Active learning is a supervised learning technique in which an algorithm actively selects which samples it wants to learn from in order to improve its performance. This differs from traditional supervised learning, in which the algorithm is given a fixed dataset to learn from without any input on which samples to choose.
The goal of active learning is to minimize the amount of labeled data required to achieve a given level of performance. This is particularly useful in cases where the cost of labeling data is high, such as when expert annotation is required or when labeling must be done by humans.
Active learning can be applied to a wide range of machine learning problems such as classification, regression, and clustering. In classification, active learning algorithms attempt to find the most informative examples to be labeled to improve the accuracy of the classifier. In regression, active learning algorithms may try to learn the shape and extent of a function with the fewest samples that minimize the overall error. In clustering, active learning can be used to select which data points to annotate when creating clusters.
One popular active learning approach is uncertainty sampling where the algorithm selects samples on the basis of the confidence or uncertainty of the current model output. The idea is that the algorithm should receive more information for the ambiguous and uncertain samples. For example, in the case of image classification, the algorithm would ask the user to label samples that are difficult for it to classify, rather than samples that it is already confident about.
Another approach is query-by-committee in which the algorithm trains multiple models, and then selects samples that result in the most disagreement among those models. This approach attempts to find the most difficult samples for the classifier to label, in order to improve its robustness.
Active learning has been successfully applied to a variety of applications, ranging from natural language processing, image and audio classification, and medical diagnosis, as well as other areas where the cost of labeling is high. In some scenarios, active learning has even been shown to outperform traditional supervised learning in which all data is labeled.