This variability was due to large trial-to-trial variations in the response of most individual neurons (Bartho et al., 2009; Hromádka et al., 2008). Despite the variability, we were able to observe

that sound intensity and identity modulated the probability of observing a population event. Tuning to pure tones could be seen at both the single neuron (Figures 2B–2D) and at the population level (Figure 2A). However, prediction of the population firing rate in response to complex sounds by a linear model based on the observed pure tone tuning was poor (Figure S2). Therefore, local populations encode sounds in a nonlinear fashion, as was reported for single neurons (Machens et al., 2004). This implies that pure tone tuning alone is not sufficient to describe sound representation in the auditory cortex. We therefore decided to use a more general framework to investigate MK-1775 mw the coding properties of local response patterns in single trials (Bathellier et al., 2008). In most local populations, we made the striking observation that despite the high

variability of response patterns, the most reliable part of the pattern seemed to be common to very different sounds GS-1101 nmr such as the different pure tones shown in Figure 2A. This suggested to us that sound evoked responses in local auditory cortex networks are constrained to a limited repertoire of functional patterns superposed on high trial-to-trial stochasticity. To obtain a quantitative account of the limited repertoire of functional patterns in the face of large variability, we systematically quantified the similarity of local response patterns elicited by large arrays of short (50–70 ms) pure tones and complex

sounds. To do so, we used a similarity metric designed to obtain an intuitive readout of single trial response separability. In short, the similarity between two sound-evoked responses was defined as the average of all pairwise correlations between the single trial response patterns of the two sounds (see matrices of single trial correlations, Figure 3A). This similarity metric was compared to a response reliability metric, which was the average of all pairwise correlations between all the single trial response Rebamipide patterns of one given sound. This reliability metric gave us a quantitative readout of the trial-to-trial variability in response to a given sound. Using these two metrics, the idea is that if the response patterns to two sounds have a lower similarity than their respective reliabilities, they will likely be discriminable on a single trial basis by an external observer. If not, the patterns can be thought to be the same. Pairwise response similarities were displayed in color-coded matrix plots after ordering the sounds with a hierarchical clustering algorithm to reveal potential underlying structures in the space of response patterns ( Figures 3B and 3C).