Fig. 7.

Decoding using timing (through the mixture-of-Poissons model) is more effective than decoding using spike count alone. A, C, Distribution across 29 (A) or 17 (C) neurons of information transmitted by the neuronal responses (spike trains) about which stimulus was shown, using expanding windows starting at stimulus onset. Dark boxes, Spike count only; light boxes, spike count and timing together. B,D, Distribution across 29 (B) or 17 (D) neurons of the percent of trials correctly decoded by guessing the stimulus with highest probability of having elicited the observed response. For comparability, figures are presented as multiples of the percent of trials that would be correctly decoded by chance (B, 1/16 = 6.25%;D, 1/128 = 0.78%). The dotted lineshows the percent of trials correctly decoded by chance. We avoided overfitting using three-way cross-validation: the available trials were divided into three subsets, with two-thirds used to fit the model and the remaining third used to test decoding. Each third of the data was used twice to fit and once to test; the results shown here are averaged over the three sets. Boxes, Median; line and indentation at center, interquartile range:bottom and top. If notchesdo not overlap, corresponding medians are different (p < 0.05). Whiskers have been eliminated, and the dark boxes have been made wider, for visibility.