Models that are too simple cannot capture the underlying patterns in data, performing poorly on both training and test data.
When a model is not complex enough to capture the true patterns in the data, resulting in poor performance everywhere.
- Model has low capacity (too few parameters or overly restrictive assumptions)
- Training error remains high because model cannot fit the data
- Test error is also high (if model can’t fit training, it can’t generalize either)
- Model makes simplistic assumptions (linear when relationship is nonlinear)
- Both train and test performance are similarly poor
- Capacity-driven: fewer parameters + complex data = higher underfitting risk
- Detectable: poor training performance is the signature
- Fixable: more complex models, more features, or less regularization
- Symmetric: opposite of overfitting (train-test gap is small, but both are bad)
- Contrasts with: Overfitting — opposite problem (model too complex)
- Related: Supervised Learning — underfitting affects all supervised models
- Related: Train-Test Split — both train and test errors reveal underfitting
- Related: Data Modeling — model selection must balance under/over-fitting
- Underfitting can masquerade as overfitting if only test error is monitored
- Excessive regularization (L1/L2 penalty too high) causes underfitting
- Feature scaling issues can make models appear to underfit
- Wrong model family: linear model for highly nonlinear data