Modelling and understanding aesthetic preferences for visual patterns, photographs and paintings: Comparing human perceivers with convolutional neural networks

In spite of the wide-spread belief that “beauty is in the eye of the beholder”, recent research in empirical aesthetics has focused on the role of statistical image properties as quasi-universal, biologically rooted factors underlying the preference for some patterns, photographs, and paintings. In the slipstream of the booming area of machine learning (deep neural networks, DNNs, and convolutional neural networks, CNNs), a new field has emerged, computational aesthetics, in which neural networks are trained to predict human aesthetic preference, with a reasonable level of success. The main goal of the present PhD project is to provide a crucial bridge between these two emerging fields of empirical and computational aesthetics. We will do this by building on the recent and ongoing work in the lab to provide a more solid empirical (sometimes psychophysical ) basis regarding the principles of perceptual organization (perceptual grouping and figure-ground organization) that are relevant to aesthetics. We can also build on large-scale online studies of human aesthetics preferences for patterns, photographs, and paintings. One of the main innovative features of the present line of work will be to develop and train Generative Adversarial Networks (GANs) that will be able to generate new images with higher aesthetic value. This will also allow to provide insight into the underlying multidimensional features in the latent space between inputs and outputs, which can be used to formulate hypotheses for additional psychophysical experiments and new online studies in empirical aesthetics.