Handbook of Bayesian Deep Learning
Bayesian deep learning (BDL) brings Bayesian reasoning to deep neural networks with the aim of making predictions useful for rational decision-making under uncertainty. From this perspective, posterior distributions over model parameters and posterior predictive distributions are not ends in themselves, but means towards approximating Bayes-optimal decisions. Deep neural networks provide powerful and flexible models for this purpose, while Bayesian inference offers a principled framework for representing evidence and uncertainty and for assessing the decisions that predictions inform.
This work develops a broad view of BDL, combining a principal route through Bayesian methods with foundational and mathematical aspects, questions of scalability to contemporary models and datasets, applications, and topical developments. Rather than attempting a comprehensive account, this work presents a snapshot of a broad and rapidly developing field at a particular moment in time. The aim is to place BDL within a wider perspective on machine learning: not as a specialised corner of the field, but as a framework for building powerful models while remaining formal about evidence, uncertainty, and the decisions that give predictions their purpose.