English | EN 
Introduction
Machine learning models often perform impressively in the lab but struggle in the real world. The main culprit? Domain shift: the difference between the data a model was trained on and the data it faces when deployed.
For example: a model trained to detect fish in underwater videos may fail when applied to aerial drone footage of the ocean; or a self-driving car trained in sunny California might not work as well in snowy Norway.
Two research directions have emerged to tackle this challenge: Domain Adaptation (DA) and Domain Generalization (DG). They share the same goal: making models robust across data distributions but take fundamentally different approaches.
🔄 What is domain adaptation?
Domain Adaptation focuses on adapting a model trained on a source domain so it performs well on a target domain, where data distributions differ. Crucially, DA assumes that we have access to data from the target domain during training, even if it’s unlabeled.
There are several flavors of DA:
- Supervised DA: target labels are available (rare in practice).
- Unsupervised DA: only target data without labels (most common).
- Semi-supervised DA: limited labels in the target domain.
How it works:
Methods typically align the feature distributions of the source and target domains. This can be done through:
- Adversarial learning (e.g., DANN by Yaroslav Ganin et al., 2015), where a domain discriminator ensures features are indistinguishable between domains.
- Statistical alignment (e.g., Correlation Alignment), matching means and covariances of features.
- Fine-tuning models with a small amount of target data.
Example: training a whale detector on underwater videos (source) and adapting it with unlabeled aerial drone images (target) so the model works for both.
🌐 What is domain generalization?
Domain Generalization addresses a harder scenario: what if we have no access to the target domain at all during training?
Here, the goal is to learn models that generalize directly to unseen domains. Instead of adapting to a known target, we make the model robust enough to handle unexpected domain shifts.
How it works:
DG strategies include:
- Data augmentation: simulating diverse conditions via style transfer, random augmentations, or synthetic data.
- Meta-Learning: training models to adapt quickly to new environments (e.g., MLDG by Da Li et al., 2018).
- Invariant feature learning: enforcing representations that remain stable across multiple training domains (e.g., IRM by Martin Arjovsky et al., 2019).
Example: Training a fish detector only on underwater images, but expecting it to work on satellite images of oceans without ever having seen them during training.
⚠️ Challenges and open questions
- Domain Adaptation: requires access to target data, which may be costly, private, or unavailable. Performance can drop if the target domain changes again.
- Domain Generalization: much harder, since unseen domains can be drastically different. Models often overfit to training distributions despite augmentations.
- For both: computationally demanding and lacking standardized benchmarks in many areas (e.g., marine ecology, rare diseases).
🔮 Future directions
The research community is actively exploring:
- Hybrid approaches: combining DA and DG by designing models that adapt if target data is available, but still generalize otherwise.
- Self-supervised learning & foundation models: pretraining on massive diverse datasets helps build invariance.
- Continual learning: allowing models to adapt over time as they encounter new domains.
Ultimately, robust AI will require models that not only adapt but also generalize seamlessly, bringing us closer to real-world deployment across diverse conditions.