Decoding momentary gain variability from neuronal populations

Neural activity in visual cortex is well described by models composed of a deterministic tuning function, a stochastic gain, and a point process. Recent work has shown that in the responses of neurons in the visual cortex of nonhuman primates, variability in the gain is tuned to the statistics of the sensory input. This suggests that gain variability may reflect signal rather than noise. Some theories of neural coding suggest that perceptual uncertainty may be encoded in neural response variability, and that, specifically, gain variability appears well suited to inform downstream uncertainty estimates. However, this notion critically requires that gain variability can be instantaneously read-out in a neurally plausible manner, and previous proposals have required a decoder to have substantial knowledge of the functional organization of the decoded populations. Here, we simulated neural population activity and studied the performance of a heuristic decoder of gain variability, derived from the expected variance of mixture distributions. This decoder exclusively relies on operations of summation, squaring, and division, and requires no knowledge of the selectivity of the input population. With increasing population size, estimates of gain variability increasingly approximate ground truth, and for input populations larger than 100, the decoder’s performance was excellent. We found the decoder to perform robustly at short timescales, across different stimulus strengths, and with realistic levels of correlated spiking variability. Together, our findings show that cross-neuron gain variability can in principle be read-out instantaneously using neurally plausible operations. We conclude that the visual cortex may employ a coding strategy whereby stimulus identity and stimulus uncertainty are encoded independently from each other.