부산대학교 도서관

Neural networks are known for their predictive capability, leading to vast applications in various domains. However, the explainability of a neural network model remains enigmatic, especially when the model comes short in learning a particular pattern or features. This work introduces a visual explainable LSTM network framework focusing on temporal prediction. The hindrance to the training process is highlighted by the irregular instances throughout the entire architecture, from input to intermediate layers and output. The framework provides interactive features to support users in customizing and rearranging the structure to obtain different network representations and perform what-if analysis. To evaluate the usefulness of our approach, we demonstrate the application of VixLSTM on the various datasets generated from different domains.