Underfitting

Definition

Underfitting happens when a model fails to learn the useful pattern.

Simple form:

bad on training data
bad on unseen data

The model is not fitting the task well enough.

Underfitting is insufficient learning of the pattern.

---

Why Underfitting Happens

Underfitting can happen for several reasons.

Examples:

model too simple
not enough training
poor features
bad optimization
wrong architecture
too much regularization

The model does not have enough useful learning capacity or training signal.

Underfitting means the model cannot capture what it needs to capture.

---

Underfitting vs Overfitting

Overfitting:

training performance high
new-data performance low

Underfitting:

training performance low
new-data performance low

Overfitting learns too much training-specific detail.

Underfitting learns too little useful structure.

Overfitting is too specific; underfitting is too weak.

---

Example

Suppose the true relationship is curved.

A simple straight-line model may not capture it.

true pattern:
nonlinear

model:
linear

The model may perform poorly even on training examples.

A model can fail because its representational form is too limited.

---

How Underfitting Is Reduced

Common responses include:

increase model capacity
train longer
improve features or representations
adjust optimization
reduce excessive regularization
use a better architecture

The goal is to give the model enough ability to learn the pattern.

Reducing underfitting means increasing useful learning ability.

---

Final Summary

• Underfitting means failing to learn the useful pattern.
• It usually performs poorly on both training and test data.
• It can come from limited capacity, poor optimization, or too much regularization.
• It is different from overfitting.
• Reducing underfitting means giving the model enough useful ability to fit the task.