Introduction: The Challenge of Distribution Shift
Deep learning models typically rely on the independent and identically distributed (i.i.d.) assumption—the belief that training and testing data share the same underlying distribution. In functional brain imaging (e.g., EEG, fMRI), this assumption is frequently violated due to "domain shift." Variations in hardware, recording environments, and individual subject physiology lead to significant performance degradation when models are deployed in real-world clinical settings. This systematic survey explores strategies to bridge this gap through Domain Adaptation (DA) and Domain Generalization (DG).
A Taxonomy of Transfer Learning and Adaptation
The field is broadly categorized based on the relationship between source and target domains, specifically regarding marginal distributions $P(X)$ and label spaces $Y$. Understanding these relationships is crucial for selecting the appropriate transfer strategy.
Domain Adaptation (DA)
DA methods assume access to some target domain data (often unlabeled) during training. Key strategies include:
- Alignment: Techniques like Domain-Adversarial Neural Networks (DANN) use adversarial training to force the model to learn domain-invariant features.
- Feature Disentanglement: Separating task-relevant brain signals from domain-specific noise.
- Pseudo-labeling: Iteratively refining model predictions on the target domain to bootstrap learning.
Domain Generalization (DG)
DG aims for "zero-shot" transfer to unseen domains by learning universal representations that work out-of-the-box. Strategies include:
- Meta-Learning: Simulating domain shifts during training to "learn to learn" across environments.
- Data Augmentation: Synthesizing diverse domains to improve robustness to noise.
- Ensemble Learning: Leveraging multiple source models to mitigate individual site biases.
Scenarios in Domain Adaptation
Beyond simple distribution shifts, the overlap between source and target label sets defines different DA challenges. While Closed Set DA assumes identical classes, more complex scenarios like Universal DA handle unknown label relationships.
Architectures and Benchmarks
The survey highlights a transition from traditional machine learning (SVMs) to deep architectures:
- CNNs: The most widely used building block, appearing in over 50% of fMRI studies reviewed.
- RNNs/LSTMs: Frequently employed to capture temporal sequences in EEG data.
- Graph Neural Networks (GNNs): Rapidly gaining ground for their ability to model the physical layout of sensors and functional brain connectivity.
Standard benchmarks include SEED and DEAP for EEG emotion recognition, and ABIDE (Autism Brain Imaging Data Exchange) for fMRI-based diagnostic generalization.
Key Insights and Future Directions
Functional images are uniquely sensitive to non-stationary noise (e.g., subject mood, caffeine intake, or electrode displacement), making "Cross-Subject" adaptation the most significant challenge. While adversarial alignment is currently the most popular strategy, there is a clear trend toward Unsupervised Domain Adaptation to circumvent the high cost of medical data labeling.
Future research must address several "frontier" areas:
- Interpretability: Moving beyond "black box" models to provide clinical justifications for predictions.
- Privacy-Preserving DA: Utilizing Federated Learning to adapt models without exposing sensitive patient data.
- Source-Free Adaptation: Developing methods that can adapt a pre-trained model to a new site without requiring access to the original training data.
Summary updated