革新知能統合研究センター

The TrustML Young Scientist Seminars (TrustML YSS) started from January 28, 2022.

The TrustML YSS is a video series that features young scientists giving talks and discoveries in relation with Trustworthy Machine Learning.

This network is funded by RIKEN-AIP’s subsidy and JST, ACT-X Grant Number JPMJAX21AF, Japan.

Timetable for the TrustML YSS online seminars from Jan. to Feb. 2022.

For more information please see the following site.
TrustML YSS

This network is funded by RIKEN-AIP’s subsidy and JST, ACT-X Grant Number JPMJAX21AF, Japan.

【The 2nd Seminar】

Date and Time: February 4th 10:00am – 12:00pm(JST)

Venue: Zoom webinar

Language: English

10:00am – 11:00am

Speaker: Vikash Sehwag (Princeton University)
Title: A generative approach to robust machine learning
Abstract: My talk will focus on the emerging direction of making machine learning reliable by incorporating data-distribution information through generative models. It is natural to ask, why would generative models help, how much generative models can even help, or which generative models are most helpful? I will present a framework catered to answering these questions. In particular, I’ll demonstrate why we should use diffusion-based generative models, instead of generative adversarial networks (GANs) in robust learning applications. Though diffusion-based generative models are an unbounded source of data, their sampling process is hard to guide toward specific regions. I will discuss the techniques we developed to enable guided sampling from regions critical to robust learning.

Short Bio:
Vikash is a PhD candidate in Electrical Engineering at Princeton University. He is interested in research problems at the intersection of security, privacy, and machine learning. Some topics he has worked on are adversarial robust supervised / self-supervised learning, adversarial robustness in compressed neural networks, self-supervised detection of outliers, robust open-world machine learning, and privacy leakage in large scale deep learning.
He is co-advised by Prateek Mittal and Mung Chiang. Before joining Princeton, he completed my undergraduate in E&ECE (with minor in CS) from IIT Kharagpur, India. He earlier had an amazing summer internship experience at Microsoft Research, Redmond. Before that, He spent a wonderful summer working with Heinz Koeppl at TU Darmstadt. He has also received Qualcomm Innovation Fellowship in 2019.

11:00am- 12:00pm

Speaker: Tonghan Wang (Tsinghua University & Harvard University)
Title: Role-Based Cooperative Reinforcement Learning
Abstract: Multi-agent reinforcement learning holds the promise to imitating the large-scale cooperation behavior of humans. A major challenge faced by this field of machine learning is scalability — the problem search space grows exponentially with the number of agents.The role concept provides a useful tool to design and understand such complex multi-agent systems, which allows agents with a similar role to share similar behaviors. However, existing role-based methods use prior domain knowledge and predefine role structures and behaviors. In contrast, multi-agent reinforcement learning provides flexibility and adaptability, but less efficiency in complex tasks. To merge the best of these two worlds, we propose to synergize these two paradigms and propose the role-oriented MARL learning framework. In this framework, roles are emergent, and agents with similar roles tend to share their learning and to be specialized on certain sub-tasks. In this talk, we will introduce two specific role-based learning algorithms and study how they achieve state-of-the-art performance on complex benchmarks like StarCraft II micromanagement.

Short Bio:
Tonghan Wang is currently working with Prof. Chongjie Zhang at Institute for Interdisciplinary Information Sciences, Tsinghua University, headed by Prof. Andrew Yao. He will join the EconCS group at Harvard University and work with Prof. David Parkes. His primary research goal is to develop innovative models and methods to enable effective multi-agent cooperation, allowing a group of individuals to explore, communicate, and accomplish tasks of higher complexity. His research interests include multi-agent learning, reinforcement learning, and reasoning under uncertainty.

All participants are required to agree with the AIP Seminar Series Code of Conduct.
Please see the URL below.
https://aip.riken.jp/event-list/termsofparticipation/?lang=en

RIKEN AIP will expect adherence to this code throughout the event. We expect cooperation from all participants to help ensure a safe environment for everybody.