Formulating Generalizable and Non-Generalizable Interactions in DNNs

Lei Cheng, Junpeng Zhang, Qihan Ren, Yuxuan He, Quanshi Zhang

Keywords: Overfitting; Generalization; Deep Learning Theory

TL;DR: This paper disentangles the distributions of generalizable interactions and non-generalizable interactions from a trained DNN.

Abstract: This paper aims to analyze the generalization power of deep neural networks (DNNs) from the perspective of interactions. Unlike previous analyses of a DNN’s generalization power in a high-dimensional feature space, we find that the generalization power of a DNN can be explained as the generalization power of the interactions. We find that generalizable interactions follow a decay-shaped distribution, while non-generalizable interactions follow a spindle-shaped distribution. Furthermore, we develop a method to disentangle these two types of interactions in a DNN. We have verified that the theoretically disentangled distributions of generalizable interactions and non-generalizable interactions can well match the real distributions in experiments.

Primary Area: interpretability and explainable AI

Submission Number: 3161