The primary objective when building a deep learning model isn't just performance on the training data; it's ensuring the model performs well on new, unseen examples. This capability is called generalization. Models that fail to generalize well often suffer from either overfitting (learning the training data too specifically, including noise) or underfitting (being too simple to capture the data's patterns).

In this chapter, we establish the foundation for understanding and improving generalization. We will define generalization, overfitting, and underfitting. You'll learn about the bias-variance tradeoff in the context of deep learning and how to use learning curves as a diagnostic tool. We will also introduce the roles of regularization and optimization in tackling generalization issues and guide you through setting up the required software environment for subsequent practical work. By the end, you'll practice identifying overfitting visually.