Definition
Machine learning is a family of methods that adjusts a model from data or experience so its performance improves against a defined objective. Instead of writing every decision rule directly, developers choose the representation, model family, objective, training procedure, and data from which the model's parameters are learned.
Supervised learning fits examples paired with labels. Unsupervised and self-supervised methods learn structure or prediction tasks from unlabeled records. Reinforcement learning adjusts behavior from rewards or other feedback produced through interaction. These categories describe where the learning signal comes from; one system can combine them.
Where the name came from
Arthur Samuel used "machine learning" in the title of his 1959 IBM paper on a checkers program that improved from experience. He is widely credited with introducing the term to computing. The paper is strong evidence for documented use in 1959, though a title cannot prove that nobody used the phrase earlier.
Samuel's work used rote learning and adjustments to a scoring function. Modern machine learning covers far more methods, but the operational idea remains recognizable: experience changes future performance.
Distinguish it from nearby terms
- Artificial intelligence also includes symbolic reasoning, search, planning, and systems whose knowledge is encoded directly.
- Deep learning is the subset based on multilayer neural networks.
- Statistics and machine learning share many methods. Machine learning usually emphasizes predictive or decision performance in computational systems, while statistical work may center more on estimation and inference about a data-generating process.
- An adaptive rule does not become machine learning merely because a programmer changes it after seeing results. The learning procedure itself must use data or experience to adjust the model.
Operational significance
The objective and data determine what improves. A model can optimize its training metric while failing the business decision, subgroup, or future distribution that matters. Record the learning signal, splits, evaluation protocol, versioned data, and conditions under which the model should be retrained or retired.
Check your understanding
A credit policy contains hand-written income thresholds. That is explicit programming. If a training procedure estimates the thresholds or other parameters from past outcomes, that component is machine learning.