Looking at the world often involves not just seeing things, but feeling things. Modern feedforward machine vision systems that learn to perceive the world in the absence of active physiology, deliberative thought, or any form of feedback that resembles human affective experience offer tools to demystify the relationship between seeing and feeling, and to assess how much of visually evoked affective experiences may be a straightforward function of representation learning over natural image statistics. In this work, we deploy a diverse sample of 180 state-of-the-art deep neural network models trained only on canonical computer vision tasks to predict human ratings of arousal, valence, and beauty for images from multiple categories (objects, faces, landscapes, art) across two datasets. Importantly, we use the features of these models without additional learning, linearly decoding human affective responses from network activity in much the same way neuroscientists decode information from neural recordings. Aggregate analysis across our survey, demonstrates that predictions from purely perceptual models explain a majority of the explainable variance in average ratings of arousal, valence, and beauty alike. Finer-grained analysis within our survey (e.g. comparisons between shallower and deeper layers, or between randomly initialized, category-supervised, and self-supervised models) point to rich, preconceptual abstraction (learned from diversity of visual experience) as a key driver of these predictions. Taken together, these results provide further computational evidence for an information-processing account of visually evoked affect linked directly to efficient representation learning over natural image statistics, and hint at a computational locus of affective and aesthetic valuation immediately proximate to perception.
