Generative Deep Learning

Generative Deep Learning by David Foster is a comprehensive, code-first introduction to the field of generative modeling using deep neural networks. Starting from first principles—what it means to model a probability distribution and why naive approaches fail at scale—the book builds systematically through variational autoencoders, generative adversarial networks, recurrent neural networks with attention, and world models. Each major architecture is introduced through an allegorical story, explained mathematically, and then implemented in Keras with fully worked Python code. The second half of the book applies these foundations to concrete creative tasks: painting in an artist's style with CycleGAN and neural style transfer, generating coherent text with LSTMs and encoder-decoder networks, composing polyphonic music with MuseGAN, and training a reinforcement learning agent to drive a car by dreaming inside its own generative world model. The final chapter surveys the cutting edge—Transformers, BERT, GPT-2, MuseNet, ProGAN, SAGAN, BigGAN, and StyleGAN—and speculates on how generative modeling may ultimately be central to artificial general intelligence.

What it argues

A causal-structural model linking architectural design choices and training conditions in generative deep learning systems to intermediate representational and behavioral states, and ultimately to output quality, diversity, and downstream task performance outcomes.