Direct perception of affective valence from vision

Affect has been posited as the basic subjective feeling component underlying all experience, arising directly from internal states of the body and the closely related proximal senses. By contrast, in the distal senses such as vision, affect is thought to be indirect, largely mediated by higher level processes. Evidence from machine learning suggests affective features may be embedded in the ecological statistics of the external environment but their causal role, if any, remains unclear. Here we provide evidence from machine and human observers that affective valence can be decoded directly from visual features. A visual valence (VV) model of low-level image statistics trained to predict Normative Valence (NV) in 8000 emotionally charged images transferred even more robustly to predicting valence of abstract paintings without conceptual content. Manipulating VV and availability of conceptual analysis, enhanced the contribution of VV to valence experience. In the brain, in contrast with NV, VV resided exclusively in early and mid-level visual areas. Employing a deep generative network, brain activity in these regions synthesized new images containing predicted positive versus negative VV. There are distinct competing modes determining valence experience, one indirectly from meaning, and the other derived from ecological visual statistics, affording direct perception of valence as an apparent objective property of the external world.