Emotional Attention for Trained Convolutional Neural Networks

Abstract

In nature, emotions allow for rapid redirection of attention. In the present work, a modular emotional attention approach consisting of two parts is proposed. The attention submodule and the emotional submodule are applied to already trained convolutional neural networks, learning to recognize difficult cases and focus on them. Experiments on a trained CNN-LSTM model for air pollution forecasting show that the proposed approach improves the performance with fast training.