Figure 5.

Dynamics of spike count prediction throughout the trial. For the GLM predicting time-resolved neural activity additional covariates were included: to account for the temporal structure of firing within trials, we introduced PSTH basis functions, as well as time-varying features for choice history (previous choice, previous target, previous win) and choice on current trial (see Materials and Methods). a, Prediction performance across 75 units when cumulatively adding predictors for the four quartiles of the trial. The gain due to choice history decreases throughout the trial, while the gain due to choice (same trial) increases, as seen in b. The grand mean prediction performance of the time-resolved model is 0.442, very similar to the prediction performance of a model that does not include additional covariates for time-resolved prediction. b, Alternative order of covariates, with choice (same trial) included in the first model.