This book presents deep learning as both a mathematical discipline and an applied field,
connecting theoretical principles with the modeling and design choices that shape modern
neural systems. It explores how neural networks are formulated and how they learn, from
multilayer perceptrons, activation and loss functions, optimization, and backpropagation
to questions of expressivity and generalization. These ideas are applied to neural
architectures for different forms of data, including convolutional, recurrent, and graph
neural networks, with attention to the mathematical principles, training challenges, and
design choices that distinguish them. The book then turns to generative models and
sequential decision-making, covering variational and adversarial models, diffusion, and
reinforcement learning.
The later parts follow the evolution of neural networks toward increasingly general systems,
exploring attention and transformer architectures, multimodal models, large language
models, retrieval-augmented systems, and autonomous agents. The book also considers
directions beyond conventional deep learning, including continuous-time and physics-informed neural networks, biologically-inspired and neuromorphic models, Bayesian
approaches to uncertainty, and quantum neural networks.
The term "theoretical" in the title, Neural Networks: A Concise Theoretical Foundation, refers to the mathematical formulation of neural networks, not to a comprehensive, proof-oriented treatment of neural network theory. Organized in four parts, the book connects mathematical descriptions with
the architectures, learning methods, and emerging directions that continue to shape neural computation.