Interpretation Support System with Visualization Interface for Classification Patterns Using Deep Learning Weights

Abstract

In recent years, AI systems based on deep learning have been rapidly increasing their applications.On the other hand, there is a major problem with deep learning: the criteria for prediction and classification by learning are unknown and incomprehensible to humans.To address this problem, research is underway in the field of XAI to understand and evaluate deep learning models.However, at present, XAI only presents some additional information to the output of deep learning, and there is still insufficient system to support the understanding of deep learning models.Therefore, in this research, we construct an interface that extracts and visualizes features that are the basis of deep learning classification as classification patterns based on the weights of deep learning that have already been learned, using a text classification problem as the subject.We will also implement some interpretation support functions to help system users understand the roles of the visualized classification patterns and the meanings of the words in the patterns.In the evaluation experiment of the interpretation support system, we will have subjects interpret a set of sentences, such as product reviews, and evaluate their validity to verify the effectiveness of the interpretation support interface in supporting the interpretation of classification patterns.