报告时间：2018年9月13日（周四）16:30 - 17:30
报告题目：Robust Deep Learning with Noisy Labels
邀请人： 宫辰 教授
It is challenging to train deep neural networks robustly with noisy labels, as the capacity of deep neural networks is so high that they can totally overfit on these noisy labels. In this talk, I will introduce two advanced and orthogonal techniques in deep learning with noisy labels, namely "training on selected samples" and "estimating the noise transition matrix". (1) Motivated by the memorization effects of deep networks, which shows networks fit clean instances first and then noisy ones, we present a new paradigm called "Co-teaching" combating with noisy labels. We train two networks simultaneously. First, in each mini-batch data, each network filters noisy instances based on memorization effects. Then, it teaches the remained instances to its peer network for updating the parameters. (2) As noisy labels are corrupted from ground-truth labels by an unknown noise transition matrix, the accuracy of classifiers can be improved by estimating this matrix. However, such estimation is often inexact, which inevitably degenerates the accuracy of classifiers. The inexact estimation is due to either a heuristic trick, or the brutal-force learning by deep networks under a finite dataset. Thus, we present a human-assisted approach called "masking". The masking conveys human cognition of invalid class transitions, and naturally speculates the structure of the noise transition matrix. Given the structure information, we only learn the noise transition probability to reduce the estimation burden.
Bo Han is pursuing his Ph.D. degree under the supervision of Prof. Ivor W. Tsang at Centre for Artificial Intelligence, University of Technology Sydney, Australia. He is currently a visiting student working with Prof. Masashi Sugiyama and Dr. Gang Niu at Center for Advanced Intelligence Project, RIKEN, Japan. His current research interests lie in machine learning and its application. Specifically, his long-term goal is to develop intelligent systems, which learn from complex (uncertain, massive, interactive, and private) data.