Machine learning
Machine learning is useful when defining fixed rules for every situation would be difficult. A model can predict values, classify information, identify similarities or generate content. To work in practice, it needs a clear goal and evaluation with data similar to real use.
Text reviewed on
← Back to the 7 pillars
When it emerged
The idea of programs improving through experience took shape before today's use of AI. In 1959, Arthur Samuel described a checkers program that adjusted its play using earlier games. This is a milestone in machine learning, not a single birth date for every method in the field.
What it is
Models adjust parameters from data or feedback to recognize patterns, predict outcomes or generate content. Development still requires human choices about data, goals and evaluation. Deep learning is a family of methods based on multilayer neural networks.
How it works
A team first defines a task and a way to measure the result. Training data then helps adjust a model, while separate data checks whether the pattern also works on new cases. In supervised learning, examples include known answers; in unsupervised learning, the system seeks groups. Neural networks are one possible family alongside trees, regression and other methods.
What it is for
Machine learning is useful when defining fixed rules for every situation would be difficult. A model can predict values, classify information, identify similarities or generate content. To work in practice, it needs a clear goal and evaluation with data similar to real use.
Where it is used
Spam filters, demand forecasts, recommendations and pattern detection in images are common uses. For a message filter, a team gathers labeled examples, trains a classifier, checks false positives and monitors changes in spam after release.
Limits and safeguards
Learning from data does not mean learning alone without design, supervision or tests. An ordinary chat does not retrain the model. A model may reproduce biases in its data or lose quality when conditions change. Training, validation and monitoring are different stages; an ordinary conversation with an assistant does not automatically change the model's parameters.