Neural variability structure in primary visual cortex is optimal for robust representation of visual similarity

How different neuronal populations construct a robust representation of the sensory world despite neural variability is a mystery. We found that neural variability in mouse primary visual cortex observe a simple rule: For a given sensory stimulus, the mean and the variance of spike counts follow a linear relationship across neurons. To understand how this neural variability structure affects the sensory representation, we artificially varied the slope of the log-mean and log-variance relationship. We found that the intrinsic structure of neural variability allows representations of distinct sensory information to be continuous while minimizing overlap, enabling the neural code to be roust while still being efficient. Further, representational similarity was maximally consistent between different sets of neurons at slope 1, both within and across mice. Thus, the neural variability structure may enable the neocortex to build robust representations of the sensory world, both within and across individuals.