HyDRA: hypergradient data relevance analysis for interpreting deep neural networks.

• Main Authors: Chen, Yuanyuan, Li, Boyang, Yu, Han, Wu, Pengcheng, Miao, Chunyan
• Other Authors: School of Computer Science and Engineering
• Format: Conference or Workshop Item
• Language: English
• Published: 2021
• Subjects: Engineering::Computer science and engineering; Interpretability; Neural Network; Machine Learning; Artificial Intelligence
• Online Access: https://aaai.org/Conferences/AAAI-21/; https://hdl.handle.net/10356/147652
• Institution: Nanyang Technological University

The behaviors of deep neural networks (DNNs) are notoriously resistant to human interpretations. In this paper, we propose Hypergradient Data Relevance Analysis, or HYDRA, which interprets the predictions made by DNNs as effects of their training data. Existing approaches generally estimate data con...