Active Learning

Active Learning allows the learning algorithm to actively request labels for a small number of examples among predictions, which is manually labeled by a human.

Process of active Learning:

1. Label most informative unlabeled data and train your model.
2. Use trained model on unlabeled data, then label a portion of worst predictions.
   Methods for discovering worst predictions or most informative instances use estimated probability(EP) of the predicted label as a measure of confidence or certainty:
   • Least Confidence Uncertainty: it picks predictions with least confident.
   • Smallest Margin Uncertainty: it picks predictions with most confusion in choice between available labels.
   • Entropy Reduction: It uses entropy as a measure of uncertainty and chooses most uncertain predictions.
3. repeat last step until predictions pass required accuracy..

Approaches in active learning:
- Query Synthesis: the learning algorithm is allowed to request a label for any instance from unlabeled data. Perform well in theoretical analysis, however in Generative Models can produce instances which cannot be labeled. therefore Pool-based and Stream-based methods are more popular.
- Pool-based: Learning algorithm is limited to requesting labeling from a limited pool of unlabeled data.
- Stream-based: A stream of unlabeled items are inputted to learning algorithm, each time it should decide to whether query the label, or never query the item.
- DAgger (Dataset Aggregation): It's an iterative algorithm for datasets of observation-action pairs that solves imitation learning problems caused by learning from previous predictions in training. It creates a dataset of observations, which is then corrected by an expert for good actions with newly labeled observations.