The decomposition
Generalization error decomposes into bias squared (how far the average prediction is from truth), variance (how much predictions wobble across training sets), and irreducible noise. Models with too little capacity have high bias. Models with too much capacity have high variance. The sweet spot minimizes the sum.
How tools change the tradeoff
- Regularization raises bias to lower variance.
- More data lowers variance without changing bias.
- Bagging (random forests) averages high-variance trees to reduce variance.
- Boosting reduces bias by sequentially correcting errors; can overfit if unchecked.
Why it matters
You cannot eliminate both at once. Knowing which one dominates your current model tells you which knob to turn next. High train accuracy and low validation accuracy = variance problem. Both low = bias problem. The framework saves weeks of guessing.