Abstract
Three recent papers based on the Natural Scenes Dataset (Allen et al, 2022) reported that two bands of cortex in the ventral visual stream respond selectively to images of food (Khosla et al, 2022; Jain et al, 2022; Pennock et al, 2023). Khosla et al (2022) applied data-driven non-negative matrix factorization (NMF) to the NSD data and discovered a consistent component across participants, defined by a response profile over stimuli and a weight matrix over voxels, that correlated strongly with the salience of food in the image. Control analyses indicated that this “food component” responded more strongly to food than nonfood images matched for low-level visual features. Here we further tested the selectivity of this “food component” with new controlled stimuli in six new subjects. We developed a component localizer using a subset of 50 NSD images which minimize the variance of the inferred food component weights in new subjects. We used this localizer to infer the food component weights in new subjects, and then computed the response of this inferred food component to independent stimuli. Our findings replicated the broad anatomical localization of the food component, as well as the previously published findings of a greater response to food than nonfood in NSD images held out from the localizer. We further found that responses to these images were very similar when presented in greyscale. We also found a significantly higher response of this component to food than nonfood “reachspace” images (e.g., tabletops), though with a smaller effect size. However, we did not find a higher response to food than nonfood in “cutout” images, in which food and nonfood objects were isolated on a white background. These results suggest that the “food component” is not broadly selective to all food images, but its selectivity may depend on surrounding context.