Cycle Self-Training for Domain Adaptation (2103.03571v3)

Abstract: Mainstream approaches for unsupervised domain adaptation (UDA) learn domain-invariant representations to narrow the domain shift. Recently, self-training has been gaining momentum in UDA, which exploits unlabeled target data by training with target pseudo-labels. However, as corroborated in this work, under distributional shift in UDA, the pseudo-labels can be unreliable in terms of their large discrepancy from target ground truth. Thereby, we propose Cycle Self-Training (CST), a principled self-training algorithm that explicitly enforces pseudo-labels to generalize across domains. CST cycles between a forward step and a reverse step until convergence. In the forward step, CST generates target pseudo-labels with a source-trained classifier. In the reverse step, CST trains a target classifier using target pseudo-labels, and then updates the shared representations to make the target classifier perform well on the source data. We introduce the Tsallis entropy as a confidence-friendly regularization to improve the quality of target pseudo-labels. We analyze CST theoretically under realistic assumptions, and provide hard cases where CST recovers target ground truth, while both invariant feature learning and vanilla self-training fail. Empirical results indicate that CST significantly improves over the state-of-the-arts on visual recognition and sentiment analysis benchmarks.

• The paper introduces Cycle Self-Training (CST), a novel algorithm improving unsupervised domain adaptation by enhancing the reliability of pseudo-labels under distributional shifts.
• CST employs a cyclical process with forward and reverse steps, using Tsallis entropy for an adaptive, confidence-friendly uncertainty measure to refine pseudo-labels.
• Empirical results show CST outperforms state-of-the-art methods on visual recognition and sentiment analysis benchmarks, validated by theoretical analysis under realistic conditions.

Cycle Self-Training for Domain Adaptation

The paper explores a novel approach for unsupervised domain adaptation (UDA) by introducing Cycle Self-Training (CST), an algorithm designed to address the inherent challenges faced by traditional domain-invariant representation learning methods under distributional shifts. Researchers have long aimed to transfer knowledge from a richly-supervised source domain to an unlabeled target domain, a task complicated by subtle changes in data distributions that can cause models to falter. While feature adaptation methods have founded success through domain alignment, they are theoretically constrained by impossibility theorems related to label and support shifts. Recently, self-training has emerged as a promising alternative, leveraging pseudo-labels from unlabeled data for model improvement. However, these pseudo-labels often prove unreliable due to distributional shifts.

CST proposes an innovative solution to improve the generalization of pseudo-labels across domains. It employs a cyclical process comprising a forward step, in which a source-trained classifier generates pseudo-labels for the target, and a reverse step, wherein a target classifier is trained using these pseudo-labels and subsequently updated to perform accurately on source data. To enhance the reliability of pseudo-labels, CST introduces a confidence-friendly uncertainty measure based on Tsallis entropy, which can adaptively minimize prediction uncertainty without manual threshold tuning.

The paper's theoretical analysis substantiates the effectiveness of CST under realistic conditions. The expansion assumption—a property that reflects good continuity within each class—is leveraged to demonstrate that the CST minimizer will have small target errors if certain regularity conditions are met. Moreover, empirical validation shows CST outperforms state-of-the-art methods in a significant majority of tasks on visual recognition and sentiment analysis benchmarks. This indicates that CST has effectively mitigated the pitfalls of standard self-training under domain shifts, leading to significant improvements in target domain adaptability.

This research carries substantial implications for both practical applications and theoretical advancement in UDA. Practically, CST offers a robust solution for adapting models to new, unlabeled domains with minimized risk from unreliable pseudo-labels, which is critical in real-world deployment scenarios. Theoretically, it paves the way for exploration into entropy-based regularization techniques in domain adaptation, diverging from traditional domain-invariance strategies. Future work should explore extending CST to other semi-supervised learning paradigms like consistency regularization and self-ensembling, which are potentially beneficial under distribution shifts.

In summary, the Cycle Self-Training framework presents an insightful shift in addressing UDA challenges by refining pseudo-label reliability and capitalizing on entropy-based measures for better cross-domain generalization. The strong performance metrics and thorough theoretical justification validate CST as a significant step forward in adaptive learning, with promising directions for future research and development in AI.