"Machine Learning" Term Coined
Arthur Samuel formally coins the term "Machine Learning", defining it as "the field of study that gives computers the ability to learn without being explicitly programmed."
Overview
In 1959, Arthur Samuel formally used the term "Machine Learning" in his research on the checkers program. His classic definition was: "Machine Learning is the field of study that gives computers the ability to learn without being explicitly programmed." This definition captured the core essence of machine learning—automatically discovering patterns from data and improving performance, rather than humans writing specific rules. Samuel's checkers program itself was an early success case of machine learning, continuously improving through self-play.
Study Notes
Background, mechanics, impact, and key takeaways in one read
Background & Context
In July 1959, Arthur Samuel published "Some Studies in Machine Learning Using the Game of Checkers" in the IBM Journal of Research and Development (vol. 3, no. 3, pp. 210–229). There he summarized two learning procedures that let a computer play better than its programmer without being explicitly programmed, and he formally coined the term "machine learning." "Artificial intelligence" had been named only three years earlier at Dartmouth, and learning was still a vague aspiration; Samuel's reproducible engineering paper turned that aspiration into a studyable, measurable subfield and laid the linguistic foundation for the next six decades, from perceptrons to deep learning.
Key Points
- 1In July 1959 Samuel's IBM journal paper first formally coined and used the term "machine learning."
- 2The paper gave the field its classic definition: enabling computers to learn without being explicitly programmed.
- 3It showed a machine could reach strong play in only 8–10 hours of self-play from the rules plus unknown weights.
- 4It summarized rote learning and generalization learning, the latter clearly foreshadowing the later temporal-difference methods.
- 5The paper was later reprinted in Feigenbaum and Feldman's 1963 classic collection "Computers and Thought."
- 6The term endures today as the common banner linking both symbolic and statistical approaches to learning.
How It Works
Samuel operationalized "learning": the program holds a "scoring polynomial" of about 38 weighted features (piece advantage, center control, etc.) whose signs and relative weights start unknown. Two mechanisms improve it. First, rote learning stores each encountered position with its minimax backed-up value; on reuse it deepens search by one ply, and a small discount per backup makes the program prefer "win sooner / lose later" paths. Second, generalization learning has two copies (Alpha, Beta) play many games and nudges weights so that successive position values stay consistent—if the gap between them is large, the weights need adjusting. In essence this updates parameters by the difference between predicted values, isomorphic to modern temporal-difference learning, except it lacks an explicit terminal-reward signal.
Impact & Legacy
Coining "machine learning" made "letting machines improve automatically from experience" a fundable, publishable, comparable mainstream direction. Samuel's checkers program was its first strong proof and previewed the reinforcement-learning paradigm: environmental feedback tunes an evaluation function. The term survived expert systems, statistical learning, and deep learning alike. The caveat: Samuel meant symbolic parameter tuning, not today's gradient-based statistical learning, yet both share the spirit of "data-driven, not hand-coded rule by rule." Decades later, the same phrase anchors billion-parameter models, showing how durable a single well-chosen term can be.
Key Takeaways
- "Learning without being explicitly programmed" is the simplest yet most powerful definition of ML.
- Self-play plus value consistency already touched the core of reinforcement learning in 1959.
- A good term can unify a field; "machine learning" outlived the specific algorithms behind it.
Glossary
Samuel's 1959 definition: giving computers the ability to learn from experience without being explicitly programmed.
Samuel's linear weighted function that scores board positions; its weights are tuned by the learning process.
Updating the current value by the difference to the next state's value; Samuel's generalization learning already hinted at it.