Discrimination of complex form by simple oscillator networks

Abstract

Natural images are rich in higher order spatial correlations. Brain scanning, psychophysics and electrophysiology indicate that humans are sensitive to these image properties. A useful tool for exploring this sense is the set of isotrigon textures. Like natural images these textures have low dimensionality relative to random images, but like random images contain no average structure in their first to third order correlation functions. Thus, the structured appearance of these textures results from higher order correlations. One way to generate the higher order products inherent in higher order correlations is recursive nonlinear processing. We therefore decided to examine if very small oscillator networks could produce a profile of activity that matches human isotrigon discrimination performance across 53 isotrigon texture types. Human performance was measured in 23 subjects. The two best network types found contained as few as 4 oscillators. The input oscillators are of a novel cubic form and the final readout oscillator was a logistic oscillator. Mean readout oscillator activity matched human performance reasonably well even though the network parameters were fixed for all 53 texture types. Overall it appears that relatively simple, short range, and biologically plausible, recursive processing could provide the basis for discrimination of complex form.