Machine learning (ML) uses algorithms to learn models from data rather than having people specify every association or parameter by hand: ML + data = model. A model may take the form of rules, a decision tree, or neural-network weights. It is an approximation, not a complete description of the world. section 1.1
A model can combine with partial data to infer a missing value: model + partial data = inference. Learning from data lets a machine search many possible settings quickly, but a human-built model may still be better when domain knowledge matters or data is scarce. section 1.1
Classifying an animal
Leg count, hair, and size can help a model infer whether an animal is a dog or a cat. “Four legs and hair” alone is too broad; useful distinctions require more information. section 1.1
Treat an ML application as a delivery project, not just a model-building task:
- Check what data exists, how it is processed, and what users need from the resulting inference. section 1.1 section 1.5
- Plan for model evaluation, production integration, and ongoing management alongside functional and nonfunctional requirements. section 1.5 section 1.6
- Assess stakeholder needs and potential harms throughout the project, including whether a proposed use suits the model. section 1.2 section 1.6
For a bike-store chain with separate sales and inventory systems, combining data in a cloud database is only a starting point. The team still needs to establish what ML could usefully do, how people would use the result, and whether the benefit justifies building and operating a model. section 1.5
