Basics

This part builds the working vocabulary of deep learning from the ground up. It opens with the preliminaries — tensors, automatic differentiation, linear algebra, calculus, and probability — the machinery every later chapter assumes. From there it develops the core supervised models in the order the ideas themselves build on one another: linear and softmax regression as the simplest trainable predictors, the multilayer perceptron as the first genuinely deep model, and a builder’s guide that turns those ideas into reusable, debuggable, device-aware code.

The second half introduces the two architectures that dominate practical work. Convolutional networks — classic and modern — exploit the spatial structure of images, and recurrent and sequence models carry information through time. By the end of this part you can define, train, regularize, and profile the standard architectures, and you have the engineering habits (initialization, numerical stability, checkpointing, GPU memory) that the Advanced part takes for granted.