By Heart or Understood?
Don't just memorise the answers!
What you'll learn in this level
- What it means when an AI only learns by heart
- Why a model has to be tested on data held back from it
- Why "works on new things too" is the real goal
When a Net Learns by Heart
Sometimes a child learns all the answers by heart and still fails the test. They understood nothing. Exactly the same thing happens to an AI.
The AI is perfect on the practice questions. On new questions it is bad.
How Good Is Your Model Really?
- 1
Underfitting
The model is too simple for the pattern. That is like not practising enough.
- 2
Learned by heart
The model is too complicated. It even remembers the mistakes in the data. That is like learning answers by heart.
- 3
Really understood
The model has learned the real pattern. It works on new data too. That is real understanding.
- 4
Regularization
Tricks that keep the AI from slipping into learning by heart.
Can you spot the difference between learning and learning by heart?
Guess first, then experiment
Model A gets everything right on the practice data. On new data it gets 6 out of 10. Model B gets about 8 out of 10 on both. Which one do you pick?