled to remove noise correlations. Different decoders, all layers, and differential effect of pupil diameter and running speed are presented in Supplemental Figs. 4–7. Cell and dataset numbers for each region and condition can be found in Supplementary Table 1. All error bars are 95% confidence intervals of mean estimate calculated via 1000 bootstraps. Statistics are calculated via Wilcoxon Sign Rank test, two tailed. Source data are provided as a Source Data file.