domain adaptation

Domain adaptation is a sub-field of transfer learning that addresses the domain shift problem: a model trained on a labelled source distribution degrades when applied to a target distribution whose marginal or conditional statistics differ. Methods fall into feature-alignment approaches (learning domain-invariant representations via adversarial training, maximum mean discrepancy minimisation, or optimal transport), instance re-weighting schemes that correct for covariate shift, and self-training or pseudo-labelling strategies that exploit unlabelled target data. It spans unsupervised, semi-supervised, and multi-source settings and underpins practical deployment of models in NLP, computer vision, speech recognition, and scientific computing wherever labelled target data is scarce or costly.

Semantic Classification

Content

Compositional Relationships (Components)

SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:hasPart ml:UnsupervisedDomainAdaptation))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:hasPart ml:SemiSupervisedDomainAdaptation))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:hasPart ml:MultiSourceDomainAdaptation))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:hasPart ml:SourceFreeDomainAdaptation))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:hasPart ml:FeatureAlignment))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:hasPart ml:InstanceReWeighting))

Dependency Relationships

SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:requires ml:TrainingData))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:requires ml:SourceDomain))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:requires ml:TargetDomain))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:dependsOn ml:PreTraining))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:dependsOn ml:RepresentationLearning))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:dependsOn ml:DeepLearning))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:dependsOn ml:NeuralNetwork))

Capability Relationships

SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:enables ml:FineTuning))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:enables ml:ZeroShotLearning))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:enables ml:LowResourceNLP))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:enables ml:SyntheticToRealTransfer))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:enables ml:ModelDeployment))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:enables ml:CrossDomainGeneralisation))

Implementation Relationships

SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:implements ml:AdversarialTraining))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:implements ml:MaximumMeanDiscrepancy))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:implements ml:OptimalTransport))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:implements ml:SelfTraining))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:implements ml:PseudoLabelling))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:implements ml:ImportanceWeighting))

Reduction Relationships

SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:reducesTo ml:TransferLearning))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:reducesTo ml:DistributionalAlignment))

Support Relationships

SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:supports ml:NaturalLanguageProcessing))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:supports ml:ComputerVision))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:supports ml:MedicalImaging))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:supports ml:AutonomousDriving))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:supports ml:FederatedLearning))

Usage Relationships

SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:uses ml:DomainAdversarialNeuralNetwork))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:uses ml:GenerativeAdversarialNetwork))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:uses ml:ParameterEfficientFineTuning))

Contrast Relationships

SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:contrastsWith ml:DomainGeneralisation))
SubClassOf(ml:DomainAdaptation
  ObjectSomeValuesFrom(ml:contrastsWith ml:ContinualLearning))

About

Domain Adaptation emerged as a formal research topic at the intersection of statistical learning theory and applied machine learning, catalysed by the observation that real-world data is non-stationary and that deploying models outside their training distribution is the rule rather than the exception. The phenomenon has both theoretical characterisation and devastating practical consequences: a speech recognition system trained on American English achieves word error rates above 30% on British regional accents; an object detection model trained on Californian dashcam footage misses pedestrians in rainy Edinburgh streets; a clinical NLP model trained on US electronic health record syntax fails to parse NHS clinical note conventions. The scale of this problem is substantial — it is estimated that a majority of AI deployment failures in regulated industries trace to unacknowledged distribution shift between training and operational environments.

The theoretical foundations were laid by Ben-David et al. (2010), who derived PAC-style generalisation bounds for the adaptation setting. Their main theorem states that the target error is bounded by the source error, half the H-divergence d_H(D_S, D_T) between source and target distributions (measured with respect to the hypothesis class H), and the error of the ideal joint hypothesis λ* that minimises the combined source and target error. Formally: ε_T(h) ≤ ε_S(h) + d_H(D_S, D_T) / 2 + λ* for any hypothesis h in H. This bound formalises three actionable design principles: (i) minimise source error (good source task performance is necessary); (ii) minimise H-divergence (align representations to reduce domain distance); and (iii) ensure the target task is related to the source task (a condition on λ* being small). The H-divergence can be estimated empirically by training a binary domain classifier — if the classifier achieves near-chance accuracy, d_H ≈ 0 and the distributions are well-aligned; if it achieves high accuracy, the distributions are far apart and adaptation is harder.

The applied landscape bifurcates cleanly between the classical era (pre-2016) and the deep learning era. Classical domain adaptation relied on hand-crafted feature mappings, kernel-based distributional divergence measures such as Maximum Mean Discrepancy (MMD, Gretton et al. 2012), and subspace alignment methods (SA, Fernando et al. 2013) that align the principal components of source and target PCA subspaces. Importance-weighting methods (Kernel Mean Matching, KLIEP, uLSIF) provided theoretically grounded covariate-shift correction by re-weighting source examples by their density ratio P_T(x)/P_S(x). These classical methods remain competitive for low-dimensional tabular data and small-sample regimes where deep learning is statistically inefficient.

The deep learning era transformed domain adaptation by making distributional alignment an end-to-end differentiable objective integrated into neural network training. Ganin and Lempitsky’s DANN (2015, later published in JMLR 2016) introduced the gradient reversal layer — a parameter-free operation that during forward passes acts as an identity transformation but during backpropagation multiplies gradients by -λ (where λ is a schedule parameter that increases during training). This routes gradients from a domain discriminator through the shared encoder with reversed sign, forcing the encoder to produce representations that fool the discriminator — i.e., are statistically indistinguishable between source and target domains. The adversarial loss is combined with the task classification loss on source data, producing an encoder that is simultaneously discriminative for the source task and domain-invariant. This adversarial formulation was rapidly extended by CDAN (conditional adversarial, conditioning on class predictions, 2018), MDD (maximum classifier discrepancy, 2018), DAAN (dynamic adversarial adaptation network, 2019), and many variants, converging on the principle that a domain-discriminative signal is exactly the supervision needed to remove domain-specific information from representations.

Optimal transport (OT) methods provide a geometrically principled alternative to adversarial alignment. The earth-mover’s distance (Wasserstein distance) measures the minimum cost of transporting probability mass from the source to the target distribution, accounting for the geometry of the feature space. Courty et al. (2017) showed that optimal transport can directly regularise domain adaptation: source samples are re-weighted by OT transport plan coefficients to match the target distribution, and the re-weighted source classifier achieves good target accuracy. DeepJDOT (Damodaran et al., 2018) minimised OT in joint feature-label space end-to-end within a deep network, simultaneously aligning features and ensuring class-conditional alignment.

With the maturation of Large Language Models and Foundation Models after 2020, domain adaptation increasingly operates at a higher level of abstraction. Rather than aligning intermediate representations of task-specific models, practitioners now adapt large pre-trained models to new domains primarily through continued pre-training on domain corpora followed by Parameter Efficient Fine Tuning (LoRA, (IA)³, adapter layers, prefix tuning). This paradigm shift — from feature-space alignment to parameter-efficient behavioural steering — is economically dominant because it amortises the enormous cost of pre-training across all downstream adaptations. The theoretical connection to classical domain adaptation is maintained: continued pre-training on target-domain text (medical notes, legal documents, scientific papers) updates the model’s prior over the target domain’s token distribution, reducing the implicit KL-divergence between model distribution and target domain distribution before task-specific fine-tuning. LoRA (Hu et al., 2021) and its federated extensions (FedEx-LoRA, LoRA-FAIR, FedMomentum — all 2024-2025) have become the standard parameter-efficient adaptation mechanism, reducing the number of trainable parameters by 10,000x compared to full fine-tuning while matching or approaching full fine-tuning performance on most domain adaptation benchmarks.

Source-free domain adaptation (SFDA) represents a practically critical emerging paradigm motivated by data-privacy constraints: the source domain data is unavailable at adaptation time — either because it has been deleted for GDPR compliance, is held by a different legal entity, or is too large to transfer — and the only resource is the pre-trained source model plus unlabelled target data. SFDA methods exploit the source model’s predictions, intermediate features, or confidence distributions to guide alignment without requiring source data. SHOT (Liang et al., 2020) freezes the source-trained classifier head and adapts only the encoder via information maximisation: the encoder is trained to simultaneously maximise mutual information between target examples and their predicted class assignments (seeking discriminative features) while minimising conditional entropy on individual examples (seeking confident predictions). Recent SFDA variants (NRC, AaD, DINE, 2021-2023) further exploit nearest-neighbour structure in target embedding space to generate reliable pseudo-labels. Results on Office-Home and DomainNet show that SFDA methods trained on source model outputs alone can match standard UDA methods that have access to source data, making SFDA the natural choice for privacy-preserving deployment in healthcare and finance.

A 2024 ICLR analysis (OpenReview, Rethinking Effectiveness of UDA Methods) raised an important challenge to the field: on strong modern backbones (ViT-L/16 pre-trained on CLIP or DINO), the benefit of explicit UDA alignment methods over simple source-pre-trained fine-tuning diminishes substantially, with standard fine-tuning on pseudo-labelled target data approaching or matching adversarial UDA methods. This suggests that the implicit domain alignment achieved by large-scale pre-training on diverse web data already reduces the practical domain gap for many standard benchmarks, and that the research frontier must move toward harder distribution shifts, open-set settings, and real-world evaluation beyond the standard Office/DomainNet suite.

Key Mechanisms and Algorithms

Feature Alignment Methods

• DANN (Domain-Adversarial Neural Network, Ganin & Lempitsky 2015) — gradient reversal layer propagates domain discriminator loss backward through the shared encoder with reversed sign, forcing domain-invariant representations. Trained jointly for source classification accuracy and domain indistinguishability.

• Deep CORAL (Sun & Saenko 2016) — aligns second-order statistics (covariance matrices) of source and target deep features; computationally cheap and surprisingly effective.

• MMD-based methods (Long et al. 2015, DAN; Long et al. 2017, JAN) — embed distributional distances into the loss as multi-kernel MMD terms; joint adaptation networks align joint distributions P(X,Y) rather than marginal P(X).

• Optimal Transport (Courty et al. 2017, JDOT; Damodaran et al. 2018, DeepJDOT; Fatras et al. 2021, WDGRL) — earth-mover’s distance provides a geometrically principled alignment objective; DeepJDOT minimises joint distribution OT in feature-label space end-to-end.

• CDAN (Conditional Adversarial Domain Adaptation, Long et al. 2018) — conditions the adversarial domain discriminator on the classifier’s output, capturing the joint relationship between feature and class predictions.

  Instance Re-Weighting Methods

• Kernel Mean Matching (Gretton et al. 2009) — directly estimates importance weights P_target(x)/P_source(x) via RKHS-norm matching without estimating either density explicitly.

• TrAdaBoost (Dai et al. 2007) — boosting variant that assigns low weights to source instances that are likely to differ from the target distribution.

• Density ratio estimation (KLIEP, uLSIF) — parametric models for the density ratio enable principled re-weighted empirical risk minimisation.

  Self-Training and Pseudo-Labelling Methods

• Self-Training — iteratively assign pseudo-labels to high-confidence (low-entropy) target predictions; re-train on augmented dataset; stabilised by confidence thresholds and consistency regularisation.

• Mean Teacher (Tarvainen & Valpola 2017) — exponential moving average of model weights creates a stable teacher that assigns pseudo-labels; student is trained with both supervised (source) and consistency (target) loss.

• SHOT (Shot for Source-Free Adaptation, Liang et al. 2020) — freezes source-trained classifier head, adapts encoder via information maximisation (mutual information + entropy minimisation) without any source data.

• NRC (Neighbourhood Reciprocal Classifier, Yang et al. 2021) — exploits local structure of target embeddings to mine reliable pseudo-labels via nearest-neighbour consistency.

  Generative Methods

• CycleGAN-based adaptation — translates source images into target visual style (or vice versa) using unpaired image-to-image translation; task model trained on style-adapted source images avoids feature-space alignment entirely.

• UNIT (Liu et al. 2017) — couples VAE and GAN to learn shared latent space for paired domain translation with implicit distribution alignment.

  Large-Model Adaptation Methods (post-2020)

• Continued pre-training on target-domain corpora to update prior language distributions before task fine-tuning; BioBERT, ClinicalBERT, LegalBERT all follow this paradigm.

• LoRA (Hu et al. 2021) — low-rank decomposition of weight-update matrices; predominant PEFT method for adapting LLMs to new domains; 2024-2025 federated variants (FedLoRA, LoRA-FAIR, FedEx-LoRA) extend LoRA to cross-client domain heterogeneity.

• Adapter modules (Houlsby et al. 2019) — small bottleneck layers inserted into frozen transformer blocks; only adapters are trained, enabling many domain adaptations to share a single backbone.

• Prefix tuning / prompt tuning — prepend trainable token vectors to frozen model input; task and domain are encoded in the soft prefix rather than in model weights.

  Applications and Use Cases

  Natural Language Processing

  Sentiment analysis transfer across product categories is the canonical NLP adaptation benchmark (Amazon Review Multi-Domain, 4 domains; then expanded to 25+ product categories). The practical motivation is that a company may have abundant labelled reviews for books and electronics but need sentiment analysis for kitchen appliances or sports equipment, where labelled data is sparse. Named entity recognition (NER) adaptation across biomedical, legal, and financial text domains bootstraps specialised models from general English NER: BioBERT adapts BERT via continued pre-training on PubMed abstracts and PubMed Central full-text articles, achieving significant NER gains for biomedical named entities (genes, proteins, diseases) without the 10,000+ manually labelled examples that training from scratch would require. Cross-lingual transfer — adapting models pre-trained on high-resource languages (English, Chinese, German) to low-resource target languages (Yoruba, Swahili, Welsh, Scottish Gaelic) — constitutes a linguistically rich form of domain adaptation in which the distribution shift encompasses vocabulary, morphology, and syntax simultaneously. Clinical NLP uses continued pre-training on de-identified medical notes to adapt GPT-class models to electronic health record (EHR) understanding without sharing raw patient data across institutions; this is the dominant technique in US Epic-integrated and NHS federated AI pipelines. Machine translation quality improvement via domain adaptation is widely practised: translation systems pre-trained on general parallel corpora are adapted to specific domains (medical, legal, technical) using small in-domain parallel datasets or monolingual in-domain text via continued pre-training, yielding 5-10 BLEU point improvements on domain-specific test sets.

  Computer Vision

  Synthetic-to-real (Syn2Real) adaptation is the canonical and most extensively studied computer vision UDA task. GTA5 (a photorealistic game engine rendering of urban driving with 24,966 annotated images) and SYNTHIA (a virtual city with 9,400 densely annotated images) serve as synthetic source domains; Cityscapes (2,975 real German urban driving images with dense panoptic annotation) and KITTI serve as target domains. The domain gap between synthetic renders and real imagery includes texture differences (game-engine materials versus real lens physics), lighting models, motion blur, and geometric distribution differences in object frequencies. Adversarial UDA methods applied to this benchmark achieve approximately 50-55 mIoU on Cityscapes semantic segmentation from a GTA5 source, compared to a lower bound of ~30 mIoU for direct transfer and an upper bound of ~70 mIoU for fully supervised models. Cross-domain medical imaging adaptation is arguably the highest-impact applied domain: CT-trained tumour segmentation models achieve dramatically degraded Dice scores on MRI, because the intensity distributions, acquisition artefacts, and tissue contrast are entirely different modalities. Adversarial adaptation, OT-based alignment, and source-free adaptation (particularly relevant given GDPR constraints on patient data sharing across hospitals) are all applied to this problem. UCL CMIC’s prostate MRI adaptation work (UCL Discovery 10165221) demonstrated that adversarial domain adaptation enables a model trained on one hospital’s Siemens MRI scanner to generalise to another hospital’s Philips scanner without any additional annotation. Remote sensing analysis transfers Sentinel-2 or Landsat-trained land-use classifiers to PlanetScope commercial imagery (3m resolution versus 10m) or historical aerial survey imagery with different spectral characteristics, temporal coverage, and spatial resolution.

  Speech and Audio

  Speaker adaptation in Automatic Speech Recognition reduces word error rates for individual speakers, particularly non-native speakers and speakers with atypical prosody or dysarthria, using very few labelled utterances (sometimes as few as 30 seconds) from the target speaker. Domain adaptation for ASR also handles cross-accent transfer (Received Pronunciation to Glaswegian Scots, or General American to African American English), cross-channel transfer (studio recordings to telephone speech or distant-field microphone), and cross-noise-condition transfer (clean speech to cafeteria noise or wind noise). Commercial ASR APIs from Amazon, Google, and Microsoft all employ speaker adaptation techniques behind their customised model APIs. For text-to-speech synthesis, speaker adaptation enables voice cloning — adapting a TTS model to a new speaker using only a few minutes of reference audio, a capability now commercially deployed in Microsoft Azure Neural TTS and Amazon Polly Custom Voice.

  Autonomous Systems and Robotics

  Sim-to-real transfer for robot policy learning is the dominant domain adaptation challenge in embodied AI. Policies trained in physically simulated environments (MuJoCo, PyBullet, IsaacGym) must adapt to real robotic hardware where sensor noise, actuation latency, friction coefficients, and contact dynamics differ from simulation. Domain randomisation (training on a distribution over simulation parameters) and explicit sim-to-real adaptation (using a small real-world interaction budget to fine-tune) are complementary techniques. In Autonomous Driving, domain adaptation handles geographic transfer (US to EU road marking conventions, right-hand to left-hand traffic), weather transfer (clear weather training to rain, fog, snow inference), time-of-day transfer (daytime to night-time), and sensor modality transfer (RGB camera to LiDAR or radar). The DVSA’s autonomous vehicle trial support programme (DVSA Business Plan 2025-26) explicitly acknowledges the cross-condition generalisation challenge as a key regulatory concern for the UK’s Automated Vehicles Act 2024.

  Scientific Computing

  Physics-Informed Neural Networks (PINNs) are adapted to new physical regimes or boundary conditions by fine-tuning on the new governing equations or domain geometry while retaining prior knowledge of physical constraints from source domains. This is relevant for computational fluid dynamics, heat transfer, and electromagnetic simulation where solving full PDEs from scratch is computationally expensive. Molecular property prediction models trained on large general chemical databases (ChEMBL, ZINC, PubChem) are adapted to specific assay types (ADMET properties, binding affinity for specific protein targets) using datasets of 50-500 molecules, orders of magnitude below the data requirements for training from scratch. Genomics models adapted from model organism data (mouse, zebrafish, Drosophila) to human clinical data exploit evolutionary conservation of molecular mechanisms while correcting for species-specific regulatory differences.

  Federated and Privacy-Preserving Settings

  Federated Learning with non-IID local distributions across clients constitutes a multi-source domain adaptation problem: each client’s local data represents a different domain with distinct distribution P_i(X, Y), and the global aggregated model must adapt to each client’s local distribution without direct data sharing. Source-free domain adaptation methods are naturally applicable in federated settings since they require only the source model (globally aggregated) and unlabelled target data (each client’s local data). FedLoRA, FedEx-LoRA, and LoRA-FAIR (all 2024-2025) combine the communication efficiency of LoRA fine-tuning (only the low-rank matrices are communicated rather than full model parameters) with domain-heterogeneity-aware aggregation strategies that weight client contributions by estimated distribution similarity to the target. The NHS England federated AI platform (in development 2024-2025) applies these techniques to enable NHS trusts to collaboratively fine-tune clinical NLP models without sharing patient records across trust boundaries, with differential privacy guarantees providing formal privacy protection against membership inference attacks.

  Key Terminology Glossary

• Domain shift (also dataset shift or distribution shift) — the statistical phenomenon in which the joint distribution P(X, Y) differs between the source (training) and target (deployment) environments. May be in the marginal input distribution P(X) (covariate shift), the class prior P(Y) (label shift), or the conditional P(Y|X) (concept drift). All three forms occur in practice, often simultaneously.

• Covariate shift — the special case of domain shift where P_S(X) ≠ P_T(X) but P_S(Y|X) = P_T(Y|X). The conditional label distribution is preserved; only the input distribution shifts. Covariate shift underlies importance-weighting methods such as Kernel Mean Matching, where re-weighting source examples by P_T(x)/P_S(x) corrects for the mismatch without architectural changes.

• Label shift (prior probability shift) — the case where P_S(Y) ≠ P_T(Y) but P_S(X|Y) = P_T(X|Y). Class frequencies differ but class-conditional feature distributions are preserved. Addressed by class weight re-weighting estimated from source model predictions on unlabelled target data via Black Box Shift Estimation (BBSE, Lipton et al., 2018).

• Concept drift — the hardest case: P_S(Y|X) ≠ P_T(Y|X). The labelling function itself changes. Requires target-labelled data or strong domain knowledge; unsupervised alignment cannot correct a changed labelling function.

• H-divergence — a distribution divergence measure defined with respect to a hypothesis class H, measuring how well a binary domain classifier from H can distinguish source from target samples. Appears in Ben-David et al. (2010) bounds as the central term measuring adaptation difficulty. Estimated empirically by training a domain classifier: near-chance accuracy implies d_H ≈ 0 (good alignment); high accuracy implies large d_H (hard adaptation).

• Maximum Mean Discrepancy (MMD) — a kernel-based statistic measuring the distance between mean embeddings of two distributions in a Reproducing Kernel Hilbert Space (RKHS). MMD(P, Q) = 0 iff P = Q (for characteristic kernels). Minimisable by stochastic gradient descent when using a differentiable kernel, implemented in the DAN and JAN methods. The kernel-based complement to adversarial discriminator-based alignment.

• Gradient reversal layer (GRL) — a pseudo-layer in DANN (Ganin & Lempitsky, 2015) that acts as identity in forward propagation but multiplies gradients by -λ in backpropagation. When placed between the shared encoder and domain classifier, forces the encoder to maximise domain classification loss (becoming domain-agnostic) while the task classifier minimises task loss — implementing the domain adaptation minimax objective in a single end-to-end pass.

• Pseudo-labelling — assigning predicted class labels (pseudo-labels) to unlabelled target examples, then using these labels as soft or hard supervision for continued training. Quality depends on confidence thresholds (keeping only predictions with softmax probability > 0.9 filters noise), calibration, and curriculum scheduling that gradually accepts lower-confidence predictions as the model improves.

• Source-free domain adaptation (SFDA) — a setting where source domain data is unavailable during adaptation — due to GDPR deletion, data ownership, or size constraints — with only the pre-trained source model and unlabelled target data. SFDA methods exploit model predictions (pseudo-labelling), internal statistics (batch norm statistics), or information-theoretic criteria (entropy minimisation, mutual information maximisation) applied to target data alone. SHOT (Liang et al., 2020) is the seminal SFDA method.

• Domain randomisation — a sim-to-real transfer technique where simulation environment parameters (textures, lighting, physics, noise) are randomised during training so that the real-world environment lies within the training distribution. Unlike post-hoc domain adaptation, domain randomisation modifies the source domain during training to implicitly cover the target domain.

• Negative transfer — the phenomenon where applying transfer learning or domain adaptation hurts target task performance compared to training on target data alone. Occurs when source and target are sufficiently dissimilar that source inductive bias is misleading. H-divergence metrics help predict and diagnose negative transfer risk before committing to adaptation.

• Optimal transport (OT) — a mathematical framework computing the minimum-cost plan for transforming one distribution into another, with cost measured by a ground metric on the sample space (typically Euclidean distance in feature space, or joint feature-label distance). Used in domain adaptation to compute correspondences between source and target samples (Courty et al., 2017; DeepJDOT, Damodaran et al., 2018), enabling feature alignment and label propagation via the transport plan.

• Parameter-efficient fine-tuning (PEFT) — a family of techniques adapting large pre-trained models to new domains while updating only a small fraction of parameters. Dominant methods: LoRA (low-rank matrix decomposition of weight updates), adapters (small bottleneck modules in frozen transformer layers), prefix tuning (prepending trainable token vectors), and (IA)³ (element-wise activation rescaling). Reduces trainable parameters by 10,000-100,000x versus full fine-tuning while approaching full fine-tuning performance on most domain adaptation tasks.

• WILDS benchmark — a collection of real-world distribution shift problems (Koh et al., 2021, Stanford): wildlife camera traps (iWildCam), pathology slide cancer detection (Camelyon17), satellite land use (fMoW), Amazon reviews, and online comment toxicity (CivilComments). Enforces standardised splits where validation and test sets come from different environments than training, providing realistic out-of-distribution evaluation with consistent protocols.

• DomainBed — a testbed (Gulrajani & Lopez-Paz, 2021) implementing 7 domain generalisation datasets with consistent protocols and fixed hyperparameter search budgets, enabling fair comparison that had been impossible due to inconsistent experimental setups across papers.

  Contrasts and Boundaries

• vs Domain Generalisation — domain generalisation trains a model to work on any unseen target domain without any target data at test time; no adaptation step is performed. Domain adaptation assumes access to (unlabelled) target data at adaptation time, making it strictly more powerful but requiring target data collection. The boundary is methodologically significant: domain generalisation must be robust to all possible target shifts, while domain adaptation can be tailored to a specific known target distribution.

• vs Multi Task Learning — MTL trains jointly on multiple related tasks and their labels simultaneously; the goal is generically better representations that serve all tasks well, not bridging a specific source-to-target distributional gap. Domain adaptation focuses on a single task (or task family) across distributions. The two can be combined: multi-task pre-training followed by domain adaptation to a specific target is a common practical pipeline.

• vs Continual Learning — continual learning addresses sequential distribution shifts over time whilst explicitly avoiding catastrophic forgetting of previous tasks; domain adaptation treats the source-to-target shift as a single two-stage process. The emerging field of continual domain adaptation addresses the combination: initial adaptation at deployment followed by continual drift handling over time.

• vs Meta Learning — meta-learning (few-shot learning-to-learn) learns a model initialisation or adaptation procedure enabling rapid adaptation from very few labelled examples by meta-training across many tasks. Domain adaptation focuses on distributional alignment, potentially without any labelled target data. The two are complementary: a meta-learned initialisation can reduce the number of labelled target examples needed for adaptation.

• vs Fine Tuning — standard fine-tuning assumes some labelled target data and simply updates model weights on it; it does not address the unsupervised case and does not explicitly minimise domain divergence. Domain adaptation provides principled handling of the unsupervised or semi-supervised case where target labels are absent or very scarce, and explicitly addresses distributional shift as an objective.

• vs Knowledge Distillation — knowledge distillation transfers knowledge from a teacher model to a student model, typically for model compression; it does not inherently address distribution shift. However, distillation is used in domain adaptation contexts: distilling a source-domain teacher into a target-adapted student, or using ensemble teacher predictions as soft labels for pseudo-labelling in source-free domain adaptation.

  Theoretical Foundations in Depth

  The theoretical framework for domain adaptation rests on three interlocking pillars: PAC-learning generalisation bounds, distributional divergence theory, and importance weighting theory. Each provides a distinct lens for understanding why adaptation works or fails and for designing algorithms with theoretical backing.

  The PAC-learning perspective was comprehensively formalised by Ben-David et al. (2010) through four influential theorems. The first establishes that for any hypothesis h in H, the target error ε_T(h) is bounded by ε_S(h) + d_H(D_S, D_T)/2 + λ*, where λ* = min_{h’ ∈ H}[ε_S(h’) + ε_T(h’)] is the error of the ideal joint hypothesis. The second shows that an empirical estimate of d_H can be computed by training a domain classifier on pooled source and target samples: d̂_H(D_S, D_T) = 2(1 - 2·err(domain classifier)). The third theorem provides sample complexity bounds for learning d_H from finite samples. The fourth extends the framework to the multi-source setting where multiple source domains with different weights contribute to a convex combination source distribution. These bounds motivate the adversarial training approach (minimise d_H by training the encoder adversarially against the domain classifier) and caution against adaptation when λ* is large (tasks are too dissimilar for any shared hypothesis to work well).

  Kernel-based divergence theory (Gretton et al., 2012) provides the MMD framework as a computationally tractable distributional distance. The squared MMD between distributions P and Q is: MMD²(P, Q) = E_{x,x’ ~ P}[k(x, x’)] - 2·E_{x ~ P, y ~ Q}[k(x, y)] + E_{y,y’ ~ Q}[k(y, y’)], where k is a positive-definite kernel. For a Gaussian RBF kernel k(x, x’) = exp(-‖x - x’‖²/2σ²), this is a characteristic kernel and MMD(P, Q) = 0 iff P = Q, providing a complete distributional comparison. The finite-sample estimator of MMD² using n source and m target samples is an unbiased estimator with concentration bounds of order O(1/√n + 1/√m), making it practical for empirical distribution alignment. The DAN method (Long et al., 2015) minimises a multi-kernel MMD variant (MK-MMD) using a mixture of Gaussian RBF kernels with different bandwidths, providing better coverage of different length scales of distributional difference.

  Optimal transport theory (Villani, 2008; Peyré & Cuturi, 2019) provides the most geometrically principled framework for distributional alignment. The p-Wasserstein distance W_p(P, Q) is defined as the p-th root of the minimum expected p-th power transport cost: W_p^p(P, Q) = inf_{γ ∈ Γ(P,Q)} E_{(x,y) ~ γ}[d(x, y)^p], where Γ(P, Q) is the set of all joint distributions (couplings) with marginals P and Q, and d is a ground metric on the sample space. For p=1, this is the Wasserstein-1 (or earth-mover’s) distance. The dual formulation (Kantorovich-Rubinstein duality) expresses W_1(P, Q) = sup_{f: Lip(f) ≤ 1} [E_P[f(x)] - E_Q[f(x)]], which is approximated in Wasserstein GAN training by the critic network constrained to be 1-Lipschitz. In domain adaptation, Courty et al. (2017) show that the optimal transport plan γ between source and target provides natural sample weights for re-weighting the empirical risk: each source sample x_i is weighted by its transport mass γ_ij to target sample x_j, so that the re-weighted source distribution approximates the target.

  Importance weighting theory addresses covariate shift directly. Under covariate shift, the optimal re-weighting is the density ratio w(x) = P_T(x)/P_S(x). Since directly estimating two densities and dividing is numerically unstable, dedicated density ratio estimation methods are used: Kernel Mean Matching (KMM, Gretton et al. 2009) directly estimates weights by matching kernel mean embeddings; KLIEP (Kullback-Leibler Importance Estimation Procedure, Sugiyama et al. 2008) minimises the KL divergence between the importance-weighted source distribution and the target distribution; uLSIF (unconstrained Least-Squares Importance Fitting, Kanamori et al. 2009) directly minimises the squared error of the ratio estimator. When the density ratio is highly skewed (heavy tail in P_T relative to P_S), importance weighting variance blows up; covariate shift can be “pathological” (Kpotufe & Martinet, 2021), motivating hybrid approaches that combine re-weighting with representation alignment.

  Self-training theory has been substantially advanced through the lens of pseudo-label quality and consistency regularisation. Sohn et al. (2020) established that FixMatch-style consistency regularisation — requiring that predictions on strongly augmented target views agree with predictions on weakly augmented views — implicitly implements a form of pseudo-labelling with built-in augmentation invariance. Lee (2013) showed that pseudo-labelling with a confidence threshold is equivalent to minimising the entropy of predictions on unlabelled data, connecting to the information-maximisation principle in SHOT (Liang et al., 2020). The mean teacher (Tarvainen & Valpola, 2017) stabilises pseudo-labels through exponential moving average weight averaging, providing a consistently more accurate teacher than the current student model, which is critical for avoiding early confirmation bias (the failure mode where early pseudo-label errors reinforce themselves across training epochs).

  Benchmarks and Evaluation Standards

• Office-31 (Saenko et al. 2010) — 31 object categories across Amazon product images, DSLR photos, and Webcam photos; the foundational small-scale UDA vision benchmark.

• Office-Home (Venkateswara et al. 2017) — 65 categories across four domains: Art, Clipart, Product, and RealWorld; harder than Office-31 due to larger domain gaps.

• VisDA (Peng et al. 2017) — large-scale synthetic-to-real benchmark; 200K+ images from 12 classes; challenging due to photorealism versus render gap.

• DomainNet (Peng et al. 2019) — ~0.6 million images from 345 categories across 6 domains (Infograph, Quickdraw, Real, Sketch, Clipart, Painting); currently the largest and most challenging multi-source UDA benchmark.

• WILDS (Koh et al. 2021, Stanford) — standardised in-the-wild distribution shift benchmark across ecology (iWildCam), medical imaging (Camelyon17), satellite imagery (fMoW), text (Amazon reviews, CivilComments); maintained with consistent splits and evaluation protocols.

• DomainBed (Gulrajani & Lopez-Paz 2021) — testbed implementing 7 popular domain generalisation datasets with consistent experimental protocols, enabling fair comparison across methods.

• Amazon Review Multi-Domain — NLP benchmark for sentiment adaptation across product review domains.

• GLUE / SuperGLUE domain-shifted splits — adapted NLP benchmarks that include out-of-domain evaluation.

  Academic Context

  Domain adaptation research has deep roots in statistical learning theory, covariate shift estimation (Shimodaira 2000), and importance weighting (Sugiyama et al. 2008). Blitzer et al. (2006) introduced structural correspondence learning for NLP adaptation. The theoretical framework was comprehensively formalised by Ben-David, Blitzer, Crammer, Kulesza, Pereira, and Wortman Vaughan (2010) in a paper that remains the canonical reference. Gretton et al.’s (2012) MMD formulation provided the kernel-based empirical counterpart.

  The adversarial era began with Ganin & Lempitsky (2015) and was rapidly followed by the feature-level and output-level extensions of Tsai et al. (2018) for semantic segmentation, Long et al.’s DAN/JAN/CDAN family, and Saito et al.’s MCD (Maximum Classifier Discrepancy, 2018). Optimal transport methods (Courty et al. 2017, Damodaran et al. 2018) provided geometric grounding. The source-free paradigm was crystallised by Liang et al.’s SHOT (2020) and subsequent variants.

  Post-2020, the field merged substantially with the PEFT literature: Hu et al. (2021) LoRA, Houlsby et al. (2019) adapters, and Lester et al. (2021) prompt tuning collectively redefined the parameter budget for domain adaptation of large models. Comprehensive book-length treatments appeared from Springer Nature in 2024 (Unsupervised Domain Adaptation: Recent Advances and Future Perspectives, Kherallah et al.), reflecting the field’s maturity.

  Recent (2024-2025) trends include: (i) rethinking whether adversarial UDA methods still provide benefit over strong backbones (OpenReview ICLR 2024 analysis); (ii) active domain adaptation that uses uncertainty or diversity criteria to select which target examples to annotate; (iii) federated domain adaptation combining PEFT methods with privacy-preserving client heterogeneity (FedLoRA, FedEx-LoRA, FedMomentum, 2024-2025); (iv) large-model domain adaptation via continued pre-training plus instruction tuning for clinical and legal LLMs.

  Current Landscape (2026)

  By 2026, the conceptual distinctions between domain adaptation and PEFT-based LLM fine-tuning have substantially blurred in practice. The dominant operational paradigm for adapting large models to new domains is: continued pre-training on a domain corpus (hundreds of millions of tokens) followed by instruction tuning and LoRA-based fine-tuning on a task-specific dataset. This is applied routinely for clinical NLP (hospital EHR systems), legal NLP (contract analysis, regulatory compliance), and financial NLP (earnings call analysis, credit risk modelling).

  For computer vision, source-free UDA methods have gained commercial relevance in regulated healthcare AI deployments where sharing training data across hospital boundaries is legally restricted under GDPR and national health data protection law. Radiograph, pathology slide, and endoscopy image models are routinely adapted to new hospital equipment without accessing the source hospital’s patient data.

  In autonomous driving, domain adaptation is embedded in the perception stacks of OEMs: synthetic-to-real adaptation (NeRF-generated training data adapted to sensor footage) and geographic transfer dominate. The UK’s Automated Vehicle Act 2024 and the DVSA’s continued support for autonomous vehicle trials (cited in DVSA Business Plan 2025-26) reflects the regulatory context within which domain adaptation for safety-critical perception is being deployed.

  Federated domain adaptation is an active intersection with UK regulatory priorities around data sharing in the NHS: models trained on NHS England data must adapt to Welsh Health Board or Scottish NHS data distributions without raw data exchange. Federated LoRA methods (FedEx-LoRA, LoRA-FAIR, FedMomentum — all 2024-2025) are directly applicable to this setting.

  Benchmark performance on DomainNet has plateaued near 90% accuracy for standard UDA tasks, driving the community toward harder multi-target, open-set, and partial domain adaptation settings where target categories may be a subset or superset of source categories.

  UK Context

  UK universities have made substantial contributions to domain adaptation research across both theory and application. UCL’s Centre for Medical Image Computing (CMIC) is a leading group for domain adaptation in medical imaging: their work on domain adaptation for novel imaging modalities, including prostate MRI (UCL Discovery ID:10165221), directly addresses the cross-scanner generalisation problem in clinical AI. UCL’s involvement in the UK Biobank and related health data platforms positions it as a key site for federated and source-free domain adaptation in genomics and imaging.

  The University of Edinburgh’s Informatics Research Institute has active research in low-resource NLP and cross-lingual adaptation, including language modelling for Scottish Gaelic, Welsh, and other UK minority languages. This work sits squarely within the domain adaptation literature for low-resource linguistic settings. Edinburgh’s Alan Turing Institute fellows have contributed to distribution shift theory and robustness.

  The Alan Turing Institute, headquartered in London, coordinates UK-wide machine learning research with explicit workstreams in distribution shift, fairness under covariate shift, and trustworthy AI deployment — all of which intersect domain adaptation. Turing’s Health Programme and Data Study Groups regularly tackle cross-site medical AI generalisation, which requires domain adaptation methodology.

  At Cambridge, the Computational and Biological Learning Laboratory (CBL) has theoretical and empirical work in generalisation under distribution shift, and the Cambridge Machine Learning Group has contributed to Bayesian treatments of domain shift. King’s College London has domain adaptation research in medical imaging (cardiac MRI across scanner protocols).

  In Northern England, Sheffield Hallam’s Industry-Academia Centre and the University of Manchester’s AI group have addressed domain adaptation for industrial sensor data — vibration and acoustic signals from rotating machinery — where sensor placement, equipment age, and operating regime all introduce distribution shift. Leeds’ Institute of Data Analytics has applied domain adaptation to crime and traffic prediction across UK cities. Newcastle’s Digital Economy researchers have studied domain adaptation in satellite image analysis for Northern English urban change detection.

  UK industry deployments include: Arm’s cross-platform neural network deployment tools that implicitly require robustness to hardware-induced distribution shift in quantised inference; BT’s network anomaly detection systems that adapt models from one regional network configuration to another; NHS England’s federated AI platform (FedGPT project, 2024-2025) that applies PEFT domain adaptation across NHS trust boundaries without sharing patient data.

  Practical Deployment Considerations

  Deploying domain adaptation in production raises engineering and operational challenges that extend well beyond the algorithm selection question. These practical considerations often determine whether domain adaptation projects succeed or fail independently of their theoretical soundness.

  Distribution shift detection is the operational precursor to domain adaptation: before adapting a model, practitioners must reliably detect that meaningful distribution shift has occurred. Population Stability Index (PSI), Jensen-Shannon divergence between feature histograms, Maximum Mean Discrepancy tests on incoming data streams, and learned covariate shift classifiers are used to continuously monitor incoming data distributions and trigger re-adaptation pipelines when shift is detected. Without monitoring, models degrade silently until business-level metrics (accuracy, error rates, customer complaints) reveal the problem — typically many weeks after the shift began.

  Target domain data curation is a critical and often underestimated step. The quality of unlabelled target data directly affects adaptation quality: web-scraped or automatically collected target data may contain distribution contamination (data from outside the true target distribution), label noise (for pseudo-labelling methods), or representational bias (if target collection systematically over-represents certain subpopulations). For medical imaging, target data curation requires DICOM metadata filtering to ensure that cases come from the correct scanner protocol; for NLP, target domain curation requires careful deduplication and near-duplicate detection at the character n-gram and sentence embedding level.

  Adaptation pipeline orchestration requires careful management of the training schedule, particularly for iterative pseudo-labelling methods where premature convergence to poor pseudo-labels (confirmation bias) is a well-documented failure mode. Best practices include: warm starting with feature extraction only for the first few epochs; gradually unfreezing deeper layers; using exponential learning rate decay; applying strong data augmentation to reduce pseudo-label confidence on out-of-distribution inputs; and reserving a small labelled validation set from the target domain for early stopping.

  Evaluation validity is a systematic problem in domain adaptation research and deployment. Practitioners frequently evaluate on the available subset of target data that happens to have labels, which may not be representative of the unlabelled target distribution the adaptation is actually trained on. Proper evaluation requires a held-out test set from the target domain that was not used in adaptation, drawn from the same distribution as the unlabelled adaptation data, with labels obtained independently (e.g., by human annotation or a gold-standard system). For WILDS-style real-world benchmarks, the train/validation/test splits are carefully designed to avoid this pitfall.

  Compute and memory budgets for domain adaptation depend critically on model scale. For adaptation of small models (BERT-base, 110M parameters), full fine-tuning costs roughly 1 GPU-hour per 10K target examples. For large models (LLaMA-3.1-70B, 70B parameters), even LoRA adaptation costs 100-400 GPU-hours per adaptation run, making iterative pseudo-labelling schemes that require multiple adaptation passes computationally expensive. Test-time adaptation methods (TENT, T3A) reduce this cost by adapting at inference time without offline training passes, but at the cost of higher per-inference latency and reduced adaptation quality compared to offline methods.

  Model versioning and rollback for domain-adapted models requires careful tracking of: the source model checkpoint used as starting point, the unlabelled target data used for adaptation (including data provenance and collection timestamp), the adaptation hyperparameters (learning rate schedule, pseudo-label threshold, number of epochs), and the evaluation metrics on a held-out target validation set. Without this provenance tracking, adapted models cannot be debugged when they fail on new target subpopulations, and regulators cannot audit the adaptation process in regulated deployments.

  Future Directions (2026-2030)

• Theoretically grounded source-free adaptation — developing generalisation bounds for SFDA settings analogous to Ben-David et al.’s classical bounds, enabling principled algorithm design without access to source data.

• Test-time adaptation at inference — adapting model predictions on a per-sample or per-batch basis during deployment without any training phase; recent TTT (Test-Time Training) and TENT variants show promise for streaming distribution shift.

• Multi-modal domain adaptation — aligning distributions jointly across text, image, audio, and sensor modalities in foundation models that process multiple modalities simultaneously.

• Causal domain adaptation — exploiting causal structure to distinguish stable causal mechanisms (invariant across domains) from unstable spurious correlations (domain-specific), enabling more principled adaptation.

• Federated + privacy-preserving adaptation at scale — convergence of federated learning, differential privacy, and PEFT domain adaptation methods for regulated sectors (healthcare, finance, government).

• Continual domain adaptation — handling ongoing, streaming distribution shift without catastrophic forgetting, merging domain adaptation with continual learning methodology.

• Adaptation for safety-critical autonomous systems — domain adaptation with formal verification bounds, relevant to the UK’s Automated Vehicles Act 2024 and evolving EU AI Act requirements for high-risk AI in mobility and healthcare.

• Foundation model adaptation efficiency — as models grow to trillion parameters, the cost of even PEFT adaptation becomes significant; research into minimum-intervention adaptation (e.g., activation steering, in-context domain specification) is growing.

  Regulatory and Ethical Dimensions

  Domain adaptation intersects regulatory frameworks in ways that are increasingly legally significant, particularly in the UK and EU contexts. The EU AI Act (fully applicable from August 2026 for high-risk AI systems) requires that AI systems deployed in high-risk categories — including medical devices, critical infrastructure, and autonomous vehicles — be validated on data representative of the target deployment environment. This validation requirement directly operationalises the domain adaptation problem: if a medical AI system was trained on hospital A’s data and deployed at hospital B, the Act requires evidence that performance on hospital B’s data distribution is acceptable. Source-free domain adaptation methods that adapt a model to hospital B’s data without sharing hospital A’s patient data are both the technically appropriate solution and the legally required approach under GDPR’s data minimisation and purpose limitation principles.

  The UK AI Safety Institute (AISI, established 2023) has included distribution shift in its evaluation frameworks for frontier AI systems. AISI’s evaluations assess whether models trained on web data generalise appropriately to specific UK government deployment contexts (public service access, immigration processing, welfare benefit determination), which constitute domain adaptation challenges in the NLP sense. The AISI’s Responsible Scaling Policy review process has highlighted domain mismatch as a key source of deployment risk for large language models in government contexts.

  In regulated industries, model validation requirements create specific demands for domain adaptation. UK FCA (Financial Conduct Authority) guidance on AI and machine learning in regulated financial services (FG/21/1 and subsequent updates) requires that model risk management frameworks explicitly assess performance degradation under distribution shift, using out-of-time and out-of-sample testing that simulate the domain gap between training and live deployment environments. NHS England’s Accelerated Access to AI (AAAI) pathway for medical AI requires pre-specified domain adaptation plans for deployments across multiple NHS trusts, typically involving prospective validation on target-site data before clinical deployment. The MHRA’s AI framework for software as a medical device (SaMD) requires post-market surveillance that effectively implements ongoing domain adaptation monitoring in production.

  Data privacy considerations shape which domain adaptation methods are legally deployable in practice. Under UK GDPR and the Data Protection Act 2018, patient data cannot be transferred between NHS trusts without patient consent or a Section 251 approval. This legal constraint makes source-free domain adaptation and federated domain adaptation the only permissible approaches for cross-trust model adaptation in NHS contexts: the source hospital’s patient data stays on-premise, and only model weights (or LoRA weight deltas) are shared. The NHS federated AI platform (FedGPT project, referenced in NHS England’s 2024-2025 AI strategy) explicitly adopts this architecture, using FedLoRA and differential privacy guarantees (ε-differential privacy applied to gradient updates) to enable collaborative model improvement across trusts without patient data leaving site.

  The Intellectual Property dimensions of domain adaptation are less well-resolved. When a model is adapted to a new domain using a company’s proprietary internal data, the adapted model weights may encode proprietary information about the training data distribution. Courts in the UK and EU are beginning to address whether model weights constitute a form of derivative work from the training data, relevant to domain adaptation when the target domain data is proprietary or copyrighted. The EU AI Act’s training data transparency requirements for general-purpose AI models may create disclosure obligations when domain-adapted models are deployed in regulated contexts.

  Research & Literature

  1. Shimodaira, H. (2000). Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of Statistical Planning and Inference, 90(2), 227-244.
  2. Blitzer, J., McDonald, R., & Pereira, F. (2006). Domain adaptation with structural correspondence learning. EMNLP 2006.
  3. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., & Wortman Vaughan, J. (2010). A theory of learning from different domains. Machine Learning, 79(1), 151-175.
  4. Sugiyama, M., Suzuki, T., Nakajima, S., Kashima, H., von Bünau, P., & Kawanabe, M. (2008). Direct importance estimation for covariate shift adaptation. Annals of the Institute of Statistical Mathematics, 60(4), 699-746.
  5. Gretton, A., Borgwardt, K. M., Rasch, M. J., Schölkopf, B., & Smola, A. (2012). A kernel two-sample test. Journal of Machine Learning Research, 13, 723-773.
  6. Ganin, Y., & Lempitsky, V. (2015). Unsupervised domain adaptation by backpropagation. ICML 2015.
  7. Ganin, Y., Ustunova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., & Lempitsky, V. (2016). Domain-adversarial training of neural networks. JMLR, 17(1), 2096-2030.
  8. Long, M., Cao, Y., Wang, J., & Jordan, M. I. (2015). Learning transferable features with deep adaptation networks. ICML 2015.
  9. Long, M., Zhu, H., Wang, J., & Jordan, M. I. (2017). Deep transfer learning with joint adaptation networks. ICML 2017.
  10. Sun, B., & Saenko, K. (2016). Deep CORAL: Correlation alignment for deep domain adaptation. ECCV Workshops 2016.
  11. Courty, N., Flamary, R., Tuia, D., & Rakotomamonjy, A. (2017). Optimal transport for domain adaptation. IEEE TPAMI, 39(9), 1853-1865.
  12. Damodaran, B. B., Kellenberger, B., Flamary, R., Tuia, D., & Courty, N. (2018). DeepJDOT: Deep joint distribution optimal transport for unsupervised domain adaptation. ECCV 2018.
  13. Long, M., Cao, Z., Wang, J., & Jordan, M. I. (2018). Conditional adversarial domain adaptation. NeurIPS 2018.
  14. Saito, K., Watanabe, K., Ushiku, Y., & Harada, T. (2018). Maximum classifier discrepancy for unsupervised domain adaptation. CVPR 2018.
  15. Liang, J., Hu, D., & Feng, J. (2020). Do we really need to access the source data? Source hypothesis transfer for unsupervised domain adaptation. ICML 2020.
  16. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., & Wang, B. (2019). Moment matching for multi-source domain adaptation. ICCV 2019 [DomainNet].
  17. Koh, P. W., Sagawa, S., Marklund, H., et al. (2021). WILDS: A benchmark of in-the-wild distribution shifts. ICML 2021.
  18. Houlsby, N., Giurgiu, A., Jastrzebski, S., et al. (2019). Parameter-efficient transfer learning for NLP. ICML 2019.
  19. Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2021). LoRA: Low-rank adaptation of large language models. ICLR 2022.
  20. Gulrajani, I., & Lopez-Paz, D. (2021). In search of lost domain generalization. ICLR 2021 [DomainBed].
  21. Rethinking effectiveness of unsupervised domain adaptation methods. OpenReview, ICLR 2024. https://openreview.net/forum?id=l9ZIU71zQK
  22. Federated Low-Rank Adaptation for Foundation Models: A Survey. IJCAI 2025. https://www.ijcai.org/proceedings/2025/1196.pdf
  23. LoRA-FAIR: Federated LoRA Fine-Tuning with Aggregation and Initialization Refinement. ICCV 2025.
  24. Kherallah, M., et al. (2024). Unsupervised Domain Adaptation: Recent Advances and Future Perspectives. Springer Nature. https://link.springer.com/book/10.1007/978-981-97-1025-6
  25. Domain Adaptation for Novel Imaging Modalities with Application to Prostate MRI. UCL Discovery. https://discovery.ucl.ac.uk/id/eprint/10165221/
  26. Lester, B., Al-Rfou, R., & Constant, N. (2021). The power of scale for parameter-efficient prompt tuning. EMNLP 2021.
  27. Saenko, K., Kulis, B., Fritz, M., & Darrell, T. (2010). Adapting visual categories to new domains. ECCV 2010 [Office-31].

Provenance