Unsupervised Domain Adaptation (UDA) Task

An Unsupervised Domain Adaptation (UDA) Task is a task generalization that evaluates the ability of algorithms or system solutions to transfer knowledge from a labeled source domain to an unlabeled target domain.

• AKA: No-Label Domain Adaptation.
• Context:
  • Task Input: Labeled source domain samples and unlabeled target domain samples.
  • Task Output: Predictions on the target domain test set.
  • Task Performance Measures: Accuracy, A-distance, target-domain error rate.
  • Task Objective: Maximize performance on the target domain without using target labels during training.
  • It can be systematically solved and automated by an unsupervised domain adaptation system.
  • It can test robustness to distribution shift across domains such as sensor drift, language style, or imaging modality.
  • It can range from synthetic-to-real transfer (e.g., MNIST → SVHN) to cross-lingual adaptation or medical imaging tasks.
  • ...
• Example(s):
  • Office-31 Benchmark, which evaluates image classification performance in UDA.
  • VisDA-2017 Challenge, a synthetic-to-real visual domain adaptation benchmark.
  • DANN Task, testing transfer learning via adversarial domain classifiers.
  • ...
• Counter-Example(s):
  • Supervised Domain Adaptation Task, which assumes access to some target labels.
  • Cross-Lingual Transfer Task with labeled target supervision.
  • ...
• See: Transfer Learning Task, Domain Adaptation Algorithm, Gradient Reversal Layer, Feature Alignment Technique.

---