1.2.1. Data, Patterns, and Generalization

💡 First Principle: A model is only as good as the patterns in its training data, and its real test is generalization — performing well on inputs it never saw during training. If it merely memorizes the training examples, it's useless on anything new.

In supervised learning, the most common setup, you provide features (the inputs, like the words in an email) and labels (the correct answers, like "spam" or "not spam"). The model adjusts itself until its predicted labels match the real ones, then you hope it generalizes to new emails. This is why data quality is paramount: gaps, errors, or bias in the training data become gaps, errors, or bias in the model's behavior. There's no magic that lets a model learn something its data never showed it.

⚠️ Exam Trap: A model performing perfectly on its training data but poorly on new data hasn't "succeeded" — it has overfit, memorizing instead of generalizing. High training accuracy alone is not evidence of a good model.

Reflection Question: If you trained a résumé-screening model only on résumés from people a company hired in the past, what pattern might it learn that you would not want it to repeat?