Real models have many settings, not just one. A partial derivative measures how the error changes when you nudge a single setting, and collecting all of them into one vector gives the gradient, which points in the direction where the error rises fastest. Going the opposite way is how a model improves.
The lessons start with partial derivatives and gradients, then introduce the chain rule, the tool for finding derivatives of functions built from other functions. That is exactly how a neural network is built, which makes the chain rule the heart of backpropagation. They finish with how frameworks like PyTorch apply it automatically through automatic differentiation.
Watch the lessons in order, then continue on to Optimization.
An explanation of partial derivatives and the gradient, the starting point for how models with many settings learn.
Video 5minLinks partial derivatives to gradient descent, the method that drives learning in machine learning models.
Video 34minExplains the chain rule and how it underlies backpropagation in deep learning.
Video 5minShows how backpropagation in neural networks works using basic calculus.
Video 24minShows how PyTorch builds a computation graph and uses torch.autograd to run backpropagation automatically.
Video 10min