Seminar Information


Speaker: Dr.Sainbayar Sukhbaatar (New York University)

Title: Memory network and asymmetric self-play

Time and Place: 14:30--15:30, Sept.26th 2017,  14th Eng.Bld.Room 522

Abstract:  In the first half this talk, I will talk about end-to-end memory network, which is a neural network with an external memory. This memory can store variable number of inputs, and its soft attention mechanism enables the network to selectively access important parts of the memory. Therefore, the memory network can process input items in out-of-order fashion, which is crucial many tasks such dialog system. In the second half, I will talk about my recent work on asymmetric self-play for intrinsic exploration. The idea is to let the agent play a two-phase imitation game with itself. In the task proposal phase, the agent is free to do anything in the environment. Then in the execution phase, the agent is asked to repeat the same thing. As the agent propose and execute increasingly challenging tasks, it learns to navigate the environment efficiently.


  1. Weston et al., Memory Networks, ICLR 2015
  2. Sukhbaatar et al., End-to-End Memory Networks, NIPS 2015
  3. Sukhbaatar et al., Intrinsic Motivation and Automatic Curricula via Asymmetric Self-Play, 2017


Speaker: Prof. Jiayu Zhou (Michigan State University)

Title: Recent Advances in Multi-Task Learning

Time and Place: 11:00--12:00,Aug.25th 2017, 14th Eng.Bld.Room 522

Abstract: The recent decade has witnessed a surging demand in data analysis, where we built machine learning models for various data analysis tasks. The multi-task learning is a machine learning paradigm that bridges related learning tasks and transfers knowledge among the tasks. The seminar reviews multi-task learning basics and recent advances, including distributed multi-task learning that provides efficient and privacy-preserving learning on distributed data sources; and interactive multi-task learning that solicits and integrates domain knowledge in multi-task learning, including human in the learning loop. The seminar is concluded by a discussion of future directions of multi-task learning. 


Biography: Jiayu Zhou is currently an Assistant Professor in the Department of Computer Science and Engineering at Michigan State University. He received his Ph.D. degree in computer science from Arizona State University in 2014. He has a broad research interest in large-scale machine learning and data mining, and biomedical informatics. He served as technical program committee members of premier conferences such as NIPS, ICML, and SIGKDD. Jiayu’s research is supported by National Science Foundation and Office of Naval Research. His papers received the Best Student Paper Award in 2014 IEEE International Conference on Data Mining (ICDM), the Best Student Paper Award at 2016 International Symposium on Biomedical Imaging (ISBI), and Best Paper Award at 2016 IEEE International Conference on Big Data (BigData). 





2016/3/22 16:00--17:30