This is the first book focused entirely on deep learning theory. Tools from theoretical physics are borrowed and adapted to explain, from first principles, how realistic deep neural networks work, benefiting practitioners looking to build better AI models and theorists looking for a unifying framework for understanding intelligence.