This evergreen history article uses authoritative archives and official records. Exact dates are used when documented; gradual inventions and rollouts are described as periods rather than being assigned a misleading single birthday.
Quick facts
- Machine learning is the practice of improving computer performance from data or experience rather than specifying every rule manually.
- Arthur Samuel used the phrase “machine learning” in connection with checkers programs at IBM in the 1950s.
- The field draws from statistics, optimisation, computer science, information theory and cognitive research.
- Supervised, unsupervised and reinforcement learning describe different feedback settings.
- Modern foundation models are a large-scale extension of the learning-from-data tradition.
Statistical roots and pattern recognition
Long before the phrase machine learning became popular, statisticians developed regression, classification and probability methods for drawing conclusions from data. Engineers studied signal detection and pattern recognition, while cybernetics explored control and feedback.
Early computers made it possible to automate calculations but had little memory and processing power. Researchers therefore focused on small, carefully structured problems. The key conceptual shift was that a program could adjust its behaviour based on examples rather than following a fixed list of human-authored instructions.
Arthur Samuel and learning through games
At IBM in the 1950s, Arthur Samuel built checkers programs that improved by playing and evaluating board positions. He used the term machine learning to describe computers gaining skill without being explicitly programmed for every situation. Games were useful laboratories because rules and outcomes were clear.
Samuel’s work showed several enduring ideas: a model needs a representation of the problem, an objective for measuring success and a procedure for improving decisions. Modern systems are vastly larger, but those three elements remain central.
Algorithms, data and the statistical turn
Machine learning expanded through nearest-neighbour methods, decision trees, linear models, Bayesian networks, hidden Markov models and clustering. During the 1990s, support-vector machines and ensemble methods demonstrated strong performance on classification tasks.
The statistical approach often outperformed hand-built symbolic rules when sufficient data existed. Search engines, spam filters, credit scoring, speech systems and recommendations became major applications. The field learned that data quality, feature design and evaluation methodology could matter as much as the choice of algorithm.
Deep learning changes feature engineering
Traditional machine-learning projects often required experts to design input features. Deep neural networks learned representations directly from images, audio and text. GPU computing, larger datasets and improved optimisation made multilayer networks practical at scale during the late 2000s and early 2010s.
Deep learning did not replace every method. Tree-based models remain powerful for structured business data, and simpler models can be cheaper, more explainable and easier to validate. The right method depends on data, risk and operational requirements.
Foundation models and operational machine learning
Large pretrained models learn from broad datasets and can be adapted to many tasks. This shifts part of machine learning from building a separate model for each problem toward selecting, prompting, fine-tuning and evaluating general models.
Production machine learning also requires data pipelines, monitoring, privacy, security, bias testing and maintenance. A model can degrade when user behaviour or the real world changes. The history of machine learning is therefore as much about reliable systems engineering as it is about algorithms.
Common misconceptions
- Machine learning and artificial intelligence are related but not identical terms.
- A model does not understand data merely because it finds predictive patterns.
- More data is not always better when the data is biased, duplicated or irrelevant.
- Deep learning is one family of machine-learning methods, not a replacement for every other algorithm.
Timeline: key years and locations
| Year | Location | Event | Why it mattered |
|---|---|---|---|
| 1800s–early 1900s | Europe and North America | Statistical regression and probability methods develop | Created mathematical tools for learning relationships from data. |
| 1950s | IBM, United States | Arthur Samuel develops learning checkers programs | Popularised the phrase machine learning. |
| 1957–1958 | Cornell Aeronautical Laboratory, New York | Perceptron work expands | Introduced a trainable pattern-classification model. |
| 1980s | International research centres | Decision trees, neural networks and probabilistic models advance | Built practical algorithm families. |
| 1990s | United States and Europe | Support-vector machines and ensemble learning grow | Improved classification in high-dimensional data. |
| 2012 | ImageNet competition | Deep learning achieves a major vision breakthrough | Accelerated representation learning across fields. |
| 2020s | Global cloud and research infrastructure | Foundation models spread | Enabled adaptation of large pretrained systems to many tasks. |
Frequently asked questions
What is the difference between AI and machine learning?
AI is the broader goal of building systems that perform intelligent tasks; machine learning is a major approach that learns patterns from data or experience.
Who coined the term machine learning?
Arthur Samuel is widely credited with popularising it through IBM checkers research in the 1950s.
Does every machine-learning project need deep learning?
No. Simpler statistical and tree-based models may be more accurate, efficient or explainable for many structured-data tasks.