EMOTIONET: A Multi-Convolutional Neural Network Hierarchical Approach to Facial and Emotional Classification using TPUs

Abstract

Machine Learning and the tasks of recognition and detection have become a salient topic in recent years. [1] This study is focused on improving the task of using deep convolutional neural networks for facial emotion recognition in the wild with a two-part hierarchical architecture made up of two deep neural networks that are focused on two different tasks. The first task uses a convolutional neural network to classify face or not face. The second task ascertains using another neural network if a face is displaying the emotion within a selected suite of emotions, depicting joyful and sad faces. Factoring the recognition task in this way has produced an observed improvement on previous published work with an overall accuracy of 90% compared to platforms on a CPU cluster and also using TPUs to increase computational power.