Machine learning
Machine learning trains software on data so it can find patterns and make useful predictions or generate content for new inputs.
Machine learning builds software by training it on data. The result is a model, a structure plus learned numbers called parameters that together produce predictions. One widely used textbook sets a single test for the algorithm behind it: it must be “able to learn from data.”
In supervised learning, each example contains two parts: features, the information given to the model, and a label, the answer it should predict. Training changes the model to reduce a measured error called loss. Evaluation then uses examples kept out of training to check whether the learned relationship generalises.
Once accepted, the model can receive a new input and produce a prediction. That stage is inference. Other forms of machine learning learn differently: unsupervised methods search for structure without labels, and generative models learn to produce new content. Reinforcement learning instead improves through rewards or penalties for the actions it takes in an environment.
Machine learning helps when useful relationships can be learned from examples.
Separate learning the pattern from using it.
- 1 · collectGather examples whose features contain information relevant to the result you want.
- 2 · trainA training algorithm adjusts the model to reduce the difference between its predictions and the known answers.
- 3 · evaluateRun the trained model on examples it never saw, and check its answers against the real ones.
- 4 · inferGive the accepted model a new input; producing its prediction is called inference.
Training changes the model; inference uses the trained model without repeating that training for each request.
| Who | What they ask | What it works with |
|---|---|---|
| Email service | “Is this new message spam?” | Words, sender details and labels from earlier messages |
| Energy planner | “How much electricity will this building use tomorrow?” | Past usage, weather and building features |
| Music app | “Which song might this listener enjoy next?” | Listening patterns and item information |
| Factory team | “Does this sensor pattern suggest a fault?” | Equipment readings and previous outcomes |
- It can predict a number, such as rainfall or travel time, from relevant input data.
- It can classify an input, such as deciding whether an email belongs to the spam category.
- It can discover groups in unlabeled data through unsupervised learning.
- It can generate new content when trained as a generative model.
- Machine learning does not remove the need to evaluate how well the trained model works.
- More examples do not help if the data lacks the range or quality needed for the intended setting.
- Success on training examples does not guarantee good predictions on unfamiliar examples.
- A model may need more rounds of training and testing before it is ready for real use.
Sources used
This explainer is written in original language. The links below support its factual claims.
- docsWhat is Machine Learning?, Google for Developers · read 27 Sept 2026
- docsSupervised Learning, Google for Developers · read 27 Sept 2026
- docsMachine Learning Glossary: ML Fundamentals, Google for Developers · read 27 Sept 2026
- officialMachine Learning - Glossary, National Institute of Standards and Technology · read 27 Sept 2026
- docsEndpoints for inference in production, Microsoft Learn · read 27 Sept 2026
- docsCross-validation: evaluating estimator performance, scikit-learn · read 27 Sept 2026
- paperDeep Learning, Chapter 5: Machine Learning Basics, Goodfellow, Bengio and Courville, MIT Press 2016 · read 27 Sept 2026