Moment Matching for Multi-Source Domain Adaptation

By Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko. 2018

2,338 citations recorded by the source corpus.

Conventional unsupervised domain adaptation (UDA) assumes that training data are sampled from a single domain. This neglects the more practical scenario where training data are collected from multiple sources, requiring multi-source domain adaptation. We make three major contributions towards addressing this problem. First, we collect and annotate by far the largest UDA dataset, called DomainNet, which contains six domains and about 0.6 million images distributed among 345 categories, addressing the gap in data availability for multi-source UDA research. Second, we propose a new deep learning approach, Moment Matching for Multi-Source Domain Adaptation (M3SDA), which aims to transfer knowledge learned from multiple labeled source domains to an unlabeled target domain by dynamically aligning moments of their feature distributions. Third, we provide new theoretical insights specifically for moment matching approaches in both single and multiple source domain adaptation. Extensive experiments are conducted to demonstrate the power of our new dataset in benchmarking state-of-the-art multi-source domain adaptation methods, as well as the advantage of our proposed model. Dataset and Code are available at http://ai.bu.edu/M3SDA/

Subfields: Artificial Intelligence, Cancer Research, 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 Cancer Research

In Computer Vision and Pattern Recognition

In Representation & Self-Supervised Learning

Citations in the corpus

Break-It-Fix-It: Unsupervised Learning for Program Repair (2021, background): “Such domain shifts typically occur across related but different datasets (Torralba & Efros, 2011; Fang et al., 2013; Venkateswara et al., 2017; Yu et al., 2018b; 2019; Peng et al., 2019; Kamath et al., 2020), as well as from synthetic data to real data (including sim2real) (Wang et al.”
Refining Language Models with Compositional Explanations (2021, background): “In another relevant thread, unsupervised domain adaptation (UDA) looks to adapt a trained model to a new domain using unlabeled data from the target domain, by updating feature representation to minimize the distribution divergence between domains [12, 11], or updating models to match the distributi”
ELEVATER: A Benchmark and Toolkit for Evaluating Language-Augmented Visual Models (2022, background): “Openset recognition allows new concepts in evaluation, while typically remains the same visual domain [87, 93]; Domain shift allows new visual domain in evaluation, while typically remains the same concept pool [63, 28].”
Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-KITCHENS-100 (2020, background): “ides a number of practical benets, most notably the elimination of labelling time and expense when collecting new videos, in the future. Related Datasets. UDA datasets have traditionally used images [86,87,88,89], with recent attempts to use video [90,91,92] adapting across public datasets (e.g. UCF”
Multiple-Source Adaptation with Domain Classifiers (2020, background): “Peng et al. [2019] proposed to align moments of feature distribution across source and target domains.”
First Session Adaptation: A Strong Replay-Free Baseline for Class-Incremental Learning (2023, methodology): “The datasets are CIFAR100, SVHN, dSprites-location, FGVC-Aircraft, Cars, and Letters while also including 4 extra datasets: CUB200 [32], CORE50 [15], iNaturalist [30], and DomainNet [22].”
CODA-Prompt: COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning (2022, neutral): “We report extended results, including standard deviations and additional parameters trained, for all benchmarks in Tables A (5-task ImageNet-R [3, 9]), B (10-task ImageNet-R), C (20-task Imagenet-R), D (10-task CIFAR100 [5]), E (5-task DomainNet [8]), and A (Dual-Shift ImageNet-R).”
S-Prompts Learning with Pre-trained Transformers: An Occam's Razor for Domain Incremental Learning (2022, methodology): “In our experiments, the total class number C is 2 for CDDB [36], 50 for CORe50 [45], and 345 for DomainNet [52] respectively.”
General Incremental Learning with Domain-aware Categorical Representations (2022, methodology): “For the ND split, every session has all the classes and each class has one incoming domain, where iCIFAR-20, iDigits and iDomainNet are split into 5, 4 and 6 sessions, respectively.”
On Generalizing Beyond Domains in Cross-Domain Continual Learning (2022, methodology): “As our experimental evaluations indicate, our methods comfortably outperform the existing methods in both class continual learning and domain generalization on challenging datasets, namely DomainNet, OfficeHome, PACS and NICO-Animal.”
Self-Supervised Models are Continual Learners (2021, methodology): “We perform experiments on 3 datasets: CI-FAR100 [35] (class-incremental), a 100-class dataset with 60k 32x32 colour images; ImageNet100 [54] (class-and data-incremental), 100-class subset of the ILSVRC2012 dataset with ≈ 130k images in high resolution (resized to 224x224); DomainNet [43] (domain-inc”
How Well Does Self-Supervised Pre-Training Perform with Streaming Data? (2021, methodology): “…with severe data distribution shifts, by splitting 1K classes of data into four chunks while maximizing the semantical dissimilarity among chunks, and the domain incremental sequence with severe domain distribution shifts, by taking each domain in DomainNet (Peng et al., 2019) as a data chunk.”
SWAD: Domain Generalization by Seeking Flat Minima (2021, positive): “Following Gulrajani and Lopez-Paz [22], we exhaustively evaluate our method and comparison methods on various benchmarks: PACS [7] (9,991 images, 7 classes, and 4 domains), VLCS [43] (10,729 images, 5 classes, and 4 domains), OfficeHome [44] (15,588 images, 65 classes, and 4 domains), TerraIncognita”
What is being transferred in transfer learning? (2020, methodology): “We use ImageNet [7] pre-training for its prevalence in the community and consider CheXpert [21] and three sets from DomainNet [33] as downstream transfer learning tasks.”
In Search of Lost Domain Generalization (2020, background): “…loaders for seven multi-domain image classification tasks: Colored MNIST [Arjovsky et al., 2019], Rotated MNIST [Ghifary et al., 2015], PACS [Li et al., 2017], VLCS [Fang et al., 2013], Office-Home [Venkateswara et al., 2017], Terra Incognita [Beery et al., 2018], and DomainNet [Peng et al., 2019].”
Meta-Learning in Neural Networks: A Survey (2020, methodology): “PACS [257] provides a good starter benchmark with Visual Decathlon [44], [206], DomainNet [258] and Meta-Dataset [207] providing larger scale alternatives.”
WILDS: A Benchmark of in-the-Wild Distribution Shifts (2020, neutral): “…on SVHN street signs (LeCun et al., 1998; Yuval et al., 2011; Tzeng et al., 2017; Hoffman et al., 2018), as well as transfers across datasets containing different renditions (e.g., photos, clipart, sketches) in DomainNet (Peng et al., 2019) and the Office-Home dataset (Venkateswara et al., 2017).”