This is an online seminar. Registration is required.
【Imperfect Information Learning Team】
【Date】2024/September/6(Fri) 16:00-17:00(JST)
*【Speaker】Yujie Zhang, University of Tokyo *
Title: Reliable Online Learning: Towards Adapting to Non-stationary Environments with Limited Feedback
Abstract:
In many applications, data are typically accumulated in an online fashion, posing challenges in processing non-stationary data whose underlying distributions shift over time. Additionally, the time-consuming labeling process makes it difficult to acquire high-quality annotations for fast-processing online data. This talk will discuss reliable machine learning methods designed to adapt to non-stationary environments with limited feedback. First, I will discuss the sequential distribution shift problem, where the data distribution in the test stream changes over time. Focusing specifically on label shift and covariate shift scenarios, we developed algorithms that can provably adapt to changing environments even with an unlabeled data stream. Next, I will discuss online learning with bandit feedback, where only partial information about the loss function is available during the learning process. Specifically, we consider the case where the feedback is randomly generated by a multinomial logit model. By exploiting the local curvature of the logit model, we propose an online learning algorithm that can effectively learn from one-point feedback with both statistical and computational efficiency.
Public events of RIKEN Center for Advanced Intelligence Project (AIP)
Join community