Analyzing the Effect of Hyper-Parameters on a Neural Network for Emotion Detection

Abstract

Human emotion prediction is a tough task. The human face is extremely complex to understand. To build an optimal solution for human emotion prediction model, setting hyper-parameter plays a major role. It is a difficult task to train a neural network. The poor performance of the model can result from poor judgment of sub-optimal hyper- parameters before training the model. This study aims to compare different hyper-parameters and their effect to train the convolutional neural network for emotion detection. We used different methods based on values of validation accuracy and validation loss. The study reveals that SELU activation function performs better in terms of validation accuracy. Swish activation function maintains a good balance between validation accuracy and validation loss. As different combinations of parameters behave differently likewise in optimizers, RMS prop gives less validation loss with Swish whereas Adam performs better with ReLU and ELU activation function.