Active learning
Motivation
If a learning algorithm can choose data it wants to learn from, then it can perform better.
Definition
Select more instances based on the data have collected so far.
Workflow
- Gether data: Representative of true distribution.
- Split into seed and unlabed datasets.
- Train the model with seed. Variables: Cross-validation, model.
- Choose unlabeled instances. Variables:
- Scnarios:
- Membership query synthetis.
- Stream-based selective sampling.
- Pool-based sampling.
- Query strategies:
- Scnarios:
- Stopping criteria.
NLP, ASR, IR.
References
https://www.datacamp.com/community/tutorials/active-learning