Predicting Preference for Art Through Low- and High-Level Features
Post by Leanna Kalinowski
What's the science?
We are surrounded by visual art, from classic paintings in a museum to photographs on social media. While navigating through this art-filled world, we constantly make judgements about whether we like or dislike a particular piece. However, the process by which we perceive art is unclear. Do prior experiences with certain features of the piece of art shape our preferences, or are the visual properties of an image more important? The answer is that both are likely important. Computational methods have previously been applied to tease apart how we develop different preferences. However, in the case of visual art, this process is much more challenging due to the visual complexity and variation of some art. This week in Nature Human Behavior, Iigaya and colleagues developed and tested a computational framework to investigate how preferences for visual art are formed.
How did they do it?
The authors first divided the properties of an image into two categories: ‘low-level’ and ‘high-level’. ‘Low-level’ (i.e., bottom-up) features included those derived from an image’s statistics and visual properties, such as hue and brightness, while ‘high-level’ (i.e., top-down) features included those that require human judgement, such as realism and emotion. Participants were asked to report how much they liked various paintings and photographs on a four-point scale, and the authors used these ratings to determine the extent to which they could predict art preferences. They also applied machine learning: a deep convolutional neural network (DCNN) that had been trained for object recognition to predict the pattern by which these visual features emerge when the brain processes visual images.
What did they find?
By engineering a linear feature summation (LFS) model, the authors first observed that visual preference for art can be predicted through a combination of low- and high-level features. This model predicted preferences for both paintings and photographs, suggesting that the features used for driving visual preferences may be universal across different mediums. They also found that their model may represent a biologically plausible computation, as their DCNN model mirrored the results from the LFS model above. Specifically, when the authors did not specify certain features for the DCNN as they did with the LFS model, they found that the DCNN model could learn to predict all of those features on its own.
What's the impact?
The findings here uncover a mechanism through which art preferences can be predicted, shedding light on how these preferences are formed in the brain. These tools have the potential to influence the arts and media industry by predicting which works of art may be more likely to be preferred, and could be extended to predict judgements and perceptions beyond art.
Iigaya et al. Aesthetic preference for art can be predicted from a mixture of how- and high-level visual features. Nature Human Behaviour (2021) Access the original scientific publication here.