Abstract
The information-theoretic approach successfully explains the properties of visual neurons in terms of the exploitation of statistical redundancies. An apparent problem for this view arises from the high number of neurons in the visual cortex, which consdiderably exceeds the number of incoming fibers. With the classical linear filter mechanisms this would lead to an overcomplete representation in which the multi-dimensional tuning functions (the selectivity range in state space) of the neurons would overlap extensively and the responses of the individual neurons would inevitably be highly correlated. We show how this effect can be avoided by nonlinear operations which increase the selectivity in state space as compared to that obtainable with linear mechanisms. We further show that a direct consequence of this nonlinear encoding is the emergence of “non-classical” effects in which the stimulation of “unresponsive” regions leads to context-dependent modifications of the neural response.
Supported by DFG: SFB TR-8 and GRK 267