[論文レビュー] Confidence Regularized Self-Training

Yang Zou, Zhiding Yu+3 more|arXiv (Cornell University)|Aug 26, 2019

Domain Adaptation and Few-Shot Learning参考文献 63被引用数 33

ひとこと要約

CRSTはソフト疑似ラベルと出力平滑化正則化を用いることで、教師なしドメイン適応を改善し、複数のベンチマークで最先端の結果を達成します。

ABSTRACT

Recent advances in domain adaptation show that deep self-training presents a powerful means for unsupervised domain adaptation. These methods often involve an iterative process of predicting on target domain and then taking the confident predictions as pseudo-labels for retraining. However, since pseudo-labels can be noisy, self-training can put overconfident label belief on wrong classes, leading to deviated solutions with propagated errors. To address the problem, we propose a confidence regularized self-training (CRST) framework, formulated as regularized self-training. Our method treats pseudo-labels as continuous latent variables jointly optimized via alternating optimization. We propose two types of confidence regularization: label regularization (LR) and model regularization (MR). CRST-LR generates soft pseudo-labels while CRST-MR encourages the smoothness on network output. Extensive experiments on image classification and semantic segmentation show that CRSTs outperform their non-regularized counterpart with state-of-the-art performance. The code and models of this work are available at https://github.com/yzou2/CRST.

研究の動機と目的

• Motivate unsupervised domain adaptation and address noisy pseudo-labels in self-training.
• Introduce a continuous pseudo-label framework to relax one-hot assumptions.
• Propose two regularization paradigms (label regularization and model regularization) to stabilize training.
• Theoretically relate CRST to regularized maximum likelihood and EM.
• Empirically validate CRST on image classification and semantic segmentation benchmarks.

提案手法

• Generalize CBST to continuous pseudo-labels on a target domain.
• Introduce label regularization (LR) with soft pseudo-labels and a negative entropy constraint (LRENT).
• Introduce model regularization (MR) with output-smoothing terms (MRL2, MRENT, MRKLD).
• Optimize via alternating steps: pseudo-label generation and network retraining.
• Provide closed-form solutions or gradients for regularizers and connect LRENT to softmax with temperature.
• Demonstrate CRST variants outperform CBST on multiple datasets.

実験結果

リサーチクエスチョン

• RQ1 Can continuous pseudo-labels improve stability and accuracy in self-training for UDA?
• RQ2 Do label and model regularizers reduce error propagation from incorrect pseudo-labels?
• RQ3 How do CRST regularizers relate to known probabilistic/softmax formulations?
• RQ4 Which regularizer(s) provide the best performance across vision tasks?
• RQ5 What is the empirical impact of CRST on classification and segmentation benchmarks?

主な発見

• CRST variants outperform their non-regularized CBST baseline across benchmarks.
• MRKLD with LRENT often yields the best or competitive results among single regularizers.
• Combining MRKLD and LRENT yields further improvements on VisDA17 and Office-31.
• CRST achieves state-of-the-art performance in GTA5→Cityscapes and SYNTHIA→Cityscapes under DeepLabv2 and other backbones.
• Soft pseudo-labels and output-smoothing regularizers reduce overconfident errors and improve domain adaptation performance.