What is learning?
🎓 Train vs test: the honest exam
Why: Learn ML's golden rule: never grade a model on questions it memorized.
Unlocks: Building your first real model — with an honest scorecard.
A student memorizes last year's exam and scores 100%. Did they learn anything? You find out with new questions. Machine learning has exactly this problem: a model can "memorize" its training data and look perfect — then fall apart in the real world. That failure has a name: overfitting.
The defense is beautifully simple: split your data before training. Study set and exam set. The model never sees the exam until grading day.
⚡ Quick challenge
A model scores 99% on its training data and 55% on new data. What happened?
⚡ Quick challenge
Why must the test set stay hidden during training?