To test this idea, we computed the mean classification ratio of all pairs for task-relevant, task-irrelevant, and novel motifs. We indeed found that task relevant motifs exhibit a higher classification ratio

than the task-irrelevant or novel motifs (Friedman test, p = 0.018; Figure 6B), consistent with our observations of the correlation structure. Even in pairs of neurons, therefore, we find that the learning-dependent change in the correlation structure Y-27632 order directly yields improved sensory coding of motifs. How does the correlation-dependent encoding in pairs of neurons translate into encoding by larger populations? Prior theoretical (Gu et al., 2011; Zohary et al., 1994) and experimental (Cohen and Maunsell, 2009) studies have demonstrated that even small changes in average noise correlations can have very large effects on neural encoding in populations as small as only 10 or 20 neurons. Furthermore, in larger populations, noise correlations can have an impact on encoding that is substantially greater than that from mean firing rates (Cohen and Maunsell, 2009; Mitchell et al., 2009). We thus asked whether the changes in correlations that we see in pairs of neurons

yield larger effects in larger populations of neurons. Our data set makes it possible to test this explicitly because many of the pairs in our data set were actually recorded as sets of Olaparib manufacturer up to eight neurons. Metalloexopeptidase Consistent with the idea that larger population sizes allow improved coding from a higher dimensionality of response space, we found that classification performance increased with population size for all classes of motifs (Figure 7A). Importantly,

classification performance increased at a faster rate for task-relevant motifs than for either task-irrelevant or novel motifs (solid lines in Figure 7A). This observation could result either from learning-dependent changes to the underlying single-neuron response properties or from the changes to the correlation structure described above. To distinguish these two sources of increased performance, we compared the classification performance without correlations (i.e., with trials shuffled, which does not alter individual neuron responses) to that with correlations intact. Shuffling trials considerably reduces classification performance for task-relevant motifs, but not to the level of task-irrelevant or novel motifs (dashed lines in Figure 7A). This suggests that the enhanced coding fidelity for task-relevant motifs results both from single-neuron response properties and from correlations between neurons. To isolate the effects of correlations on coding, we computed the classification ratio for each class of motif and for each population size (Experimental Procedures).