Active Learning

Published:

Active learning is a supervised learning workflow built for situations where unlabeled data is plentiful, but labeling is expensive. Instead of labeling everything, the model repeatedly selects specific items to be labeled by an “oracle” such as a human annotator, then retrains using the new labels. The key idea is that not all examples are equally useful at a given moment. Many systems pick data points that the model is uncertain about, or examples that appear representative of unexplored regions of the data.

Common setups include pool-based selection from a large unlabeled set and stream-based selection as data arrives. This approach can cut labeling costs, but it can also create blind spots if the sampling strategy focuses too narrowly.

Follow us on Facebook and LinkedIn to keep abreast of our latest news and articles