Backpropagation from mathematical first principles to Python implementation and autonomous driving

Most deep learning texts rely leave practitioners dependent on framework abstractions that hide what's happening or equipped with theory but no working code. Deep Learning in the Visual Domain: Backpropagation in Theory, Code and Practice for Convolutional Networks and Visual Transformers, written by two experienced AI researchers, closes that gap by deriving the s of each layer and implementing it in Python from scratch – so readers see exactly how networks learn.

Coverage progresses from image filter fundamentals and neural network building blocks through convolutional networks and visual transformers, culminating in the design and implementation of autonomous driving models that navigate vehicles through a synthetic cityscape. Dedicated chapters derive the forward and backward passes for each layer type and pair them with corresponding Python code, making backpropagation itself – not just its effects – visible at every step.

Readers will also find:

• Python code implementations of the CNN and ViT architectures, providing full transparency into weight updates and gradient flow without framework abstraction
• Step-by-step backpropagation derivations for foundational layers, including the attention mechanism
• The e2e_driver simulation environment for designing, training, and evaluating end-to-end autonomous driving models in a virtual cityscape
• An active online community and supplementary content maintained by the authors to support ongoing learning and development

Deep learning professionals, graduate students, and senior undergraduates studying neural networks will find this book uniquely suited to building architectures from first principles. By uniting mathematical exposition with transparent code and a working autonomous driving application, it delivers the depth required to design networks from the ground up.

José Solomon, PhD, is the lead architect of the deep learning framework for an analog AI accelerator chip at Sagence AI. He holds two patents for deep learning algorithms in the autonomous vehicle space, co-founded a medical imaging start-up, and previously deployed models on Graphcore's IPU. He is a Sloan Foundation and National Science Foundation fellowship recipient.

François Charette, PhD, is a retired AI/ML Senior Research Scientist formerly with the DeepDSP Group at Palo Alto Greenfield Labs and Ford Motor Company. His career spans applied deep learning research in industrial and automotive domains, contributing directly to production-level neural network architectures and optimization strategies.