Correlation and decorrelation in natural scenes

David J. Field

Recent work on the study of natural scenes has demonstrated that much of the 'interesting' structure relevant to visual coding is not found in the pair-wise correlations between pixels. In this talk, we look at the insights gained by understanding these pairwise statistics, and consider ways of extracting 'sparse structure' from an image using nonlinear decorrelation techniques. It will be argued that a completely decorrelated solution is not necessarily the most efficient. However, it will be demonstrated that a biologically plausible network which de-correlates thresholded units is capable of exploiting the sparse structure found in natural scenes and it will be shown how the threshold plays a key role in allowing non-orthogonal solutions.

Back to the Natural Scenes Meeting Agenda.