Speaker: Yatao Bian, Machine Learning Center in Tencent AI Lab
Time: 10:00-11:00 , Jun .09
Location: Tencent Meeting
Meeting ID：722-112-167
Host: Ziping Zhao
Abstract:
Valuation problems, such as feature interpretation, data valuation and model valuation for ensembles, become increasingly important in many machine learning applications. Such problems are commonly addressed via well-known game-theoretic criteria, such as the Shapley value or Banzhaf value.
In this talk, the spaker will present a novel energy-based treatment for cooperative games, with a theoretical justification by the maximum entropy principle. Surprisingly, by conducting mean-field variational inference of the energy-based model, it is able to recover classical game-theoretic valuation criteria through conducting one-step fixed point iteration for maximizing the ELBO objective. This observation also verifies the rationality of existing criteria, as they are all attempting to decouple the correlations among players. By running the fixed point iteration for multiple steps, the proposed method achieves a trajectory of the variational valuations, among which we can define the valuation with the best conceivable decoupling error as the Variational Index. It can be proved that under uniform initializations, these variational valuations all satisfy a set of game-theoretic axioms. Finally, he will empirically demonstrate that the proposed Variational Index enjoys lower decoupling error and better valuation performance on certain synthetic and real-world valuation problems.
In this talk, the spaker will present a novel energy-based treatment for cooperative games, with a theoretical justification by the maximum entropy principle. Surprisingly, by conducting mean-field variational inference of the energy-based model, it is able to recover classical game-theoretic valuation criteria through conducting one-step fixed point iteration for maximizing the ELBO objective. This observation also verifies the rationality of existing criteria, as they are all attempting to decouple the correlations among players. By running the fixed point iteration for multiple steps, the proposed method achieves a trajectory of the variational valuations, among which we can define the valuation with the best conceivable decoupling error as the Variational Index. It can be proved that under uniform initializations, these variational valuations all satisfy a set of game-theoretic axioms. Finally, he will empirically demonstrate that the proposed Variational Index enjoys lower decoupling error and better valuation performance on certain synthetic and real-world valuation problems.
Bio:
Yatao Bian is a senior researcher of Machine Learning Center in Tencent AI Lab. He received the Ph.D. degree from the Institute for Machine Learning at ETH Zurich. He has been an associated Fellow of the Max Planck ETH Center for Learning Systems since June 2015. Before the Ph.D. program he obtained both of his M.Sc.Eng. and B.Sc.Eng. degrees from Shanghai Jiao Tong University. He is now working on graph representation learning, interpretable ML, energy-based learning, drug AI, OOD generalization and social network analysis. He has won the National Champion in AMD China Accelerated Computing Contest 2011-2012. He has published several papers on machine learning top conferences/Journals such as NeurIPS, ICML, ICLR, T-PAMI, AISTATS etc. He has served as a reviewer/PC for conferences like ICML, NeurIPS, AISTATS, CVPR, AAAI, STOC and journals such as JMLR and T-PAMI.