The seminar by T.Ide at IBM: Decentralized multi-task learning and data privacy

Speaker: Dr. Tsuyoshi Ide (IBM Thomas J. Watson Research Center)

Since the advent of Bitcoin, cryptocurrencies with distributed and decentralized architectures have been creating a lot of excitement in the information technology industry. Although machine learning under the distributed setting has a long history of research, we revisit the topic from a perspective of decentralized architecture.
In this talk, we will share our recent work on decentralized secure multi-task density estimation [1]. Multi-task learning is to build a better model through knowledge sharing in a sense. On the other hand, data privacy is almost always a critical requirement in the distributed setting. The question is if we can meet those two conflicting goals at once. We show that distributed inference of a mixture of the exponential family can be decentralized by combining the idea of multi-agent coordination. We also discuss an approach of privacy preservation and an interesting connection to algebraic graph theory.
[1] Ide et al., Efficient Protocol for Collaborative Dictionary Learning in Decentralized Networks, IJCAI 19.

Bio
Dr. Tsuyoshi ide ("Ide-san") is a Senior Technical Staff Member with IBM T. J. Watson Research Center, New York, USA. He received his Ph.D. from the University of Tokyo in condensed matter physics in 2000. Since around 2003, he has been working on data mining and machine learning research through a variety of real-world applications. Currently, he is part of the Trusted AI group in IBM Research. His recent research interests include anomaly detection, tensors, and collaborative learning. For more detail, see his website: http://ide-research.net/