Interpreting CNN Models for Apparent Personality Trait Regression

Abstract

This paper addresses the problem of automatically inferring personality traits of people talking to a camera. As in many other computer vision problems, Convolutional Neural Networks (CNN) models have shown impressive results. However, despite of the success in terms of performance, it is unknown what internal representation emerges in the CNN. This paper presents a deep study on understanding why CNN models are performing surprisingly well in this complex problem. We use current techniques on CNN model interpretability, combined with face detection and Action Unit (AUs) recognition systems, to perform our quantitative studies. Our results show that: (1) face provides most of the discriminative information for personality trait inference, and (2) the internal CNN representations mainly analyze key face regions such as eyes, nose, and mouth. Finally, we study the contribution of AUs for personality trait inference, showing the influence of certain AUs in the facial trait judgments.