Underfitting: Definition and Examples

Underfitting is a fundamental problem in machine learning that occurs when a model fails to properly learn the relationships and structures present in the training data. Unlike overfitting, where the model memorizes the data, an underfitting model simply does not have the capacity or resources to understand the underlying patterns.

This phenomenon can have several causes: an overly simple model (e.g., using linear regression for non-linear data), an insufficient number of parameters, too short training, or excessive regularization that overly constrains the model. The result is a model that produces imprecise and generic predictions, unable to distinguish nuances in the data.

In prompt engineering, the concept of underfitting translates into how you formulate your instructions. An overly vague or generic prompt can be seen as a form of underfitting: it does not provide enough context or constraints for the language model to produce a precise and tailored response. Just like an undertrained model, an underspecified prompt generates superficial results that do not meet expectations.

To detect underfitting, one typically observes high error on both the training set and the test set. The solution involves increasing model complexity, enriching features, extending training, or reducing regularization. In prompt engineering, this translates to enriching your prompts with examples, context, and more detailed instructions.