Overfitting, Ridge and Lasso: When the Best Fit Predicts Worst

Overfitting, Ridge and Lasso regression and cross-validation on real car data. A more flexible model fits the data better and better, and predicts worse and worse. Why? Lesson 3 of 10 in "Machine Learning in Under 2 Hours": a complete course built from real problems. We make the line bend with new features, watch it overfit, hold it down with Ridge and the Lasso, and choose how flexible it should be by cross-validation, keeping the test set for the very end. ▶ All lessons in order:    • Machine Learning in Under 2 Hours   ◀ Builds on Lesson 2:    • Linear Regression From Scratch: Why Adding...   ▶ Next, Lesson 4:    • Logistic Regression: Why Linear Regression...   📄 Slides and lab for this lesson: https://ml.wxing.me/course/lesson-03/ Chapters 0:00 3.1 Problem: the line keeps missing 1:12 3.2 Model: make the line bend 2:45 3.3 Model: put the columns on one scale 3:34 3.4 Test: more flexible is not better 4:36 3.5 Loss: Ridge 6:02 3.6 Loss: Lasso 6:43 3.7 Test: choose with cross-validation 8:21 3.8 From code back to the car Created by Dr Wei Xing. Narration: synthetic (text-to-speech) voice. Theme music: AI-generated. Animations made in Python (Manim, matplotlib). #Overfitting #Regularization #MachineLearning