循环神经网络的可视化解释

Recurrent neural networks have been successfully applied to various natural language processing tasks, and have achieved better results than conventional methods. However, the lack of understanding of the mechanism behind its effectiveness limits the further improvement of its architecture. In this article, in order to clearly explain the RNN classification model with visualization method, we improve the model. By outputting the intermediate prediction results at each time step, we visualize the changes in the prediction results over time and evaluate the importance of the input. And in the backpropagation mechanism, backpropagation through the output results of each step can better explain the importance of the input.