説明
Imperfect Information Learning Team (https://aip.riken.jp/labs/generic_tech/imperfect_inf_learn/) at RIKEN AIP
Speaker 1: Masashi Sugiyama
Title: Overview of Imperfect Information Learning Team
Abstract: Machine learning has been successfully employed in various important application domains where massive high-quality labeled data is available. However, learning in more realistic situations where training data is noisy, biased, weakly supervised, and even adversarial is still challenging. The Imperfect Information Learning Team is developing fundamental algorithms for learning from such imperfect data, elucidating their mathematical properties, and demonstrating their practical usefulness in experiments.
Speaker 2: Gang Niu
Title: Robust Learning against Label Noise
Abstract: The label noise problem belongs to the inaccurate supervision—one of the three typical types of weak supervision. The label noise may exist in many real-world applications where there are limited budgets for labeling the raw data. This talk will briefly review our recent advances in robust learning against label noise. More specifically, it includes the following parts:
1. the classical problem setting of label noise;
2. the loss correction approach for learning under label noise;
3. the sample selection approach for learning under label noise.
Speaker 3: Voot Tangkaratt
Title: Sequential Decision Making from Noisy Demonstrations
Abstract: Sequential decision making aims to learn a good policy that makes good sequences of decisions. Researchers have shown that a good policy can be learned efficiently from high-quality demonstrations. However, demonstrations in the real-world often have lower quality due to noise or insufficient expertise of demonstrators, and this makes learning a good policy highly challenging. This talk presents our recent works that tackle this challenge in the context of imitation learning.
Specifically, we present three approaches for imitation learning from noisy demonstrations: a variational inference approach, a divergence minimization approach, and a risk optimization approach.
Speaker 4: Shuo Chen
Title: Robust Distance Metric Learning
Abstract: Metric Learning is a supervised learning problem, where the similarities between pairwise instances are used for supervision instead of the class labels. Metric learning has been widely applied in many recognition tasks such as face verification, person re-identification, and image retrieval. However, learning a distance metric that is robust to high-dimensional data is still challenging. This talk will introduce our recent works on robust metric learning. Specifically, we will briefly review the main existing metric learning models, and then propose a new robust metric learning algorithm for high-dimensional data.
