Spike rates for all multi-units within a given hemisphere were used as independent features in a linear classifier that decoded which of four task conditions—two objects × two upper/lower locations—was present in each trial. Classification was performed independently on spike rate data from each time point (50 ms window). All reported classification accuracies were obtained via 5-fold cross-validation, in which trials were randomly split into five non-overlapping subsets and each classifier was trained on four of these, while its accuracy was evaluated on the final, untrained one. This process was repeated five times with each subset acting as the test set once, and the final results were averaged across the five folds. The same procedure was used for the cross-classification analysis (Figure 6), except that training and testing trials were selected from different task conditions. For the cross-temporal analysis (Figure S3), cross-classification was also performed using all possible combinations of pairs of time points for training and testing. All decoding analysis was performed with a linear discriminant classifier with optimal covariance shrinkage (Ledoit and Wolf, 2003), using the Python scikit-learn library.
