Underfitting
When a model is too simple or trained too little to capture the structure in the data — high error on both training and test sets.
In one line
The model can’t even fit the training data — both train and test error are bad.
What it actually means
Underfitting is the mirror image of overfitting. Where overfitting means the model memorized noise, underfitting means the model hasn’t learned the signal. Common causes: the model family is too weak (linear model on nonlinear data), regularization is too strong, the learning rate is too small, training stopped too early, or the features don’t contain enough information for the task. Diagnosis is easy: train loss is high and test loss is about the same.
Why it matters
Underfitting is the easy bug to miss. Everyone worries about overfitting, trains for a while, sees training and validation loss both flat and roughly equal, and calls it done. That “flat” might be underfitting, not convergence. Always check whether a bigger model or more training would keep improving — if it would, your current model is underfitting.
Example
Train RMSE: 4.5 Test RMSE: 4.6 → likely underfitting
Train RMSE: 0.2 Test RMSE: 4.5 → overfitting
Train RMSE: 1.2 Test RMSE: 1.4 → probably a good fitYou’ll hear it when
- Diagnosing a model that refuses to learn.
- Reviewing learning curves in a model report.
- Deciding whether to add capacity or regularization.
- Talking about the bias–variance tradeoff.