Underfitting

Overview

When people hear about machine learning problems, they often hear about models that learn too much. However, models can also learn too little.

This problem is known as underfitting.

Imagine a teacher asks a student to predict future exam scores but only allows them to look at a student’s favorite color. While the information may be interesting, it is unlikely to help make accurate predictions. Important factors such as study habits, attendance, and previous grades are being ignored.

AI models can make similar mistakes.

If a model is too simple, uses the wrong features, or is not trained long enough, it may fail to recognize meaningful patterns within the data. Instead of learning useful relationships, it produces overly simplistic predictions that perform poorly.

A helpful way to think about underfitting is that the model has not learned enough from its experience.

It has seen the data, but it has not understood what matters.

As a result, the model struggles during training and continues to struggle when presented with new information.

Why It Matters

Underfitting prevents AI systems from making reliable predictions because the model never develops a strong understanding of the problem it is trying to solve.

Real-World Example

A sales forecasting model predicts nearly identical sales figures every month despite major seasonal changes in customer demand.

Related Concepts

• Model Training
• Model Evaluation
• Overfitting
• Machine Learning
• Bias-Variance Tradeoff