Executive Summary : | Weak supervision and noisy supervision are significant challenges in building AI models. Human data annotation often leads to labeling errors due to subjectivity, which can be costly. On the other hand, weak labels can be acquired with minimal cost, such as learning from bandit feedback, partially labeled data, positive and unlabeled data, label proportions, and pairwise similarities. Effective learning algorithms have been proposed to learn AI models in weak supervision settings, but there could be labeling errors (noises) even in these settings. For example, in partial label learning, actual labels may not be present in the candidate label set. In the case of bandit feedback, noisy feedbacks can be received, such as when showing ads on an e-commerce site. Efficient and robust models under noisy weak supervision are yet to be developed. This project aims to build robust deep models in the presence of noisy weak supervision, specifically considering three weak supervision settings: learning with bandit feedback, learning with partial labels, and learning with pairwise similarities. Existing robust models are primarily designed for full information cases, but there is a need to extend them to work in noisy weak supervision cases. The project plans to develop end-to-end algorithms that can learn robust classifiers in noisy supervision settings and build a generative model (conditional GAN) in noisy pairwise similarity label settings. |