Table of Contents

🎓 Intended learning outcomes

At the end of this lesson, students are expected to:

• Be able to explain and justify the role of generalization in machine learning
• Put the various concepts covered in generalization into context with one another
• Connect learning theory to generalization in practice
• Describe and implement practical steps that can be taken to ensure good generalization
• Explain why key tenets of generalization — the fundamental theorem of PAC learning and bias-variance tradeoff have recently been called into question
• Describe domain shift and its implications for generalization

👀 A review of what we have learned ★★☆

To end our module on ML & generalization, let’s look back and recap what we’ve learned. When we developed our avocado picking app, we learned that simply finding parameters that minimize an objective function on the training data is not enough. In ML, our models need to make good predictions on unseen data — they need to generalize.

We learned that underfitting happens when the model is insufficiently expressive to model the data, and overfitting happens when it has more expressiveness than is necessary and it unintentionally captures some of the natural variance.