Trial-averaged metrics, e.g. in the form of tuning curves and population response vectors, are a basic and widely accepted way of characterizing neuronal activity. But how relevant are such trial-averaged responses to neuronal computation itself? Here we present a simple test to estimate whether average responses reflect aspects of neuronal activity that contribute to neuronal processing in a specific context. The test probes two assumptions inherent in the usage of average neuronal metrics: 1) Reliability: Neuronal responses repeat consistently enough across single stimulus instances that the average response template they relate to remains recognizable to downstream regions. 2) Behavioural relevance: If a single-trial response is more similar to the average template, this should make it easier for the animal to identify the correct stimulus or action. We apply this test to a large publicly available data set featuring electrophysiological recordings from 42 cortical areas in behaving mice. In this data set, we show that single-trial responses were less correlated to the average response template than one would expect if they simply represented discrete versions of the template, down-sampled to a finite number of spikes. Moreover, single-trial responses were barely stimulus-specific — they could not be clearly assigned to the average response template of one stimulus. Most importantly, better-matched single-trial responses did not predict accurate behaviour for any of the recorded cortical areas. We conclude that in this data set, average responses do not seem particularly relevant to neuronal computation in a majority of brain areas, and we encourage other researchers to apply similar tests when using trial-averaged neuronal metrics.
Does the brain care about averages? A simple test.
