Thread by @svpino

Two of the most significant problems you have to deal with when building machine learning models:

• Overfitting
• Underfitting

Here is a quick mental model to help identify when one of them is happening.

↓

2. Let's start with a quick and simple definition:

Overfitting happens when your model is too complex for your dataset.

For example, a very deep neural network trying to learn a few dozen samples with a couple features.

3. Underfitting happens when your model is too simple for your dataset.

For example, a linear regression model trying to learn a large dataset with hundreds of features.

4. How can you identify whether your model is overfitting or underfitting?

You can look at the loss on both your train and test data.

Let's look at a few different patterns:

5. First, let's focus on the train set:

• If the loss is low, your model is either overfitting or is working well (you can't say.)

• But if the loss is high, your model is underfitting!

If the train loss is low, you can then look at the test loss.

6. Let's now look at the loss on the test set:

• If the loss is low, your model is working well.

• If the loss is high, your model is either overfitting or underfitting (you can't say.)

You now have everything you need to make a decision:

7. Let's summarize by looking at the train and test losses:

• Are both high? You are underfitting.

• Are both low? Your model is working well.

• Is train loss low, and test loss high? You are overfitting.

8. A quick suggestion:

Always start by overfitting your data. Make sure everything works, and only then, start simplifying and regularizing until you get a good model.

Looking for more practical machine learning content? Follow me @svpino for weekly tips and stories.