Pseudo-Labeling

Pseudo-labeling operates through a self-training loop. A model is first trained on a small set of labeled source data. It then predicts labels for a large pool of unlabeled target data. Only predictions exceeding a confidence threshold are retained as pseudo-labels. The model is then retrained on a combined dataset of original labeled data and the newly pseudo-labeled data, iteratively refining its adaptation to the target domain.

• Key Steps: 1) Train on source labels, 2) Predict on target data, 3) Filter by confidence, 4) Retrain on combined set.
• Primary Goal: To leverage the structure within the unlabeled target data to reduce domain shift.

A critical component is the confidence threshold, which acts as a quality filter. Only predictions where the model's output probability exceeds this threshold (e.g., 0.9 or 0.95) are converted into pseudo-labels. This prevents confirmation bias and error accumulation, where incorrect pseudo-labels would corrupt the training process.

• High Thresholds ensure high-precision labels but may leave most data unused.
• Curriculum Learning approaches may start with a high threshold and gradually lower it to incorporate more data.
• Soft Pseudo-Labels, using the predicted probability distribution rather than a hard class, can provide richer training signals.

Pseudo-labeling is a fundamental strategy within Unsupervised Domain Adaptation (UDA), where labeled source data and unlabeled target data are available. It directly addresses the reality gap by generating target-domain-specific supervisory signals.

It is often combined with other UDA techniques:

• Domain Alignment: Used alongside methods like Domain-Adversarial Neural Networks (DANN) that learn domain-invariant features.
• Consistency Regularization: Enforcing that the model makes consistent predictions for different augmentations of the same unlabeled target sample.
• This hybrid approach provides both feature-level and label-level adaptation.

While powerful, pseudo-labeling introduces specific engineering challenges that require careful mitigation.

• Confirmation Bias: The model can reinforce its own errors. Mitigated by high confidence thresholds, label smoothing, and using an ensemble of models for more robust pseudo-label generation.
• Class Imbalance: The model may be over-confident on majority classes, leading to biased pseudo-labels. Addressed by class-balanced thresholds or re-sampling the pseudo-labeled dataset.
• Domain-Specific Overfitting: The model may overfit to noisy pseudo-labels. Countered by early stopping, strong data augmentation on target data, and temperature scaling in the softmax function to calibrate confidence.

Pseudo-labeling is distinct from but related to other semi-supervised and domain adaptation methods.

• vs. Domain-Adversarial Training: DANN aligns feature distributions; pseudo-labeling generates target-domain labels. They are complementary.
• vs. Test-Time Adaptation (TTA): TTA adapts a model during inference using target data. Pseudo-labeling is a training-time technique.
• vs. Source-Free Domain Adaptation: Pseudo-labeling is a key method within source-free DA, as it uses only the pre-trained model and target data.
• vs. Simple Data Augmentation: Augmentation creates variations of existing labeled data. Pseudo-labeling creates new labeled data from a different distribution.

Pseudo-labeling is integral to domain adaptation with synthetic data. A model trained on perfect, labeled synthetic data (source) must adapt to noisy, real-world data (target).

• Sim-to-Real Transfer: Generates pseudo-labels for real sensor data (images, LiDAR) using a model trained in simulation, enabling iterative adaptation.
• Closing the Reality Gap: By iteratively re-training on confidently labeled real data, the model's performance on the target domain improves.
• Benchmarks: Commonly evaluated on domain adaptation benchmarks like VisDA, where models trained on synthetic renderings must classify real object photographs.

Domain adaptation is a subfield of transfer learning focused on training a model on a labeled source domain so it can perform effectively on a different, related target domain with little or no labeled data. The core challenge is overcoming domain shift, where the data distributions between source and target differ.

• Key Scenarios: Supervised (labeled target data), Unsupervised (unlabeled target data), and Semi-Supervised (mix of labeled/unlabeled).
• Common Approaches: Include aligning feature distributions (via MMD or adversarial training), learning domain-invariant representations, and leveraging synthetic data to simulate the target domain.

Unsupervised Domain Adaptation is the specific scenario where a model has access to labeled source data and unlabeled target data, but no labels for the target. This is the primary setting for pseudo-labeling. The goal is to leverage the unlabeled target data's structure to adapt the model.

• Contrast with Pseudo-Labeling: Pseudo-labeling is a specific technique often used to solve the UDA problem.
• Other UDA Methods: Include adversarial methods (DANN, ADDA), discrepancy minimization (using MMD), and reconstruction-based methods.

A Domain-Adversarial Neural Network is an adversarial architecture for learning domain-invariant features. It consists of a feature extractor, a label predictor, and a domain classifier. The feature extractor is trained to produce features that both predict the label accurately and confuse the domain classifier about whether the features originated from the source or target domain. This is enabled by a Gradient Reversal Layer (GRL).

• Adversarial Objective: The feature extractor and domain classifier play a minimax game.
• Relation to Pseudo-Labeling: DANN aligns distributions at the feature level, while pseudo-labeling creates a supervised signal on the target data; they can be combined.

Test-Time Adaptation is a domain adaptation strategy where a pre-trained model is adapted using only a stream of unlabeled data from the target domain at inference time. The model updates its parameters (often via self-training or entropy minimization) on-the-fly to better suit the immediate test distribution.

• Key Difference from Standard UDA: TTA does not require an upfront adaptation phase with a large batch of target data; it adapts continuously during deployment.
• Pseudo-Labeling in TTA: Online pseudo-labeling on incoming test samples is a common TTA technique, though it risks error accumulation if not carefully regularized.

Domain generalization is a more challenging paradigm than domain adaptation. Models are trained on data from multiple, diverse source domains with the goal of performing well on previously unseen target domains, without any access to target data during training. The objective is to learn fundamental, domain-invariant features that capture the core task.

• Contrast with Adaptation: Domain adaptation uses target data (often unlabeled) during training; generalization explicitly forbids it.
• Techniques: Include Invariant Risk Minimization (IRM), meta-learning, and data augmentation/randomization to simulate unseen domains.

Sim-to-real transfer is a critical application domain for adaptation techniques, particularly in robotics and autonomous systems. It involves training a model in a high-fidelity synthetic or simulated environment (the source domain) and deploying it in the physical world (the target domain). The reality gap—discrepancies in visuals, physics, and sensor noise—causes significant domain shift.

• Key Techniques: Domain randomization (varying simulation parameters extensively) and domain adaptation (using real-world unlabeled data) are primary strategies to bridge this gap.
• Pseudo-Labeling Role: Can be used to iteratively fine-tune a sim-trained model on unlabeled real-world sensor data.