Abstract
The early visual system is hypothesized to reduce the redundancy inherent to natural images. To test this theory, Independent Component Analysis (ICA) has been applied to monochromatic natural images, yielding independent component (IC) filters, which maximally reduce redundancy. The receptive fields of these ICs strongly resemble V1 receptive fields, suggesting that redundancy reduction is a good model for V1. ICA has more recently been applied to colored natural images. The resulting ICs appear chromatically similar to visual neurons (Hoyer & Hyvarinen 2000, Wachtler, Lee & Sejnowski 2001). However, no quantitative analysis of the relationship between color sensitivity and luminance sensitivity has been done. We do so by ‘stimulating’ ICs with colored sinusoidal gratings of differing contrast, comparing the results against physiological data (Lennie et al. 1990). We find that the color tuning of ICs strongly resembles the color tuning of V1 simple cells, but not the unoriented cells found in V1 color blobs. Both simple cells and ICs are dominated by red-green opponent filters and blue-yellow opponent filters. Red-green opponency is not perfectly balanced, and so those filters respond more strongly to changes in luminance than purely chromatic changes. Many blue-yellow opponent filters, however, show significant response to purely chromatic changes. These properties are observed in both simple cells and ICs, providing further evidence for the view that simple cells are redundancy reducing filters. The mismatch between the color tuning of ICs and unoriented blob cells suggests that V1 color analysis cannot be entirely modelled by straightforward ICA. This suggests that blob cells may be part of a separate pathway that does not simultaneously reduce spatial and chromatic redundancy, but perhaps only local color redundancy.
HoyerP.O.HyvarinenA.(2000), Network 11:191–210.
WachtlerT.LeeT.W.SejnowskiT.J.(2001), J. Opt. Soc. Am. A, 18:65–77.
LennieP.KrauskopfJ.SclarG.(1990), J. Neurosci. 10:649–669.