Variance Normalization by Adaptive Rescaling of Neural Responses
Adaptation is a widespread property of nervous systems, which provides flexibility to function under varying external conditions. Here we relate an adaptive property of a sensory neuron directly to an optimization principle of information theory. We show that the response function (input-output relation) of a neuron in a dynamical environment changes with the statistical properties of this environment. In particular, at high signal to noise ratio the system can ``learn'' the standard deviation of the distribution from which sensory stimuli are drawn, and use it as the unit for measuring fluctuations. This property is expected to be particularly useful in processing natural signals whose statistics are non-stationary or intermittent. The rescaling of the response function in terms of the standard deviation is observed in two different regimes of the neural code, suggesting some universality of the effect. We provide direct experimental evidence that the rescaling degree of freedom enables the system to match its response to the statistics of the environment and maximize information transmission.