An interactive lecture series on computing derivatives correctly, checking your work, and letting the computer do it for you.
Why matrix calculus is error-prone and how to check your work with finite differences. Interactive demos of the error "V" shape and optimal step sizes.
Start here →Worked examples from simple to complex. Chain rule, product rule, and a derivative workshop where you can try any function.
Explore →Minka's differential approach: work with $d\mathbf{X}$ instead of partial derivatives. Cleaner, more mechanical, and less error-prone.
Explore →The tape, forward vs reverse, and why AD gives exact derivatives. Overview and comparison of both modes.
Discover →Dual numbers, tangent propagation, seed directions, and passes. Step through traced computations interactively.
Explore →Adjoints, the backward pass, backpropagation, and a GCN example. How to differentiate millions of parameters in one pass.
Explore →