Memorising vs understanding: overfitting explained
Big idea: a model must work on new examples, not just the ones it learned from.
The story in short
Arjun, the new boy from Pune, memorised five years of question papers. He scored 100 in practice, and 38 in the real unit test. Riya built a “super neuron” for the cricket match with 50 clues, including the colour of a kurta, a scooter and the dosa for breakfast. It got all 150 old Sundays right, but only 2 of 6 new ones. Bunny’s simple neuron, with 5 clues, got 146 of 150 old Sundays and 5 of 6 new ones. So they sealed 30 Sundays in an envelope as a test, and the neuron scored 28 of 30 (93%). Arjun stopped memorising and scored 79 next time.
The idea
- Overfitting: memorising training examples so closely that new ones go wrong (too wiggly).
- Underfitting: a rule too simple to capture the pattern, even in training data.
- Generalising: doing well on new examples. That is the goal: just right.
- Spurious pattern: an accident in the data that looks like a real rule (the dosa).
- Test data / held-out set: examples kept hidden until the end, like a sealed exam paper.
- Accuracy: the percentage of answers a model gets right.
Why it matters to you
When a company says its AI is “99% accurate”, ask: on what? A score on data the AI has already seen proves very little.
Remember: understand, don’t just memorise. The only fair test is one you’ve never seen.
Check yourself
Riya's 50-clue neuron used the colour of a kurta to predict cricket. What is that kind of clue called?
A company says its AI is 99% accurate on the data it trained on. What is the best question to ask?
The neuron scored 28 out of 30 on the sealed Sundays. What is that as a percentage (roughly)?
More for you: 🎒 Students · 🧑🏫 Parents and teachers