Machine learning is how most modern AI gets built. Instead of a programmer spelling out every rule, a learning algorithm studies examples and adjusts a model until the model’s outputs match what the examples show. The finished model is then used on inputs it has never seen.
Arthur Samuel coined the term in 1959, describing a checkers program that learned to play better than the person who wrote it. Tom Mitchell later gave the textbook framing: a program learns when its performance on a task, judged by some measure, improves with experience.
The three main kinds
- Supervised learning: learn from labeled examples (this email is spam, that one isn’t), then predict labels for new inputs.
- Unsupervised learning: find structure in unlabeled data, such as groups of similar customers.
- Reinforcement learning: learn by trial and error from rewards for good actions.
A fourth, self-supervised learning, creates its own labels from raw data, for example by predicting the next word. It’s how large language models are pre-trained.
Where deep learning fits
AI is the broad goal of machines doing things that seem intelligent. Machine learning is the dominant way of getting there. Deep learning is machine learning with many-layered neural networks, and it powers today’s language, vision and speech models.
The catch
A model only knows what its training data showed it. It can fail quietly on inputs unlike anything it saw, and it can learn the biases in its data as readily as the patterns. That’s why models are judged on data held back from training, a test set, rather than on the examples they learned from.
