Calculus for ML

Most people meet calculus as a page of limits, then a page of rules, and by the time anything is actually differentiated the point has gone missing. Then they meet machine learning, where one number — how much the loss moves when you nudge one weight — turns out to be the entire mechanism of training.

This course runs the other way. Every idea starts from a number a training loop needs, and the calculus arrives as the way to get that number.

Where this course is right now

All five lessons are live. The course is complete.

• Rate of change and the derivative — live. What a slope measures, how to compute one from a nudge, and why its sign and its size are the only two things an optimizer ever asks for.
• Derivative rules for ML — live. The handful of rules that cover almost every function a model is built from, so you stop nudging and start reading the answer off.
• The chain rule — live. What happens to a slope when one function feeds another, which is the single fact backpropagation is made of.
• Partial derivatives and the gradient — live. One knob at a time, and then all of them at once as a vector.
• The Jacobian and products of derivative matrices — live. Many inputs and many outputs, and why a backward pass is a chain of matrix products.

Start here

Rate of change and the derivative takes about twelve minutes and assumes only that you can read a function and its graph.