Beyond individual features: Identifying computational principles in perceiving visual beauty

Prior research in empirical aesthetics has related beauty to visual features like curvature, color, or shape. Although these feature-based accounts have identified reliable visual correlates of beauty, the literature on feature preferences is highly fragmented, with individual features studied for individual stimulus domains. Here, we argue that making further progress hinges on complementing feature-based approaches with approaches aimed at distilling computational principles that underlie the perception of beauty. Recent advances in computational neuroscience can drive this new line of research by linking visual beauty to the coding of visual information in artificial neural networks and the biological brain. We highlight two examples of how this approach has identified overarching computational principles. First, metabolic processing costs, both in artificial and biological vision, are inversely related to perceived beauty, suggesting that beauty may reflect energy-efficient, fluent visual coding. Second, integrative processing in artificial and biological vision captures the efficient spatial representation of image elements across space and predicts beauty ratings. Synthesizing these findings, we derive a neuro-computational framework for assessing computational principles in the perception of beauty, which uses artificial neural networks for exploration at scale and then validates candidate principles using human behavior and brain data. We discuss candidate principles that could be explored using this approach, like spatiotemporal predictability, and delineate how the new insights could be reconciled with feature-based approaches. In conclusion, complementing feature-based studies with studies targeting computational principles promises a comprehensive and mechanistic understanding of why certain visual stimuli are perceived as beautiful, and others are not.