When deploying machine learning models in the wild, we routinely encounter two distinct challenges simultaneously: domain shift (where the visual style, lighting, or background of the data changes) and category shift (where the set of labels in the deployment environment does not perfectly match the training set).
Traditional Unsupervised Domain Adaptation (UDA) operates under a strict "closed-set" assumption. As surveyed by Mei Wang and Weihong Deng in Deep Visual Domain Adaptation: A Survey, standard methods assume the source and target domains share an identical label space. When these methods encounter an "open-set" scenario—where the target domain contains novel, unseen categories—they fail catastrophically.
Understanding why they fail, and how modern open-set and universal domain adaptation methods fix this, requires looking at the geometry of feature alignment.
The Failure Mode of Closed-Set Alignment
In standard domain adaptation, a model minimizes the distance between the source and target feature distributions using metrics like Maximum Mean Discrepancy (MMD) or adversarial domain classifiers.
Imagine a source domain trained to classify cars and trucks. If the target domain contains cars, trucks, and bicycles, a closed-set alignment mechanism does not know bicycles exist. It views the target domain as a single monolithic distribution and forces the entire target feature space to overlap with the source feature space.
Standard Alignment (Failure):
Source Space: [ Cars ] [ Trucks ]
▲ ▲
│ (Forced Global Alignment)
Target Space: [ Cars ] [ Bicycles ] [ Trucks ]
This forces the novel target instances (bicycles) into the clusters of the known source classes (cars or trucks). This phenomenon, known as negative transfer, warps the decision boundaries of the shared classes, degrading the model's accuracy on categories it actually saw during training. As noted by Debjeet Majumdar, forcing this mapping in open-set environments severely limits a model's real-world utility.
The Dual Objective of Open-Set Adaptation
To resolve this, open-set domain adaptation splits the problem into a dual objective:
- Align the shared classes across the source and target domains to overcome the domain shift.
- Isolate and reject the unknown classes in the target domain to prevent negative transfer.
As argued by Bo Fu et al. in Learning to Detect Open Classes for Universal Domain Adaptation, isolating the unknown samples is the linchpin of the entire process. Once the model successfully flags and separates the target instances belonging to novel categories, aligning the remaining data reduces to a standard, manageable closed-set adaptation problem.
Open-Set Alignment (Success):
Source Space: [ Cars ] [ Trucks ]
▲ ▲
│ (Selective Alignment) │
Target Space: [ Cars ] [ Bicycles ] [ Trucks ]
│
▼
[ Rejected as Unknown ]
Discovering the Unknown Without Labels
How can a model identify a completely novel category without any target labels? Researchers leverage the intrinsic structure and uncertainty profiles of the target data.
Calibrated Uncertainty
Known and unknown target samples behave differently when passed through a source-trained classifier. Shared classes typically yield higher confidence scores or cluster tightly around source centroids. Unknown classes often fall into the low-confidence valleys between decision boundaries or exhibit high entropy across different data augmentations.
To exploit this, Bo Fu et al. developed a framework called Calibrated Multiple Uncertainties (CMU). By combining metrics for entropy, confidence, and prediction consistency across an ensemble of classifiers, the system constructs a robust filter that separates shared target categories from private, novel ones.
Exploiting the Geometry of the Unknown
Rather than treating unknown samples merely as noise to be discarded, recent strategies treat them as an informative structural boundary. In Beyond the Known: Enhancing Open Set Domain Adaptation with Unknown Exploration, Lucas Fernando Alvarenga e Silva et al. demonstrate that the geometric distribution of unknown samples contains implicit information about where the known categories end. Actively exploring and clustering these unknown instances helps the model draw sharper, more accurate boundaries around the known classes.
Dynamic Prototype Expansion
In dynamic scenarios like test-time adaptation, where data arrives in a stream, models must flag novel classes on the fly. Yushu Li et al. address this in On the Robustness of Open-World Test-Time Training: Self-Training with Dynamic Prototype Expansion. Their approach maintains feature "prototypes" (centroids) for known classes. When a stream of target data arrives, the model uses adaptive pruning to identify outliers that do not fit any existing prototype, dynamically expanding its internal representation to isolate these out-of-distribution intruders before they can corrupt the model's core parameters.
Universal and Source-Free Adaptability
The ultimate realization of this research direction is Universal Domain Adaptation (UniDA). In a universal setting, the practitioner does not know beforehand whether the target domain contains more classes (open-set), fewer classes (partial-set), or a completely disjoint set of categories relative to the source.
Jogendra Nath Kundu explores this in Universal Source-Free Domain Adaptation, introducing a constraint where the model must adapt without even having access to the original source data (often due to privacy or proprietary restrictions). The author demonstrates that by preparing the source model with a "foresighted" multi-head architecture during initial training, the model can later evaluate target data autonomously. It infers the relationship between the source and target label spaces purely by analyzing feature-space anomalies, allowing it to successfully categorize knowns and reject unknowns without any prior structural knowledge or source-side re-calibration.