Confidence Regularized Self-Training

By Yang Zou, Zhiding Yu, Xiaofeng Liu, B. V. Vijaya Kumar, Jinsong Wang. 2019

917 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Representation & Self-Supervised Learning

Citations in the corpus

Big Self-Supervised Models are Strong Semi-Supervised Learners (2020, methodology): “One family of highly relevant methods are based on pseudo-labeling [11, 15] or self-training [12, 24, 59].”
Unlabeled Data Improves Adversarial Robustness (2019, neutral): “Another line of work [59, 60] is based on iterative self-training coupled with refinements such as class balance or confidence regularization.”
Hierarchical Multi-Scale Attention for Semantic Segmentation (2020, background): “[2], other semi-supervised self-training in semantic segmentation [30, 31, 32, 33, 34], and other approaches based on pseudo label such as [4, 35, 36, 3].”
Towards Discriminability and Diversity: Batch Nuclear-Norm Maximization Under Label Insufficient Situations (2020, methodology): “8 CRST [57] 88.”
PixMatch: Unsupervised Domain Adaptation via Pixelwise Consistency Training (2021, methodology): “MinEnt [52], MaxSquare [7], CRST [62], PyCDA [34]) our method performs substantially better.”
Unsupervised Intra-Domain Adaptation for Semantic Segmentation Through Self-Supervision (2020, methodology): “Other approaches [38, 37] involve generating pseudo labels for target data and conducting refinement via an iterative self-training process.”
ST++: Make Self-trainingWork Better for Semi-supervised Semantic Segmentation (2021, background): “Recently, it is increasingly attracting attention from multiple fields, such as fully-supervised image recognition [43, 53, 54, 63], semi-supervised learning [7, 17, 48, 55], and domain adaptation [30, 64, 65].”
Generalizing Hand Segmentation in Egocentric Videos With Uncertainty-Guided Model Adaptation (2020, background): “Another line of work for unsupervised domain adaptation is based on the idea of self-training where predictions from a previously trained model are exploited as pseudo-labels for training a model of focus [49, 48].”
Instance Adaptive Self-Training for Unsupervised Domain Adaptation (2020, methodology): “8) [20,35] > AT (43.”
Minimum Class Confusion for Versatile Domain Adaptation (2019, methodology): “Some methods [39,16,55,56] also utilize the less-reliable self-training or pseudo labeling, e.”
ST3D: Self-training for Unsupervised Domain Adaptation on 3D Object Detection (2021, positive): “Recently, self-training has emerged as a simple and effective technique for UDA, attaining state-of-the-art performance on many image recognition tasks [51, 54, 21].”
Deep Unsupervised Domain Adaptation: A Review of Recent Advances and Perspectives (2022, methodology): “Unlike approaches that reduce domain discrepancy with a divergence measure, self-training is proposed as an alternative training scheme, by utilizing unlabeled target domain data to achieve domain adaptation (Zou et al., 2019).”
FSDR: Frequency Space Domain Randomization for Domain Generalization (2021, background): “The third is self-training based which employs ”pseudo label” to guide unsupervised learning over unlabelled target-domain data [88, 57, 83, 19, 86, 17, 24].”
DACS: Domain Adaptation via Cross-domain Mixed Sampling (2020, methodology): “Additional difﬁculties of erroneous pseudo-labels owing to the domain shift have prompted later research to add modules for uncertainty estimation [18, 22].”
A Review of Single-Source Deep Unsupervised Visual Domain Adaptation (2020, background): “In order to solve this issue, Zou et al. [96] proposed a conﬁdence regularized self-training framework, in which pseudo-labels are treated as continuous latent variables jointly optimized via alternating optimization.”
Reliable Weighted Optimal Transport for Unsupervised Domain Adaptation (2020, methodology): “Besides, we also take into account the discriminative feature learning to beneﬁt the adaptation performance, which has been widely used to learn the deep invariant features [49, 52].”