Learning Deep Neural Networks from Limited Data

Abstract

Constructing deep neural networks (DNNs) from limited data is one of the essential tasks in machine learning. In this talk, we talk about a new learning method using between-class examples to train DNNs from limited supervised data. Then, we introduce several unsupervised domain adaptation methods that transfer knowledge from source to target domain and can reduce a lot of annotation costs in the target domain. We also present a new knowledge transfer method for GANs that can train generative models from extremely small-sized data. Besides, we will briefly introduce various topics that we are working on in our team.

Biography

Tatsuya Harada is a Professor in the Research Center for Advanced Science and Technology at the University of Tokyo. His research interests center on visual recognition, machine learning, and intelligent robot. He received his Ph.D. from the University of Tokyo in 2001. He is also a team leader at RIKEN AIP and a vice director of Research Center for Medical Bigdata at National Institute of Informatics, Japan.