#weight-decay

Question 1: Single choice

You train a linear model with weight decay by SGD on house data but skip standardizing the numeric features: lot area is in the tens of thousands of square feet, while the number of bathrooms is between 1 and 4. What goes wrong?

The squared loss stops being convex in the weights, so SGD gets stuck in local minima1
The one-hot encoded categorical columns no longer work together with the numeric ones2
Large-scale features dominate the gradients, and the penalty treats features unevenly3
A linear model can no longer represent the same predictions as with standardized features4

Timed test