JZUS - Journal of Zhejiang University SCIENCE

Frontiers of Information Technology & Electronic Engineering

Accepted manuscript available online (unedited version)

A survey towards the first principle of explaining DNNs:interactions explain the learning dynamics

Author(s): Huilin ZHOU, Qihan REN, Junpeng ZHANG, Quanshi ZHANG^‡
Affiliation(s): School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
Corresponding email(s): zqs1022@sjtu.edu.cn
Key Words: The first-principal explanation; Theory of equivalent interactions; Two-phase dynamics of interactions; Dynamics of the generalization power of a DNN

Share this article to： More <<< Previous Paper \|Next Paper >>>

Huilin ZHOU, Qihan REN, Junpeng ZHANG, Quanshi ZHANG^‡. A survey towards the first principle of explaining DNNs:interactions explain the learning dynamics[J]. Frontiers of Information Technology & Electronic Engineering,in press.https://doi.org/10.1631/FITEE.2401025

@article{title="A survey towards the first principle of explaining DNNs:interactions explain the learning dynamics",
author="Huilin ZHOU, Qihan REN, Junpeng ZHANG, Quanshi ZHANG^‡",
journal="Frontiers of Information Technology & Electronic Engineering",
year="in press",
publisher="Zhejiang University Press & Springer",
doi="https://doi.org/10.1631/FITEE.2401025"
}

%0 Journal Article
%T A survey towards the first principle of explaining DNNs:interactions explain the learning dynamics
%A Huilin ZHOU
%A Qihan REN
%A Junpeng ZHANG
%A Quanshi ZHANG^‡
%J Frontiers of Information Technology & Electronic Engineering
%P
%@ 2095-9184
%D in press
%I Zhejiang University Press & Springer
doi="https://doi.org/10.1631/FITEE.2401025"

TY - JOUR
T1 - A survey towards the first principle of explaining DNNs:interactions explain the learning dynamics
A1 - Huilin ZHOU
A1 - Qihan REN
A1 - Junpeng ZHANG
A1 - Quanshi ZHANG^‡
J0 - Frontiers of Information Technology & Electronic Engineering
SP -
EP -
%@ 2095-9184
Y1 - in press
PB - Zhejiang University Press & Springer
ER -
doi="https://doi.org/10.1631/FITEE.2401025"

Abstract
Chinese Summary
Academic Network
Reviewer Comment

Abstract: Most explanation methods are designed in an empirical manner, so whether there exists a first-principal explanation of deep neural networks (DNNs) becomes the next core scientific problem in explainable artificial intelligence (XAI). Although it is still an open problem, in this paper, we discuss whether the interaction-based explanation can serve as the first-principal explanation of a DNN. The strong explanatory power of interaction theory comes from the following aspects: (1) it establishes a new axiomatic system to quantify the decision-making logic of a DNN into a set of symbolic interaction concepts; (2) it simultaneously explains various deep learning phenomena, such as generalization power, adversarial sensitivity, representation bottleneck, and learning dynamics; (3) it provides mathematical tools that uniformly explain the mechanisms of various empirical attribution methods and empirical adversarial-transferability-boosting methods; and (4) it explains the extremely complex learning dynamics of a DNN by analyzing the two-phase dynamics of interaction complexity, which further reveals the internal mechanism of why and how the generalization power/adversarial sensitivity of a DNN changes during the learning process.

Darkslateblue:Affiliate; Royal Blue:Author; Turquoise:Article

Reference

Open peer comments: Debate/Discuss/Question/Opinion

<1>

- Go to

Darkslateblue:Affiliate; Royal Blue:Author; Turquoise:Article

Reference