universal domain adaptation

**Universal Domain Adaptation (UniDA)** is a domain adaptation setting where the source and target domains may have different label sets—with categories that are private to the source, private to the target, or shared between both—and the algorithm must automatically identify which categories are shared and adapt only for those while rejecting unknown target samples. UniDA is the most general and realistic domain adaptation scenario, requiring no prior knowledge about the label set relationship. **Why Universal Domain Adaptation Matters in AI/ML:** Universal domain adaptation addresses the **unrealistic assumptions of standard DA**, which presumes identical label sets across domains; in real-world deployment, target domains often contain novel categories absent from training (open-set) or lack some source categories (partial), making UniDA essential for robust model deployment. • **Category discovery** — UniDA models must automatically determine which classes are shared between source and target without explicit specification; this is typically achieved through clustering target features and measuring their similarity to source class prototypes or through entropy-based thresholding • **Sample-level transferability** — Each target sample is assigned a transferability weight indicating whether it belongs to a shared class (high weight, should be adapted) or a private/unknown class (low weight, should be rejected); these weights gate the domain alignment process • **OVANet (One-vs-All Network)** — Trains one-vs-all classifiers for each source class, using the maximum activation to determine if a target sample belongs to any known class; samples with low maximum activation are classified as unknown • **DANCE (Domain Adaptative Neighborhood Clustering)** — Uses neighborhood clustering in feature space to identify shared categories: target samples that cluster near source class centroids are considered shared, while isolated target clusters are treated as private target categories • **Evaluation protocol** — UniDA methods are evaluated on H-score: the harmonic mean of accuracy on shared classes and accuracy on identifying unknown/private samples, balancing both recognition and rejection performance | DA Setting | Source Labels | Target Labels | Relationship | Challenge | |-----------|--------------|---------------|-------------|-----------| | Closed-Set DA | {1,...,K} | {1,...,K} | Identical | Distribution shift only | | Partial DA | {1,...,K} | {1,...,K'}, K'