Universal Domain Adaptation from Foundation Models

Foundation models (e.g., CLIP or DINOv2) have shown their impressive learning and transferring capabilities on a wide range of visual tasks, by training on a large corpus of data and adapting to specific downstream tasks. It is, however, interesting that foundation models have not been fully explored for universal domain adaptation (UniDA), which is to learn models using labeled data in a source domain and unlabeled data in a target one, such that the learned models can successfully adapt to the target data. In this paper, we make comprehensive empirical studies of state-of-the-art UniDA methods using foundation models. We first demonstrate that, while foundation models greatly improve the performance of the baseline methods that train the models on the source data alone, existing UniDA methods generally fail to improve over the baseline. This suggests that new research efforts are very necessary for UniDA using foundation models. To this end, we propose a very simple method of target data distillation on the CLIP model, and achieves consistent improvement over the baseline across all the UniDA benchmarks. Our studies are under a newly proposed evaluation metric of universal classification rate (UCR), which is threshold- and ratio-free and addresses the threshold-sensitive issue encountered when using the existing H-score metric.

Adversarial Style Augmentation for Domain Generalization

It is well-known that the performance of well-trained deep neural networks may degrade significantly when they are applied to data with even slightly shifted distributions. Recent studies have shown that introducing certain perturbation on feature statistics (\eg, mean and standard deviation) during training can enhance the cross-domain generalization ability. Existing methods typically conduct such perturbation by utilizing the feature statistics within a mini-batch, limiting their representation capability. Inspired by the domain generalization objective, we introduce a novel Adversarial Style Augmentation (ASA) method, which explores broader style spaces by generating more effective statistics perturbation via adversarial training. Specifically, we first search for the most sensitive direction and intensity for statistics perturbation by maximizing the task loss. By updating the model against the adversarial statistics perturbation during training, we allow the model to explore the worst-case domain and hence improve its generalization performance. To facilitate the application of ASA, we design a simple yet effective module, namely AdvStyle, which instantiates the ASA method in a plug-and-play manner. We justify the efficacy of AdvStyle on tasks of cross-domain classification and instance retrieval. It achieves higher mean accuracy and lower performance fluctuation. Especially, our method significantly outperforms its competitors on the PACS dataset under the single source generalization setting, \eg, boosting the classification accuracy from 61.2\% to 67.1\% with a ResNet50 backbone. Our code will be available at \url{https://github.com/YBZh/AdvStyle}.

Counterfactual Supervision-based Information Bottleneck for Out-of-Distribution Generalization

Learning invariant (causal) features for out-of-distribution (OOD) generalization has attracted extensive attention recently, and among the proposals invariant risk minimization (IRM) (Arjovsky et al., 2019) is a notable solution. In spite of its theoretical promise for linear regression, the challenges of using IRM in linear classification problems yet remain (Rosenfeld et al.,2020, Nagarajan et al., 2021). Along this line, a recent study (Arjovsky et al., 2019) has made a first step and proposes a learning principle of information bottleneck based invariant risk minimization (IB-IRM). In this paper, we first show that the key assumption of support overlap of invariant features used in (Arjovsky et al., 2019) is rather strong for the guarantee of OOD generalization and it is still possible to achieve the optimal solution without such assumption. To further answer the question of whether IB-IRM is sufficient for learning invariant features in linear classification problems, we show that IB-IRM would still fail in two cases whether or not the invariant features capture all information about the label. To address such failures, we propose a \textit{Counterfactual Supervision-based Information Bottleneck (CSIB)} learning algorithm that provably recovers the invariant features. The proposed algorithm works even when accessing data from a single environment, and has theoretically consistent results for both binary and multi-class problems. We present empirical experiments on three synthetic datasets that verify the efficacy of our proposed method.

Gradual Domain Adaptation via Self-Training of Auxiliary Models

Domain adaptation becomes more challenging with increasing gaps between source and target domains. Motivated from an empirical analysis on the reliability of labeled source data for the use of distancing target domains, we propose self-training of auxiliary models (AuxSelfTrain) that learns models for intermediate domains and gradually combats the distancing shifts across domains. We introduce evolving intermediate domains as combinations of decreasing proportion of source data and increasing proportion of target data, which are sampled to minimize the domain distance between consecutive domains. Then the source model could be gradually adapted for the use in the target domain by self-training of auxiliary models on evolving intermediate domains. We also introduce an enhanced indicator for sample selection via implicit ensemble and extend the proposed method to semi-supervised domain adaptation. Experiments on benchmark datasets of unsupervised and semi-supervised domain adaptation verify its efficacy.

Semi-supervised Models are Strong Unsupervised Domain Adaptation Learners

Unsupervised domain adaptation (UDA) and semi-supervised learning (SSL) are two typical strategies to reduce expensive manual annotations in machine learning. In order to learn effective models for a target task, UDA utilizes the available labeled source data, which may have different distributions from unlabeled samples in the target domain, while SSL employs few manually annotated target samples. Although UDA and SSL are seemingly very different strategies, we find that they are closely related in terms of task objectives and solutions, and SSL is a special case of UDA problems. Based on this finding, we further investigate whether SSL methods work on UDA tasks. By adapting eight representative SSL algorithms on UDA benchmarks, we show that SSL methods are strong UDA learners. Especially, state-of-the-art SSL methods significantly outperform existing UDA methods on the challenging UDA benchmark of DomainNet, and state-of-the-art UDA methods could be further enhanced with SSL techniques. We thus promote that SSL methods should be employed as baselines in future UDA studies and expect that the revealed relationship between UDA and SSL could shed light on future UDA development. Codes are available at \url{https://github.com/YBZh}.

On Universal Black-Box Domain Adaptation

In this paper, we study an arguably least restrictive setting of domain adaptation in a sense of practical deployment, where only the interface of source model is available to the target domain, and where the label-space relations between the two domains are allowed to be different and unknown. We term such a setting as Universal Black-Box Domain Adaptation (UB$^2$DA). The great promise that UB$^2$DA makes, however, brings significant learning challenges, since domain adaptation can only rely on the predictions of unlabeled target data in a partially overlapped label space, by accessing the interface of source model. To tackle the challenges, we first note that the learning task can be converted as two subtasks of in-class\footnote{In this paper we use in-class (out-class) to describe the classes observed (not observed) in the source black-box model.} discrimination and out-class detection, which can be respectively learned by model distillation and entropy separation. We propose to unify them into a self-training framework, regularized by consistency of predictions in local neighborhoods of target samples. Our framework is simple, robust, and easy to be optimized. Experiments on domain adaptation benchmarks show its efficacy. Notably, by accessing the interface of source model only, our framework outperforms existing methods of universal domain adaptation that make use of source data and/or source models, with a newly proposed (and arguably more reasonable) metric of H-score, and performs on par with them with the metric of averaged class accuracy.

Label Propagation with Augmented Anchors: A Simple Semi-Supervised Learning baseline for Unsupervised Domain Adaptation

Motivated by the problem relatedness between unsupervised domain adaptation (UDA) and semi-supervised learning (SSL), many state-of-the-art UDA methods adopt SSL principles (e.g., the cluster assumption) as their learning ingredients. However, they tend to overlook the very domain-shift nature of UDA. In this work, we take a step further to study the proper extensions of SSL techniques for UDA. Taking the algorithm of label propagation (LP) as an example, we analyze the challenges of adopting LP to UDA and theoretically analyze the conditions of affinity graph/matrix construction in order to achieve better propagation of true labels to unlabeled instances. Our analysis suggests a new algorithm of Label Propagation with Augmented Anchors (A$^2$LP), which could potentially improve LP via generation of unlabeled virtual instances (i.e., the augmented anchors) with high-confidence label predictions. To make the proposed A$^2$LP useful for UDA, we propose empirical schemes to generate such virtual instances. The proposed schemes also tackle the domain-shift challenge of UDA by alternating between pseudo labeling via A$^2$LP and domain-invariant feature learning. Experiments show that such a simple SSL extension improves over representative UDA methods of domain-invariant feature learning, and could empower two state-of-the-art methods on benchmark UDA datasets. Our results show the value of further investigation on SSL techniques for UDA problems.

Unsupervised Multi-Class Domain Adaptation: Theory, Algorithms, and Practice

In this paper, we study the formalism of unsupervised multi-class domain adaptation (multi-class UDA), which underlies some recent algorithms whose learning objectives are only motivated empirically. A Multi-Class Scoring Disagreement (MCSD) divergence is presented by aggregating the absolute margin violations in multi-class classification; the proposed MCSD is able to fully characterize the relations between any pair of multi-class scoring hypotheses. By using MCSD as a measure of domain distance, we develop a new domain adaptation bound for multi-class UDA as well as its data-dependent, probably approximately correct bound, which naturally suggest adversarial learning objectives to align conditional feature distributions across the source and target domains. Consequently, an algorithmic framework of Multi-class Domain-adversarial learning Networks (McDalNets) is developed, whose different instantiations via surrogate learning objectives either coincide with or resemble a few of recently popular methods, thus (partially) underscoring their practical effectiveness. Based on our same theory of multi-class UDA, we also introduce a new algorithm of Domain-Symmetric Networks (SymmNets), which is featured by a novel adversarial strategy of domain confusion and discrimination. SymmNets afford simple extensions that work equally well under the problem settings of either closed set, partial, or open set UDA. We conduct careful empirical studies to compare different algorithms of McDalNets and our newly introduced SymmNets. Experiments verify our theoretical analysis and show the efficacy of our proposed SymmNets. We make our implementation codes publicly available.