Applied AI for Generation and Control
Diffusion, search and optimization in science, software, robotics and media
-
- $8.99
-
- $8.99
Publisher Description
Diffusion models draw the molecule, a language model proposes the program, a policy chooses the torque. This book shows how to build, evaluate and trust such systems, with derivations you can check and code you can run. Applied AI for Generation and Control, Volume 2, starts with the machinery: diffusion and flow models derived from stochastic differential equations, symmetry and equivariance, Bayesian optimization and active learning, reinforcement learning with verifiers and search, and the statistics of evaluating generative models.
The machinery is then applied to four fields. In science, you generate candidate molecules with an equivariant diffusion model, score them with calibrated surrogates, and choose which experiments to run under a budget. In software engineering, you treat a code model as a sampler filtered by tests, with an unbiased pass@k estimator, cost models for agent loops, and a guarantee on what passing tests does and does not tell you. In robotics, you connect diffusion policies to classical control, prove why behaviour cloning compounds error, and add a barrier-function safety filter. In media, you build conditional image and audio generators, a statistical watermark with a detection test, and a memorization audit.
Every chapter follows the same plan: the problem and its stakes, a precise formalization, the method with its derivation and proofs, a complete implementation in Python, a worked example with printed output, an evaluation with confidence intervals, the ways it fails, and exercises. Models are small, written in NumPy, and the examples are designed to run offline on a CPU.
Written for engineers, data scientists, quantitative researchers and graduate students who know calculus, linear algebra, probability and basic Python. Chapter 0 recaps Volume 1, so this volume can be read on its own.
Generative AI was used under human direction to produce manuscript text and companion code. The cover illustration was drawn by a script.