Skip to main content
. 2012 Feb 1;54(5-2):49–60. doi: 10.1016/j.visres.2011.12.008

Fig. 3.

Fig. 3

Frequency tiling of receptive fields generated by the BCM, NBCM and ICA algorithms. (A) Fourier-space plot showing the half-maximum contour for each of the BCM receptive fields shown in Fig. 2A. Although 256 receptive fields are represented here, they have very similar Fourier spectra, and so only four different spectral profiles are discriminable. Also, there are noticeable gaps in the orientation coverage of these profiles. (B and C) Fourier-space plots showing the half-maximum contour for the NBCM receptive fields shown in Fig. 2B and C, respectively. Unlike the BCM receptive fields, there is a wide variety of NBCM receptive fields, and, at least for low spatial frequencies, all orientations are well sampled. (D) Receptive fields produced by ICA. Only the first 128 fields are shown, because later fields do not have Gabor-like structure. The ICs cover a range of orientations, though there is a gap at low spatial frequencies. (E) Ability of the standard BCM receptive fields to support reconstruction of sinusoids. Although the reconstruction of horizontal and vertical orientations is reasonable, there are gaps in the coverage of all other orientations. (F and G) In contrast, similar maps for the NBCM (corresponding to B and C, respectively) show that all orientations can be reconstructed accurately. The only gaps in coverage are at high spatial frequencies which were not present in the input images (due to DoG filtering). (H) A similar map for ICA confirms that there is a gap in coverage at low spatial frequencies.