报告时间：2018年9月13日（周四）14:30 - 15:30
报告题目：When weakly-supervised learning meets deep learning
Abstract: In semi-supervised learning (SSL), we have a small set of labeled data coming from all classes based on which we can evaluate the risk, so that all unlabeled (U) data are simply used for regularization purposes. Nevertheless, in many problem settings of weakly-supervised learning (WSL) other than SSL, our labeled data at hand fail to cover all classes and are insufficient to evaluate the risk, so that U data have to be used for risk-evaluation purposes. A typical example is positive-unlabeled (PU) learning, where a binary classifier is trained from only P and U data without any negative data. In this talk, I will raise two fundamental questions in WSL and answer them by using PU learning as a case study. Q1: how to correctly evaluate the risk in WSL? A1: this can be achieved by carefully designed unbiased risk estimators. Q2: when deep learning is involved, is A1 still the right way to go? A2: surprisingly, no! So what should we do when weakly-supervised learning meets deep learning? Please come to my talk to find our solution.
Dr. Gang Niu is currently a research scientist at RIKEN Center for Advanced Intelligence Project. He obtained his master and PhD in computer science from Nanjing University and Tokyo Institute of Technology in 2010 and 2013 respectively. Before joining RIKEN AIP, he was a senior RD engineer at Baidu NLP and then an assistant professor at the University of Tokyo. He has been working on weakly-supervised learning for many years, and has published more than 30 journal articles and conference papers, including 2 JMLR articles, 5 NIPS papers (1 oral presentation), and 7 ICML papers.