Machine learning

Machine learning is a branch of artificial intelligence focused on creating and analyzing statistical algorithms that improve by processing data and applying what they’ve learned to new, unseen examples, allowing them to carry out tasks without explicit programming. Deep learning, a major advance in this area, has enabled neural networks—a specific type of statistical algorithm—to outperform many older machine learning methods. The field is built on statistics and mathematical optimization, and it overlaps with data mining, which emphasizes exploratory data analysis through unsupervised learning. Theoretically, probably approximately correct learning offers a mathematical and statistical framework for machine learning, with most traditional and deep learning algorithms functioning as empirical risk minimization within that framework.

The term “machine learning” was coined in 1959 by Arthur Samuel, an IBM employee known for his work in computer gaming and artificial intelligence; “self-teaching computers” was also used at the time. The first machine learning program appeared in the 1950s, when Samuel created a checkers program that calculated each side’s odds of winning. However, the field’s history stretches back further, rooted in decades of studying human cognition. In 1949, psychologist Donald Hebb published *The Organization of Behavior*, introducing a theoretical neural structure based on interactions among nerve cells. Hebbian theory—where connected artificial neurons adjust their connection strengths based on data—laid the groundwork for many machine learning algorithms. Other researchers, including Walter Pitts and Warren McCulloch, contributed by proposing the first mathematical model of neural networks, with algorithms that mimic human thought. By the early 1960s, Raytheon Company had developed Cybertron, an experimental “learning machine” using punched tape memory and rudimentary reinforcement learning to analyze sonar signals, electrocardiograms, and speech patterns. It was repetitively trained by a human operator, who could press a “goof” button to make it reconsider incorrect decisions. Nils Nilsson’s 1960s book *Learning Machines* focused on pattern classification, and interest in pattern recognition continued into the 1970s, as documented by Duda and Hart in 1973. In 1981, a report described using teaching strategies to train an artificial neural network to recognize 40 characters (letters, digits, and symbols) from a computer terminal.

Tom M. Mitchell offered a widely cited, formal definition of machine learning algorithms: “A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E.” This definition is operational rather than cognitive, echoing Alan Turing’s proposal in “Computing Machinery and Intelligence” to replace the question “Can machines think?” with whether machines can convincingly imitate a human in conversation. In 2012, AlexNet—created by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton—dramatically improved results in the ImageNet competition, boosting the adoption of deep neural networks. In 2013, Tomáš Mikolov and colleagues introduced word2vec, a technique for efficiently learning distributed vector representations of words from large text collections. In 2014, Ian Goodfellow and colleagues developed generative adversarial networks (GANs), a framework for training generative models through an adversarial process. In 2016, AlphaGo became the first program to defeat a professional human Go player on a full-sized board without handicaps, using deep neural networks and reinforcement learning. In 2017, Ashish Vaswani and colleagues introduced the Transformer, a neural network architecture based primarily on attention rather than recurrence or convolution.

Machine learning grew out of the quest for artificial intelligence. Early AI researchers explored having machines learn from data using symbolic methods and what were then called “neural networks”—mostly perceptrons and other models later recognized as reinventions of statistical generalized linear models. Probabilistic reasoning was also applied, particularly in automated medical diagnosis. But as AI shifted toward logical, knowledge-based approaches, a rift formed. Probabilistic systems faced theoretical and practical problems with data acquisition and representation. By 1980, expert systems dominated AI, and statistics fell out of favor. Work on symbolic and knowledge-based learning continued within AI, leading to inductive logic programming, but the statistical line of research moved outside AI proper, into pattern recognition and information retrieval. Neural network research was largely abandoned by AI and computer science around the same time; this subfield, called “connectionism,” was kept alive by researchers in other disciplines.