Deep neural networks generate facial metrics that overcome limitations of previous methods and predict in-person attraction

Abstract

Here we introduce deep neural networks (a form of artificial intelligence) as a novel method for quantifying facial characteristics such as averageness, masculinity, and similarity. Previous methods have quantified facial characteristics using subjective ratings, or objective landmark methods which ignore much of the information we use to perceive faces (e.g. skin colour and contrast, hair, eye colour). We obtained facial images and in-person ratings of facial attractiveness and kindness from 682 speed-dating participants. We find that facial measures derived from neural networks similarly predict in-person ratings compared to facial measures derived from both manual and automatic landmarks. Using neural network-derived measures, we find robust evidence for the attractiveness of masculinity in males, as well as some evidence for assortative preferences for masculinity. Past findings were supported regarding facial similarity as a cue of prosociality. Correlations between neural network and landmark measures were significant but small, and we found that neural network measures captured information beyond face shape. Neural network measures of masculinity had little to no correlation with facial pitch (head tilt) on measures of masculinity, overcoming a major limitation of landmark measures, which were substantially correlated with facial pitch.