Wednesday March 08, 12pm1pm, 1116E Klaus
Learning from Conditional Distributions via Dual EmbeddingsBo Dai
Advisor: Prof. Le Song


ABSTRACT
Many machine learning tasks, such as learning with invariance and policy evaluation in reinforcement learning, can be characterized as problems of learning from conditional distributions. In such problems, each sample x itself is associated with a conditional distribution p(zx) represented by samples {zi}Mi=1, and the goal is to learn a function f that links these conditional distributions to target values y. These learning problems become very challenging when we only have limited samples or in the extreme case only one sample from each conditional distribution. Commonly used approaches either assume that z is independent of x, or require an overwhelmingly large samples from each conditional distribution. To address these challenges, we propose a novel approach which employs a new minmax reformulation of the learning from conditional distribution problem. With such new reformulation, we only need to deal with the joint distribution p(z,x). We also design an efficient learning algorithm, EmbeddingSGD, and establish theoretical sample complexity for such problems. Finally, our numerical experiments on both synthetic and realworld datasets show that the proposed approach can significantly improve over the existing algorithms.
BIO
Bo Dai is a Ph.D. student in Computational Science and Engineering at Georgia Tech, supervised by Prof. Le Song. He is working on developing efficient optimization algorithms for statistical methods, including kernel methods and nonparametric graphical models, to learn from massive volume of complex, uncertain and highdimensional data. In particularly, He is trying to make nonparametric models scalable with provable statistical and computational guarantees.