What is active learning?

Active learning is an approach to data acquisition which get us better models with less labeled data. Active learning doesn’t just select the “most promising” candidates for the next batch of labeling. It also prioritizes candidates which would be most informative, i.e., which would improve our model best (which ultimately leads to finding promising candidates sooner).

Given a fixed model architecture (eg., the number of layers, the type of neurons, and the connectivity), machine learning fits the model parameters to the data. With limited data, we will have limited confidence in the precise values of the parameters. In an ideal (Bayesian) case, model fitting would not give precise values for the parameters; rather, it would give a posterior distribution on parameter-space.

Many active learning selection strategies prioritize samples whose labeling would most reduce uncertainty in the model.