Understanding Underfitting in Detail

In this lesson, we will take a closer look at underfitting.

Underfitting occurs when an AI model fails to learn the patterns of the training data sufficiently, resulting in poor performance on both the training data and new data.

Let's assume a child is beginning to learn about dinosaurs.

At first, when they hear the word Tyrannosaurus, they only learn the image of a large animal with big teeth that walks on two legs.

If you show this child a few pictures of dinosaurs and ask, Can you pick out the Tyrannosaurus from these?, they might misidentify it if they have learned too little about dinosaurs. For example, they might think that a small dinosaur or a four-legged dinosaur is also a Tyrannosaurus.

Since the child has not learned enough information about Tyrannosaurus, they cannot correctly distinguish dinosaurs. This state is called underfitting.

The following methods can help resolve underfitting:

Increasing the complexity of the model can help it learn data patterns more effectively. For example, using a model with more features or a more structured neural network can be beneficial.

Underfitting can be addressed by adjusting the following hyperparameters:

Learning Rate

Adjust the learning rate appropriately so the model can learn sufficiently.

If the learning rate is too high, learning can become unstable, but if it is too low, underfitting may occur.

Batch Size

Setting the batch size appropriately ensures the model can sufficiently learn the data patterns.

If the batch size is too small, learning can become unstable, but if it is too large, underfitting may occur.

Number of Epochs

Increasing the number of epochs allows the model to learn sufficiently.

Too few epochs can cause underfitting.