What is the best approach to learning representations of aesthetics? with Terence Broad
In the mathematics and computing literature, research in aesthetics is for the most part treated as formal (or computational) measures of beauty. More recently in computer vision research, algorithms have been developed to follow the rules of thumb of photography, or use machine learning to predict the aesthetic quality of photographs and paintings. However, aesthetics as a field of research in philosophy and the arts is much broader than simply the visual beauty of images. There are a wide range of aesthetic qualities that could be learned and utilised in games, such as emotional qualities (e.g. bleak), compositional qualities (e.g. balanced) or taste qualities (e.g. garish).
In this talk, I will give a brief overview of the existing approaches that use deep learning to learn representations of aesthetics. I will also detail my own experiments that try to fine-tune generative neural networks to have the aesthetic features of different datasets. I will discuss the limitations of these approaches, and whether it is even possible for contemporary deep learning algorithms to approximate the way humans perceive aesthetics.
To conclude the talk, I will describe some proposed methods for my upcoming research that will try to tackle these issues. Such as exploring new architectures, different sources of data or multi-modal aesthetic representations that could be better utilised for games.