Machine Learning

Debugging

Fix ML problems faster with a structured checklist.

Overfitting vs Underfitting

  • Overfit: train high, test low → simplify model, regularize, get more data.
  • Underfit: both low → better features, better model, less regularization.

Sanity Checks

  • Shuffle labels: performance should drop near baseline.
  • Train on tiny subset: model should overfit it (if not, bug in pipeline).
  • Verify preprocessing is not fitted on test.