Main Concept
Domain Adaptation is a specific type of Transfer Learning
- It is used in scenarios where task is the same, but the data distribution has shifted between the source and target datasets.
- The OBJECTIVE of transferring on model from one domain to other
- Domain adaptation can be achieved thought fine-tuning, where the model is retrained on labeled or unlabeled data from the target domain to align its feature representation.
Example
For example, a model trained to detect cars in sunny weather (input), might need to be adapted to detect cars in rainy conditions (target domain)
Type of Domain Adaptation:
A. Unsupervised Domain Adaptation
- No labels in the target domain
- Use: Continued Pre-training
- Example: Adapting from general news → medical news
B. Supervised Domain Adaptation
- With labels in the target domain
- Use: Fine-tuning with labels
- Example: General sentiment classification → sentiment in medical reviews