The “target images” (M1, M2, M3, 100% A, and 100% B) were followed by a 500 ms blank, after which the names of the two persons of the corresponding stimulus pair were shown and the subject had to indicate which one (s)he perceived with the left/right arrow key (Figure 1A). From the continuous wide-band data, spike detection and sorting were carried out using “Wave_Clus,” an adaptive and stochastic clustering algorithm (Quian Quiroga Gemcitabine in vivo et al., 2004). As in previous works (Quian Quiroga et al., 2009), a response was considered significant if, for the presentation of the “target images”—either

for the 100% A, 100% B (when available), the “recognized A” or “recognized B” presentations (pulling together the responses for the three morphs)—it

fulfilled the following criteria: (1) the firing in the response period (defined as the 1 s interval following the stimulus onset) was significantly larger than in the baseline period (the 1 s preceding stimulus onset) according to a paired t test with p < 0.01; (2) the median number of spikes in the response period was at least 2; (3) the response contained at least five trials (given that the number of BMN 673 order trials in the conditions “recognized A” and “recognized B” was variable). For the average population plots (Figure 3), we normalized each response by the maximum response across conditions (100% A, 100% through B, M1, M2, M3, separated according to the decision: A or B). Statistical comparisons were performed using nonparametric Wilcoxon rank-sum tests (Zar, 1996). A linear classifier was used to decode the subjects’ decision upon the presentation of the ambiguous morphed images (recognized picture A or B) in those cases where we had at least five trials for each decision. One at a time, the firing in each trial was used to test the classifier, which was trained with the remaining trials (all-but-one cross-validation). As in previous works (Quian Quiroga et al., 2007 and Quian Quiroga

and Panzeri, 2009), to evaluate the statistical significance of decoding performance, we used the fact that since the outcomes of the predictions of each decision are independent trials with two possible outcomes, success or failure, the probability of successes in a sequence of trials follows the Binomial distribution. Given a probability p   of getting a hit by chance (p = 1/K  , K  : number of possible decisions), the probability of getting k   hits by chance in n   trials is P(k)=(nk)pk(1−p)n−k, where (nk)=n!(n−k)!k! is the number of possible ways of having k   hits in n   trials. From this, we assessed statistical significance and calculated a p value by adding up the probabilities of getting k   or more hits by chance: p-value=∑j=knP(j). We considered a significance level of p = 0.05, thus expecting 5% of the responses to reach significance just by chance.