On Network Science and Mutual Information for Explaining Deep Neural Networks

• Published in: ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 8399 - 8403
• Main Authors: Davis, Brian, Bhatt, Umang, Bhardwaj, Kartikeya, Marculescu, Radu, Moura, Jose M.F.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-05-2020
• Subjects: Deep learning; information theory; interpretability; Mutual information; network science; Neural networks; Neurons; Predictive models; Signal processing; Speech processing
• ISSN: 2379-190X
• DOI: 10.1109/ICASSP40776.2020.9053078

In this paper, we present a new approach to interpret deep learning models. By coupling mutual information with network science, we explore how information flows through feedforward networks. We show that efficiently approximating mutual information allows us to create an information measure that quantifies how much information flows between any two neurons of a deep learning model. To that end, we propose NIF, Neural Information Flow, a technique for codifying information flow that exposes deep learning model internals and provides feature attributions.