Lesson 2 of 6 9 min · 99 words
The bias–variance trade-off in practice
Learning curves and what to change for each problem.
Once you know which problem you have, the remedies differ:
- High bias (underfitting): more features, a more flexible model, less regularisation (smaller λ), train longer, deeper/wider networks.
- High variance (overfitting): more data, data augmentation, fewer features, stronger regularisation (larger λ), dropout, early stopping, simpler model, ensembling (bagging).
A common mistake is collecting more data for a model that underfits — it won't help. Another is adding complexity to a model that already overfits.
Formula to remember
Ridge closed form
λI keeps the matrix invertible.
Quick check · +5 XP