Iterative Pseudo-Labeling for Unsupervised Domain Adaptation

647 concepts499 papers1 proposal

When deploying machine learning models to the real world, they frequently encounter a "domain gap"—a shift in data distribution between the training environment (source domain) and the deployment environment (target domain). For instance, a vision model trained on daytime images may struggle at night, or a natural language processing model trained on simulated dialogues may fail on human-to-human conversations. As surveyed by Mei Wang et al. in Deep Visual Domain Adaptation: A Survey and Alan Ramponi et al. in Neural Unsupervised Domain Adaptation in NLP - A Survey, Unsupervised Domain Adaptation (UDA) seeks to bridge this gap using only labeled source data and unlabeled target data.

Among the most intuitive and effective strategies for UDA is iterative pseudo-labeling (also known as self-training). Instead of relying on complex adversarial distribution alignment, this approach turns the model into its own teacher. By leveraging its existing knowledge to label the target domain, the model bootstraps its own adaptation process.

The Core Intuition: Leveraging Baseline Recognition

Why should a model trust its own predictions on a domain it has never seen? The fundamental insight is that a domain shift rarely destroys a model's performance entirely. As noted by Paolo Russo et al. in From Source to Target and Back: Symmetric Bi-Directional Adaptive GAN, a model typically retains a baseline level of recognition accuracy across the domain gap. It will still be highly confident and correct about the "easiest" or most domain-invariant examples in the target dataset.

By isolating these high-confidence predictions and treating them as ground-truth "pseudo-labels," we can construct a synthetic training set for the target domain. When the model is retrained on these pseudo-labels, it is forced to update its internal parameters to accommodate the target domain's specific feature distributions.

This parameter shift can be directly observed. In Predicting Out-of-Distribution Error with the Projection Norm, Yaodong Yu et al. demonstrate that fine-tuning a model on its own pseudo-labels causes its weights to move in a precise direction (the Projection Norm) that correlates strongly with the model's true out-of-distribution performance. This confirms that pseudo-labeling provides a genuine, structured adaptation signal rather than merely reinforcing random noise.

The Self-Training Loop

In practice, iterative pseudo-labeling operates as a cyclical optimization process:

  1. Label Generation: The model, initially trained on the source domain, runs inference on the unlabeled target domain.
  2. Confidence Filtering: To prevent the model from learning from its own mistakes, a selection mechanism filters out low-confidence predictions. For example, Jitin Krishnan et al. in Attention Realignment and Pseudo-Labelling for Interpretable Cross-Lingual Classification of Crisis Tweets use a strict probability threshold to ensure only the most certain predictions are retained.
  3. Model Retraining: The filtered pseudo-labels are combined with the source data to retrain the model. As explored by Woojae Chang et al. in Domain-Specific Batch Normalization for Unsupervised Domain Adaptation, this step can be treated as a multi-task learning problem where domain-specific batch normalization layers capture the unique noise and statistics of each domain while sharing core semantic features.

Each iteration of this loop refines the model's decision boundaries. As the model becomes more familiar with the target domain's style, its accuracy improves, allowing it to generate higher-quality pseudo-labels in the next iteration. This bootstrapping mechanism is highly effective across diverse tasks, from adapting dialogue act taggers (Shachi Paul et al. in Towards Universal Dialogue Act Tagging for Task-Oriented Dialogues) to evaluating large-scale benchmarks (Shiori Sagawa et al. in Extending the WILDS Benchmark for Unsupervised Adaptation).

Breaking the Bias: Uncertainty and Ensembles

The primary vulnerability of self-training is confirmation bias—the risk that a model will generate a high-confidence but incorrect pseudo-label, and subsequently reinforce its own error during retraining. To mitigate this, modern approaches move beyond simple logit thresholds.

One strategy is to incorporate explicit uncertainty estimation. In Generative Self-training for Cross-domain Unsupervised Tagged-to-Cine MRI Synthesis, Xiaofeng Liu et al. utilize Bayesian uncertainty masks alongside pseudo-labels. By retraining the network only on pixels where the model is both highly confident and certain, they prevent erroneous synthesis artifacts from corrupting the adaptation process.

Another robust defense against confirmation bias is the use of multi-perspective consensus. When multiple independent views or models agree on a prediction, the likelihood of that prediction being correct increases significantly:

Progressive Adaptation and Curriculums

When the domain gap is exceptionally wide, attempting to pseudo-label the target domain immediately can lead to catastrophic failure because the initial baseline accuracy is too low. To overcome this, researchers structure the adaptation progressively.

One approach is a self-paced curriculum. In Cross-Domain Adaptation for Animal Pose Estimation, Jinkun Cao et al. implement a strategy that begins with a highly conservative confidence threshold. The model initially trains only on the easiest target samples, and as its competence grows, the threshold is gradually relaxed to incorporate increasingly complex anatomical structures.

Alternatively, the curriculum can be dictated by natural environmental transitions. In Dark Model Adaptation: Semantic Image Segmentation from Daytime to Nighttime, Luc Van Gool et al. tackle the extreme shift from day to night by using "twilight" as an intermediate bridge. The model first generates pseudo-labels for the twilight domain, adapts to it, and then uses that intermediate model to pseudo-label the pitch-black nighttime domain. By breaking a large distribution shift into a sequence of smaller, manageable steps, progressive pseudo-labeling prevents the model from diverging.

Source-Free and Latent Space Alignment

Ultimately, retraining on pseudo-labels does more than just adjust the final classification layer; it forces the model to reorganize its latent space. As shown by Zhengming Ding et al. in Deep Low-Rank Coding for Transfer Learning, iteratively updating pseudo-labels helps align class-specific feature distributions across domains, directly minimizing the geometric mismatch between how categories are represented.

This capability is particularly vital for source-free domain adaptation, where privacy constraints or data sizes prevent access to the original source dataset during adaptation. In Learning to Cluster Under Domain Shift, Elisa Ricci demonstrates that treating high-confidence target predictions as fixed pseudo-labels allows a model to cluster target data and refine its internal representations entirely on its own. Similarly, in speech processing, Alfonso Ortega in Unsupervised adaptation of PLDA models for broadcast diarization shows that using initial clustering outputs as pseudo-labels allows generative speaker verification models to specialize their parameters to new acoustic environments without needing any manual annotations.

By treating its own confident predictions as a steady anchor, a model can successfully navigate unfamiliar data distributions, turning unsupervised adaptation into a structured process of self-discovery.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_iterative_pseudo_labeling_iterative_pseu_5897cd2c,
  title = {iterative pseudo-labeling (iterative pseudo-label refinement)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/iterative-pseudo-labeling-iterative-pseudo-label-refinement/rdr_5897cd2c39ca49ccb920a8251fe5838c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_iterative_pseudo_labeling_iterative_pseu_5897cd2c_v1,
  title = {iterative pseudo-labeling (iterative pseudo-label refinement)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/iterative-pseudo-labeling-iterative-pseudo-label-refinement/rdr_5897cd2c39ca49ccb920a8251fe5838c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake