When one line will not do,
you stack layers
/ml lesson 11 ended with try a linear model before you reach for deep learning. These twenty lessons are what comes next.
Start from four points no line can separate, write backpropagation by hand (and check it against numerical differentiation), build attention in six lines, and see why AlphaFold needed that architecture. No PyTorch โ you only understand the inside once you have written it.
How you'll learn
The chain rule is all it takes. Twenty lines, then checked against numerical differentiation โ matching to nine decimals.
We try 220,000 straight lines and confirm that none separates XOR. The reason for stacking layers comes from there.
Why attention suits proteins, and how AlphaFold reached 92.4 at CASP14.
The last lesson asks when not to. On tabular data with few samples, logistic regression wins.