Hyperparameter optimization in CNN for learning-centered emotion recognition for intelligent tutoring systems

摘要

An intelligent tutoring system is used as an efficient self-learning tutor, where decisions are based on the affective state of the user. These detected emotions are what experts call basic emotions and the best-known recognition technique is the recognition of facial expressions. A convolutional neural network (CNN) can be used to identify emotions through facial gestures with very high precision. One problem with convolutional networks, however, is the high number of hyperparameters to define, which can range from a hundred to a thousand. This problem is usually solved by an expert experience combined with trial and error optimization. In this work, we propose a methodology using genetic algorithms for the optimization of hyperparameters of a CNN, used to identify the affective state of a person. In addition, we present the optimized network embedded into an intelligent tutoring system running on a mobile phone. The training process of the CNN was carried out on a PC with a GPU and the trained neural network was embedded into a mobile environment. The results show an improvement of 8% (from 74 to 82%) with genetic algorithms compared to a previous work that utilized a trial and error method.