Probabilistic Deep Learning with Python, Keras and TensorFlow Probability

Probabilistic Deep Learning demystifies the statistical foundations that underpin nearly every neural network, showing practitioners that all traditional DL amounts to maximum likelihood estimation and that extending models to output full probability distributions—rather than single-point predictions—is both theoretically principled and practically achievable. The book moves from neural network architectures and gradient descent through maximum likelihood loss derivation, TensorFlow Probability, normalizing flows, and finally Bayesian neural networks via variational inference and MC dropout. Rich with Jupyter notebook exercises, real case studies (Bavarian roadkills, CIFAR-10 novel-class detection), and state-of-the-art examples (WaveNet, PixelCNN++, Glow), it equips readers to build models that not only predict accurately but also know when they don't know—an essential capability for safety-critical and decision-support applications.

What it argues

A causal-structural model describing how design choices in neural network architecture, outcome distribution selection, loss function construction, and Bayesian extension propagate through training dynamics and predictive distribution quality to produce downstream outcomes of prediction accuracy, calibration, and uncertainty quantification.