Performance in decoding natural visual stimuli using various decompositions of single-trial neural population activity.

Decoding accuracy (% of correct decoding) is shown for spatiotemporal PCA (orange), ICA (red), FA (dark green), Bayes Poisson Factor (black), orthogonal Tucker-2 (magenta), spatiotemporal NMF (cyan) and space-by-time NMF (blue). (A, B) Decoding accuracy as a function of the number of training trials per stimulus averaged over all image datasets (A) or all movie datasets (B). The light green lines show the performance of LDA applied to the binned population spike trains. (C, D) Robustness of the methods with respect to the number of training stimuli. The number of stimuli that were used for training the components/modules is varied on the x-axis and averaged over all image datasets (C) or all movie datasets (D). Performance was evaluated on different stimuli. Lines and shaded areas indicate mean and SEM over all recordings sessions, respectively. In all panels, chance level is at 1/60 = 1.67%.