Subjective feelings are thought to arise from conceptual and bodily states. We examine whether the valence of feelings may also be decoded directly from objective ecological statistics of the visual environment. We train a visual valence (VV) machine learning model of low-level image statistics on nearly 8000 emotionally charged photographs. The VV model predicts human valence ratings of images and transfers even more robustly to abstract paintings. In human observers, limiting conceptual analysis of images enhances VV contributions to valence experience, increasing correspondence with machine perception of valence. In the brain, VV resides in lower to mid-level visual regions, where neural activity submitted to deep generative networks synthesizes new images containing positive versus negative VV. There are distinct modes of valence experience, one derived indirectly from meaning, and the other embedded in ecological statistics, affording direct perception of subjective valence as an apparent objective property of the external world.
© 2024. The Author(s).