MEPX
Chapter 4 of 10All chapters
  1. 1Framing the problem
  2. 2Data
  3. 3Training and testing
  4. 4Underfitting and overfitting
  5. 5Measuring performance
  6. 6Common algorithms
  7. 7How training works
  8. 8The pipeline
  9. 9After deployment
  10. 10Common pitfalls

Chapter 4 of 10

Underfitting and overfitting

Two ways to be wrong.

The balance

An underfitted model is too simple and does poorly everywhere. An overfitted one memorises the training data, scoring well there and badly on anything new.

  • A large gap between training and test scores means overfitting.
  • More data, simpler models and regularisation all pull against it.

Complexity

Complexity should be justified by measured improvement on held-out data. Starting simple gives a baseline that surprisingly often survives.