Neural networks are the model family behind most of modern AI, from image recognition to language models. At heart, a neural network is a large function built from layers of simple units called neurons, whose weights and biases are tuned until the network gives good answers. This module builds that picture step by step, using handwritten digit recognition as the running example.
The lessons follow 3Blue1Brown's classic series in order: first what a neural network is, then how it learns by gradient descent on a cost function, and finally backpropagation, the algorithm that efficiently works out how to nudge every weight.
Watch the lessons in order. If a video runs long, feel free to treat it as a reference you dip back into later rather than something to finish in one sitting.
3Blue1Brown introduces neurons, layers, weights, and biases using handwritten digit recognition, and shows why a network is really just a function with thousands of tunable parameters.
Video 18minHow a network learns from labeled examples: the cost function measures how wrong it is, and gradient descent nudges the weights downhill to reduce it.
Video 20minAn intuitive walkthrough of backpropagation, the algorithm that works out how every weight and bias should change, and how stochastic gradient descent speeds it up.
Video 12min