Underfitting
Learn what happens when a neural network is too simple to capture the real pattern.
Underfitting explained visually
Follow Laila as she diagnoses a model that performs poorly even on familiar training examples.
What you will learn
Diagnose underfitting, understand bias and variance, and prepare features with normalization.
When the model learns too little
Underfitting happens when a model cannot represent the useful relationship in its training data. Both training and validation performance remain weak.
Capacity and useful features
A model may need more suitable layers, neurons, training time, or better features. The goal is enough capacity to learn the signal without memorizing noise.
Regularization can go too far
Strong regularization can restrain a network so much that even important patterns disappear. Validation results help reveal whether the model needs more or less flexibility.
Find the useful middle
Underfitting is too limited and overfitting is too specialized. Effective learning sits between these extremes.
Quick check: test your understanding
What is underfitting?
A model is too limited to learn the useful training pattern.
How does it usually affect training performance?
Training performance is poor as well as validation performance.
What is the goal?
Enough capacity to learn the signal without memorizing noise.
Accessible referenceLesson 23 transcript369 words · Matches the published narration and captionsOpen transcript
In Lesson 22, we learned how data augmentation creates realistic variations that help a neural network generalize. But sometimes the problem is not memorization. Sometimes the network has not learned enough. Meet Laila, a student building her first animal-recognition system. Laila begins with one simple rule: if an animal has four legs, classify it as a dog.
Her rule recognizes some dogs, but it immediately creates problems. Cats, horses, cows, and many other animals also have four legs. Even when Laila tests the system on the original examples, it makes many mistakes. Its rule is simply not detailed enough to understand the differences between the animals.
This is similar to underfitting. A neural network underfits when it cannot capture the important patterns in its training data. Its performance is poor on the examples used for training, and its performance on new examples is usually poor as well. Now imagine Laila working with a factory camera that must separate scratched products from clean ones. A model with too few layers or neurons may notice basic brightness differences but fail to recognize the small shapes and textures created by scratches.
Providing more examples may not solve the problem if the model does not have enough capacity to represent what it sees. Laila can improve the model by increasing its complexity. She might add layers, increase the number of neurons, or select layers that are better suited to images.
Convolutional layers, for example, can learn edges, textures, and shapes more effectively than a very simple network. She can also provide more useful input features. A machine-monitoring model may improve when it receives vibration and temperature instead of temperature alone. Excessive regularization can cause underfitting too. Too much dropout removes too much information during training. An overly strong L1 or L2 penalty can prevent useful connections from developing.
Laila's goal is not to build the largest possible model. She needs a model that is complex enough to learn the real pattern without becoming so complex that it memorizes noise. Underfitting means the model is too limited. Overfitting means it is too specialized. Finding the correct balance requires understanding two closely related ideas: bias and variance. That is where we continue in Lesson 24.