Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Humans and variational autoencoders agree: attractive faces are average and feminine, not symmetric and young

Francisco M. Lopez¹, Jochen Triesch²; ¹Frankfurt Institute for Advanced Studies, ²Goethe University

Presenter: Francisco M. Lopez

Why do we find some faces more attractive than others? Four properties are frequently identified as sources of beauty: symmetry, youth, femininity, and averageness. Recent experiments show that only the latter two determine facial attractiveness. Here, we test whether a neural network trained with unsupervised learning can reproduce and explain this phenomenon. We train a variational autoencoder (VAE) on face images and estimate its preference judgments as the compression of a face in the latent space. We find that, like humans, the VAE’s preferences are significantly correlated with the averageness and the femininity of the faces but not their symmetry or youth. Furthermore, the VAE correlates with human attractiveness judgments. In sum, this work suggests that human aesthetic face preferences can be explained by the efficiency with which a face can be encoded.

Topic Area: Object Recognition & Visual Attention

Extended Abstract: Full Text PDF