noun
Underfitting occurs when a machine learning model is too basic to capture the underlying trends in training data, leading to poor performance on both training and real-world data.
It typically happens when the model lacks complexity (e.g., too few parameters or overly rigid assumptions) or when training is insufficient. For example, a linear regression model trying to fit a complex, non-linear dataset will underfit.
For practitioners, underfitting means wasted resources—time and compute power spent on a model that fails to generalize. It forces developers to either increase model complexity, gather more data, or refine feature engineering. A real-world case: a chatbot trained on a simple keyword-matching model (underfit) might fail to understand nuanced queries, costing user trust.
A language model trained on just 10,000 sentences struggles to generate coherent responses, missing subtle linguistic patterns.
In other languages
See also: Overfitting