A sparse set of spikes corresponding to reliable correlations is highly informative of visual stimulus on single trials

Spike trains in cortical neuronal populations vary in number and timing trial-to-trial, rendering a viable single trial coding scheme for sensory information elusive. Correlations between pairs of neocortical neurons can be segmented into either sensory or noise according to their stimulus specificity. Here we show that pairs of spikes, corresponding to reliable sensory correlations in imaged populations in layer 2/3 of mouse visual cortex are particularly informative of visual stimuli. This set of spikes is sparse and exhibits comparable levels of trial-to-trial variance relative to the full spike train. Despite this, correspondence of pairs of spikes to a specific set of sensory correlations identifies spikes that carry more information per spike about the visual stimulus than the full set or any other matched set of spikes. Moreover, this sparse subset is more accurately decoded, regardless of the decoding algorithm employed. Our findings suggest that consistent pairwise correlations between neurons, rather than first-order statistical features of spike trains, may be an organizational principle of a single trial sensory coding scheme.