Revealing spatial-frequency channels in an ensemble encoding model of human fMRI

Humans recognize objects by means of a narrow octave-wide spatial-frequency filter called a ``channel''. Behaviorally, this channel can be revealed using critical band masking, a classic psychophysical method that measures sensitivity of recognition accuracy to noise added at each spatial frequency. We want to learn how this channel emerges from the activity of the brain areas involved in object recognition. Critical band masking relies on measured human accuracy in categorizing filtered-noise-perturbed images. To get an analogous accuracy from the brain response, we use N-way Representational Classification Accuracy (N-RCA). For each noise condition used to perturb images in critical band masking, our score measures how often the brain response to the noisy image is more correlated with the response to the original image when compared with responses to $N-1$ other images. This captures how well the brain area's activity informs categorization of a noisy image. We apply the critical band masking paradigm to these accuracies to reveal spatial frequency channels. We then characterize the spatial frequency channel in each visual ROI of an ensemble fMRI encoding model. Our long term goal is to measure channels directly from human fMRI data. Here, we find that the channel bandwidth equals the 1-octave human channel bandwidth and is conserved across model visual ROIs: V1 to V4 and category-selective areas.