Fig. 2.

Receptive fields generated by the standard and contrast-normalized BCM algorithms, with Independent Components Analysis (ICA) for comparison. (A) The standard BCM algorithm, when applied to natural images, generates receptive fields containing patterns of parallel light and dark bars. Each receptive field is reminiscent of the tuning found in the simple cells of mammalian V1. However, when viewed as a population, many of the BCM receptive fields are very similar to one another, so that the same visual information (primarily horizontal and vertical structure) is represented redundantly by many receptive fields. (B) The NBCM algorithm is similar to the standard BCM, but incorporates contrast normalization. This results in a wider variety of receptive fields, and much lower redundancy (here, α = 1 and β = 2). (C) With a different choice of parameters (α = 2, β = 4), the fields have high spatial frequencies, but similar structure. (D) Receptive fields produced by ICA of the same images cover a range of positions and orientations, but are of generally high spatial frequency. (E–H) Fourier power spectra of the receptive fields in (A–D). The uniformity of the standard BCM receptive fields (D) is particularly noticeable in their power spectra.