Information Theoretic Measure of Stimulus Significance is Not Confounded by Stimulus Correlations or Non-Linearities

Information theory provides tools to analyze neural ensembles that do not depend on assumptions of linearity or correlations in the stimulus ensemble. Here, we use a newly derived information theoretic measure, the stimulus-specific information (SSI), to analyze the important stimuli to simulated visual neurons. Using full-frame flicker stimuli as input to a simulated LGN neuron, we find that the SSI returns the linear kernel of the neuron. For correlated input, linear reconstruction is confounded by correlations in the stimulus. However, we show that the SSI reliably finds the linear kernel of the neuron independent of stimulus correlations. Thus, the SSI is a general tool to discern the important stimuli of neurons, independent of stimulus correlations. It should also be useful for neurons with non-linear response properties