Personalized machine learning for facial expression analysis

Abstract

For this MEng Thesis Project, I investigated the personalization of deep convolutional networks for facial expression analysis. While prior work focused on population-based ("one-size-fits-all") models for prediction of affective states (valence/arousal), I constructed personalized versions of these models to improve upon state-of-the-art general models through solving a domain adaptation problem. This was done by starting with pre-trained deep models for face analysis and fine-tuning the last layers to specific subjects or subpopulations. For prediction, a "mixture of experts" (MoE) solution was employed to select the proper outputs based on the given input. The research questions answered in this project are: (1) What are the effects of model personalization on the estimation of valence and arousal from faces? (2) What is the amount of (un)supervised data needed to reach a target performance? Models produced in this research provide the foundation of a novel tool for personalized real-time estimation of target metrics.