Purposeful Regularization with Reinforcement Learning for Facial Expression Recognition In-the-Wild

Abstract

Facial Expression Recognition (FER), an essential aspect of emotion analysis through artificial intelligence, is a crucial research area. Although traditional approaches utilizing Convolutional Neural Networks (CNNs) for analyzing human emotions from facial expressions achieve superior accuracy over conventional machine learning methods, overfitting—especially arising severely from data collected in uncontrolled, In-the-wild settings—significantly impedes CNNs performance. This is due to the data scarcity and inherent noise inside In-the-wild data. To address this challenge, this paper introduces a novel regularization method that employs Reinforcement Learning (RL) to adaptively apply regularization hyperparameters appropriate for the evolving state of trained CNNs. Through experiments on various datasets such as CIFAR100, FER2013, and AffectNet datasets including diverse perspective analysis such as graphical, Grad-CAM and numerical analysis, it is demonstrated that the suggested method can alleviate memorization of noise in training data and promote learning of essential features. The significance of the suggested method lies in its demonstrated remarkable effectiveness in enhancing CNNs’ generalization and reducing overfitting.