Definition
In machine learning, particularly in transfer learning, a domain is defined by two components: the feature space (the set of all possible inputs) and the marginal probability distribution of those inputs. For example, images taken in daylight and images taken at night constitute different domains due to their distinct distributions, even if they share the same feature space (pixels). Understanding domains is critical for addressing domain shift, where a model trained on one domain performs poorly on another, necessitating techniques like domain adaptation to bridge the gap between source and target distributions.
Summary
A domain represents a specific context or distribution of data characterized by its feature space and underlying probability distribution.
Key Concepts
- Feature Space
- Marginal Distribution
- Domain Shift
- Source vs Target
Use Cases
- Adapting models trained on synthetic data to real-world scenarios
- Handling cross-cultural or cross-language NLP tasks
- Medical imaging analysis across different hospital equipment