SOSELETO: A Unified Approach to Transfer Learning and Training with Noisy Labels

von · Mai 6, 2019 · 102 Besichtigungen ·

ICLR 2019

Modern representation learning techniques like deep neural networks have had a major impact on a wide range of tasks, achieving new state-of-the-art performances on benchmarks using little or no feature engineering. However, these gains are often difficult to translate into real-world settings because they usually require massive hand-labeled training sets. Collecting such training sets by hand is often infeasible due to the time and expense of labeling data; moreover, hand-labeled training sets are static and must be completely relabeled when real-world modeling goals change. Increasingly popular approaches for addressing this labeled data scarcity include using weak supervision---higher-level approaches to labeling training data that are cheaper and/or more efficient, such as distant or heuristic supervision, constraints, or noisy labels; multi-task learning, to effectively pool limited supervision signal; data augmentation strategies to express class invariances; and introduction of other forms of structured prior knowledge. An overarching goal of such approaches is to use domain knowledge and data resources provided by subject matter experts, but to solicit it in higher-level, lower-fidelity, or more opportunistic ways. In this workshop, we examine these increasingly popular and critical techniques in the context of representation learning. While approaches for representation learning in the large labeled sample setting have become increasingly standardized and powerful, the same is not the case in the limited labeled data and/or weakly supervised case. Developing new representation learning techniques that address these challenges is an exciting emerging direction for research [e.g., 1, 2]. Learned representations have been shown to lead to models robust to noisy inputs, and are an effective way of exploiting unlabeled data and transferring knowledge to new tasks where labeled data is sparse.