Interpretability research of deep learning: A literature survey

Abstract

Deep learning (DL) has been widely used in various fields. However, its black-box nature limits people's understanding and trust in its decision-making process. Therefore, it becomes crucial to research the DL interpretability, which can elucidate the model's decision-making processes and behaviors. This review provides an overview of the current status of interpretability research. First, the DL's typical models, principles, and applications are introduced. Then, the definition and significance of interpretability are clarified. Subsequently, some typical interpretability algorithms are introduced into four groups: active, passive, supplementary, and integrated explanations. After that, several evaluation indicators for interpretability are briefly described, and the relationship between interpretability and model performance is explored. Next, the specific applications of some interpretability methods/models in actual scenarios are introduced. Finally, the interpretability research challenges and future development directions are discussed.