active learning
In plain wordsA way to train a model with fewer labeled examples by letting it choose which unlabeled examples a person should label next, usually the ones it is most unsure about.
A learning setting in which the learner helps decide what gets labeled, nominating the examples it expects to learn most from, so that a target accuracy is reached with far fewer labels than random annotation would need. In the common pool-based form, a model trained on a small labeled set scores a large pool of unlabeled examples, the most informative are sent to an annotator (the 'oracle'), and the model is retrained, repeating until the labeling budget is spent. Selection rules range from estimating which labels would most change the model or cut its error, through ensemble disagreement (query by committee) and batch diversity, to the simplest and commonest, uncertainty sampling of the least confident or highest-entropy predictions. Active learning suits domains where labels are expensive, such as medical images, legal documents and rare events. Its pitfalls: queried samples cluster near the decision boundary and so misrepresent the data distribution, and gains over random sampling can vanish when strong pretrained models are used. Weak supervisionweak supervision, n.: Training AI with labels that are cheap but imperfect, such as rough rules of thumb, partial tags or noisy guesses, instead of carefully hand-checked answers. and model-assisted labeling are complementary ways to cut data labelingdata labeling, n.: The work of attaching correct answers to examples — often done by large paid workforces, sometimes by experts — so that AI systems can learn from them or be tested against them. costs.
Active learning got the defect classifier to 95% recall with 1,800 labeled images instead of the 10,000 we had budgeted.
Origin
Grew out of statistical optimal experimental design and Dana Angluin's 1988 learning-theory work on queries. H. Sebastian Seung, Manfred Opper and Haim Sompolinsky introduced query by committee in 1992; David Cohn, Les Atlas and Richard Ladner's 'Improving Generalization with Active Learning' (1994) helped name the field, and David Lewis and William Gale proposed uncertainty sampling the same year. Burr Settles's 2009 survey became the standard reference.
Related
- data labelingdata labeling, n.: The work of attaching correct answers to examples — often done by large paid workforces, sometimes by experts — so that AI systems can learn from them or be tested against them.
- annotationannotation, n.: Adding information to raw data by hand or with software — marking objects in photos, tagging names in sentences, transcribing speech or ranking chatbot answers — so machines can learn from it.
- weak supervisionweak supervision, n.: Training AI with labels that are cheap but imperfect, such as rough rules of thumb, partial tags or noisy guesses, instead of carefully hand-checked answers.
- semi-supervised learningsemi-supervised learning, n.: Machine learning that combines a small set of labeled examples with a much larger pile of unlabeled ones, which are cheaper to collect, to learn better than from the labeled data alone.
- online learningonline learning, n.: Learning from data that arrive one piece at a time, updating the model after each new example instead of training once on a fixed collection, as a spam filter does while emails stream in.
- human-in-the-loophuman-in-the-loop, adj.: Describes AI systems in which a person checks, corrects or approves the machine's work at key points, rather than letting it act entirely on its own.
- multi-armed banditmulti-armed bandit, n.: A puzzle named after a gambler choosing between slot machines: you must keep picking among options whose payoffs are unknown, learning which is best while wasting as few tries as possible on the others.
Sources
- Settles, B. (2009). Active Learning Literature Survey. Computer Sciences Technical Report 1648, University of Wisconsin–Madison.
- Cohn, D., Atlas, L. & Ladner, R. (1994). Improving Generalization with Active Learning. Machine Learning 15(2).
- Lewis, D. D. & Gale, W. A. (1994). A Sequential Algorithm for Training Text Classifiers. Proc. SIGIR 1994.
Mentioned in
curriculum learningcurriculum learning, n.: Teaching a model the way schools teach people: starting with easy examples and gradually moving to harder ones, instead of presenting everything in random order., data labelingdata labeling, n.: The work of attaching correct answers to examples — often done by large paid workforces, sometimes by experts — so that AI systems can learn from them or be tested against them., data-centric AIdata-centric AI, n.: An approach to building AI that focuses on improving the data — fixing labels, adding missing cases, cleaning inconsistencies — rather than endlessly tweaking the model., entropyentropy, n.: A measure of unpredictability. A fair coin toss has more entropy than a rigged one; an AI that is unsure of its next word has high entropy., human-in-the-loophuman-in-the-loop, adj.: Describes AI systems in which a person checks, corrects or approves the machine's work at key points, rather than letting it act entirely on its own., information gaininformation gain, n.: How much a question reduces uncertainty about the answer. A decision tree asks first the question that removes the most uncertainty, measured in bits., online learningonline learning, n.: Learning from data that arrive one piece at a time, updating the model after each new example instead of training once on a fixed collection, as a spam filter does while emails stream in., semi-supervised learningsemi-supervised learning, n.: Machine learning that combines a small set of labeled examples with a much larger pile of unlabeled ones, which are cheaper to collect, to learn better than from the labeled data alone., data flywheeldata flywheel, n.: A cycle in which a product's users generate data that improves its AI, the better AI attracts more users, and those users generate still more data., decision theorydecision theory, n.: The study of how to make the best choice when outcomes are uncertain, by weighing how likely each result is against how much it matters., weak supervisionweak supervision, n.: Training AI with labels that are cheap but imperfect, such as rough rules of thumb, partial tags or noisy guesses, instead of carefully hand-checked answers.