Most ML curricula split into two tribes: the statisticians who reason in distributions and the deep-learning practitioners who reason in layers and gradients. This book's bet is that the split is artificial — treat everything as inference under uncertainty, and a linear classifier, a Gaussian mixture, and a deep net all become instances of the same recipe. That single lens is what makes hundreds of methods feel connected rather than memorized.
Key Findings
- It rebuilds the standard toolkit (regression, classification, clustering, dimensionality reduction) on a shared probabilistic spine, so the relationships and tradeoffs between methods become explicit rather than incidental.
- Every figure ships with reproducible code in Python, JAX, and TensorFlow — you can rerun and modify the exact experiment behind each plot, which turns passive reading into something you can poke at.
- It deliberately stops at the introductory frontier and hands off advanced material (deep generative models, variational inference, causality) to a companion volume, keeping this one focused rather than encyclopedic.
Who It's For
A strong fit if you want one coherent narrative connecting statistics and deep learning, and you're comfortable with multivariate calculus and linear algebra — it assumes mathematical maturity and pays it back with depth. Look elsewhere if you want a quick applied cookbook or a gentle non-mathematical primer; the reasoning here is derivation-first, and the companion "Advanced Topics" volume, not this one, is where research-level methods live.
