Using deep learning and a large labelled dataset, we are able to obtain excellent performance on a range of important tasks. Often, however, we only have access to a large unlabelled dataset. For example, it is easy to acquire lots of stock photos, but labelling these images is time-consuming and expensive. This excludes many applications from benefiting from recent advances in deep learning.

In Active Learning we only ask experts to label the most informative data points instead of labelling the whole dataset upfront. The model is then retrained using these newly acquired data points and all previously labelled data points. This process is repeated until we are happy with the accuracy of our model.

To perform Active Learning, we need to define some measure of informativeness, which is often done in the form of an acquisition function. This measure is called an “acquisition function” because the score it computes determines which data points we want to acquire. We send unlabelled data points which maximise the acquisition function to an expert and ask for labels.