Definition
Underfitting is the opposite failure mode to overfitting. It occurs when a model is too simple, or trained too briefly, to capture the underlying structure of the data. An underfit model performs poorly not only on new inputs but also on the very data it was trained on, a telltale sign that distinguishes it from overfitting.
Common causes
Underfitting usually traces back to a model with too little capacity for the task, too few training passes over the data, overly aggressive constraints, or features that do not carry enough signal. A linear model asked to fit a curved relationship, for example, will systematically underfit no matter how long it trains. The model has high bias: it makes strong, simplifying assumptions that the data does not support.
Fixing it
Remedies are generally the reverse of the overfitting toolkit: increase model capacity, train for more passes, add more informative features, or relax constraints that are suppressing the model's flexibility. The goal in every case is to find the balance point where the model is complex enough to learn real patterns but not so complex that it memorizes noise. This trade-off is often called the bias-variance trade-off, and navigating it is a core skill in self-hosted model training.
For the related concepts, see our entries on overfitting, regularization, and the role of the training epoch in deciding how long to train.
In Simple Terms
Underfitting is the opposite failure mode to overfitting. It occurs when a model is too simple, or trained too briefly, to capture the underlying structure…