Semi-Supervised Domain Adaptation via Minimax Entropy

By Kuniaki Saito, Stan Sclaroff, Trevor Darrell, Kate Saenko. 2019

738 citations recorded by the source corpus.

Contemporary domain adaptation methods are very effective at aligning feature distributions of source and target domains without any target supervision. However, we show that these techniques perform poorly when even a few labeled examples are available in the target domain. To address this semi-supervised domain adaptation (SSDA) setting, we propose a novel Minimax Entropy (MME) approach that adversarially optimizes an adaptive few-shot model. Our base model consists of a feature encoding network, followed by a classification layer that computes the features' similarity to estimated prototypes (representatives of each class). Adaptation is achieved by alternately maximizing the conditional entropy of unlabeled target data with respect to the classifier and minimizing it with respect to the feature encoder. We empirically demonstrate the superiority of our method over many baselines, including conventional feature alignment and few-shot methods, setting a new state of the art for SSDA. Our code is available at \url{http://cs-people.bu.edu/keisaito/research/MME.html}.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

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

Citations in the corpus

A Theory of Multiple-Source Adaptation with Limited Target Labeled Data (2020, methodology): “Saito et al. (2019) proposed to use a minmax entropy method for domain adaptation.”
Domain-Adaptive Few-Shot Learning (2020, background): “Given that they are mainly designed for distribution alignment across domains (recently focusing on per-class alignment [47, 32, 5, 25, 46]), they are intrinsically unsuited for FSL whereby the target classes are completely different from the source ones: either global or per-class distribution alig”
Tent: Fully Test-Time Adaptation by Entropy Minimization (2021, background): “Entropy Minimization Entropy minimization is a key regularizer for domain adaptation (Carlucci et al., 2017; Shu et al., 2018; Saito et al., 2019; Roy et al., 2019), semi-supervised learning (Grandvalet & Bengio, 2005; Lee, 2013; Berthelot et al.”
Reducing Domain Gap via Style-Agnostic Networks (2019, methodology): “This approach can be applied to both labeled and unlabeled examples in UDA or SSDA settings.”
Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation (2021, background): “Entropy minimization and its variants [9, 46, 57] encourage the network to make sharp predictions on the unlabeled target data, and the resulting network is prone to be over-confident on false predictions.”
Unsupervised Domain Adaptation in Semantic Segmentation via Orthogonal and Clustered Embeddings (2020, methodology): “Avoiding the need for target pseudo-labels, [43, 44] propose a self-supervised clustering technique to discover target modes without any form of supervision.”
Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation (2020, positive): “IM would work better than conditional entropy minimization (Grandvalet & Bengio, 2005) widely used in prior DA works (Vu et al., 2019; Saito et al., 2019), since IM can circumvent the trivial solution where all unlabeled data have the same one-hot encoding via the fair diversity-promoting objective…”
Learning to Diversify for Single Domain Generalization (2021, background): “Domain adaptation has been widely studied in various settings, such as semi-supervised [18, 38, 46] and unsupervised scenarios [26, 28], which leverage the partially labelled or unlabelled target domain in the training process.”
Self-Supervised Pretraining Improves Self-Supervised Pretraining (2021, methodology): “We used this model to initialize the feature encoder in a Minimax Entropy (MME) Domain Adaptation (DA) method [52].”
Contrastive Test-Time Adaptation (2022, methodology): “It is worth to know that since the original DomainNet has noisy labels, we follow the authors’ followup work [45] to use a subset of it that contains 126 classes from 4 domains (Real, Sketch, Clipart, Painting), which we refer to as DomainNet-126.”
A Broader Study of Cross-Domain Few-Shot Learning (2019, background): “In the semi-supervised few-shot learning [35,45,2,37,49] the unlabeled data comes in addition to the support set and is assumed to have a similar distribution to the target classes (although some unrelated samples noise is also allowed).”
Adversarial Learning and Self-Teaching Techniques for Domain Adaptation in Semantic Segmentation (2019, background): “A curriculum-style learning approach is proposed in [16], where ﬁrstly the easier task of estimating global label distributions is learned and then the segmentation network is trained forcing that the target label distribution is aligned to the previously computed properties.”
Reducing Domain Gap by Reducing Style Bias (2021, methodology): “Saito et al. [44] showed that common UDA methods [11, 34, 45] often fail to improve their performance in such scenario, and proposed minimax entropy (MME) setting the state-of-the-art for SSDA.”
Delving Deep into the Generalization of Vision Transformers under Distribution Shifts (2021, methodology): “Thus we further design Generalization-Enhanced ViTs (GE-ViTs) from the perspectives of adversarial training [10], information theory [24] and self-supervised learning [34], named as T-ADV, T-MME, and T-SSL respectively.”
A Survey of Unsupervised Deep Domain Adaptation (2018, background): “, semi-supervised learning [203] or weakly-supervised learning [221]), domain adaptation may bring performance gains to other problems as well.”