What is underfitting in deep learning?

Underfitting happens when a deep-learning model is too simple, insufficiently trained, or otherwise unable to learn the important patterns in the training data.

A common sign is that the model performs poorly on both the training data and unseen validation data.

For example:

Training accuracy   = 68%
Validation accuracy = 65%

If both are low, the model may not have learned enough from the data.

Why does underfitting happen?

Common causes include:

• The model does not have enough capacity for the problem.
• Training stops too early.
• The learning rate or optimization setup prevents effective learning.
• The input features or data representation are not informative enough.
• Regularization is too strong.

How is it different from overfitting?

• Underfitting: the model performs poorly even on the training data.
• Overfitting: the model performs very well on training data but poorly on unseen data.

See “What is overfitting in deep learning?” for the complementary case.

Underfitting can often be reduced by increasing model capacity, training longer, improving optimization, or using better input representations.

In short, underfitting means the model has not learned the underlying patterns well enough, so performance remains poor even on the data it was trained on.