The strategy by which the visual system encodes our environment has long been a topic of debate. One compelling hypothesis is that the tuning characteristics of cells in the visual system provide a sparse representation of natural scenes (Barlow, 1972; Field, 1994). While recent work by Olshausen and Field (1996) supports this hypothesis, other studies suggest alternate explanations (Law and Cooper, 1994; Fyfe and Baddeley, 1995). To address this question, we employ exploratory projection pursuit to investigate the statistical structure of natural scenes—the images the visual system evolved to represent. Applying projection pursuit to over 130,000 natural image patches, we find that searching for sparse and other non-normal structure results in oriented, band-pass, localized projections. Our results suggest that VI simple cell receptive fields directly reflect the statistical structure in natural scenes, consistent with Field's hypothesis (Field, 1994). We relate our technique to that of Olshausen and Field (1996), as well as compare our results to those of similar efforts (Law and Cooper, 1994; Fyfe and Baddeley, 1995). In addition, we demostrate how projection pursuit can be used to investigate non-linear processing found in the visual system, such as on-off channel segregation and receptive fields derived from feed-forward projections.
展开▼
