Underfitting happens when a model learns a process that is too simple to capture the pattern. It then performs poorly not only on unseen data, but often on its training data too. section 1.27
A useful contrast
A model that memorizes particular training sentences or images suffers from overfitting. A model that learns an overly simple rule fails to handle those training examples effectively and also struggles with new sentences or images. Neither model has learned a useful general process. section 1.27
The practical distinction is simple: generalization means working beyond the examples seen during training, while underfitting means the learned process is not capable enough to work reliably even on the data used to teach it. section 1.27
