Machine learning: fit a rule instead of writing every case

Machine learning (ML) builds models whose behaviour is fitted from data. Instead of explicitly writing every condition for recognizing an image or predicting a value, we choose a model family and a learning objective, then fit its parameters.

The result is still a computation. Learning changes where its behaviour comes from, not whether it follows mathematical operations.

The central test is a new example

A model can memorize its training examples and fail on new ones. The useful achievement is generalization: exploiting patterns that continue to hold outside the fitting data.

That is why data selection, held-out evaluation and the objective matter alongside the algorithm. A perfectly optimized proxy can still predict the wrong thing for the application.

Supervised learning uses target labels; self-supervised learning constructs targets from the data itself; reinforcement learning learns from interaction and rewards. These are different sources of learning signal, not separate meanings of “intelligence.” Deep learning is a subset using multilayer neural networks.