Teach With This Book
This book is free, open, and built to be taught from. If you run a course in probability, probabilistic machine learning, computational cognitive science, or probabilistic programming, you are welcome to adopt any part of it — whole, or a Part at a time. This page is the orientation for instructors.
What it is
A single narrative textbook that carries a reader from counting all the way to LLMs and alignment, using one continuous story (Chibany, the Chiba Tech mascot, and a year of bento mysteries) and one continuous toolkit (a generative model, a prior, an inference). Every chapter is scenario first, math second, runnable code third.
Three things distinguish it from a lecture-note PDF:
- Runnable, validated code in every chapter. Code cells are executed in CI and checked against their printed output, so what you see on the page is what runs. The probabilistic-programming spine is GenJAX (a JAX-based PPL); a dedicated chapter explains why GenJAX rather than Stan or PyMC, honestly, including where Stan/PyMC are the better tool.
- ~25 interactive in-browser widgets — conditioning by crossing out outcomes, an MCMC explorer, a Q-learning gridworld, a POMDP belief updater, a matrix-transform / ReLU-fold visualizer, an attention lookup, and more — driven live in lecture or explored by students.
- A Colab capstone per chapter, modeled on real assignments and papers (e.g. the in-context-learning chapter’s capstone runs both sides of the “is ICL Bayesian?” debate — the Xie/Ye exact-Bayes toy and the Falck martingale test on it).
It also has an in-page English ⇄ Japanese toggle (press L) — the concept
chapters are translated, which some students genuinely need.
Two reading tracks
The same book serves two audiences; a sidebar switch dims off-track chapters and each chapter’s footer points to the next stop on your track.
- 🤖 Beginner — probability theory → probabilistic ML → probabilistic computing, through to neural networks, LLMs, and ML ethics, from a light background. Roadmap.
- 🧠 CogSci — Bayesian cognitive science: probabilistic models as theories of minds (generalization, memory search, sampling hypotheses, goal inference, theory of mind). Roadmap.
The nine Parts
| Part | Covers | Capstones anchor on |
|---|---|---|
| I. Foundations | events, counting, conditioning, Bayes’ rule | counting the outcome space |
| II. The Tools (GenJAX) | simulate, trace, condition, infer in a PPL | the taxicab problem in code |
| III. Continuous & Bayesian Learning | densities, Gaussians, conjugacy, mixtures | Gaussian conjugate updating |
| IV. Structure | generalization, Bayes nets, causality, information, hierarchies | Tenenbaum & Griffiths, causal inference |
| V. Chains, Walks & Sampling | Markov chains, random walks, MC, particle filters, MCMC | Sanborn & Griffiths MCMC-with-people |
| VI. Decisions & RL | decision theory, MDPs, Q-learning, inverse RL, POMDPs | Baker goal inference, Daw two-step |
| VII. Model Complexity | bias–variance, Bayesian nonparametrics, GPs | double descent, DPMM |
| VIII. Deep Networks | vectors→nets→transformers→LLMs, in-context learning as hierarchical Bayes | FGSM, attention, the ICL debate |
| IX. Ethics, Fairness & Safety | adversarial examples, fairness formalisms, bias, alignment | the fairness impossibility, RLHF reward hacking |
Part I assumes no math background and Part VIII opens with a linear-algebra-from-zero chapter, so the beginner track is genuinely self-contained.
Adopting it in a course
The book grew alongside a 12-week graduate course, Human and Machine Learning (Chiba Institute of Technology, School of Design & Science), so it maps cleanly onto a semester — roughly one Part every one to two weeks. The course homepage has the full schedule, slide decks, readings, and assignments — a worked example of teaching from this book. Where a chapter parallels a lecture, its closing “From the lecture” box links the week’s deck, and the in-chapter self-check quizzes are the same commit-before-reveal polls used live.
Mix and match freely: a one-term probabilistic-ML course can run Parts I–III then VII–IX; a cognitive-modeling seminar can run the CogSci track; a single guest lecture can lift one chapter and its widget.
Coming from probmods?
If you have taught with Probabilistic Models of Cognition, much of the Bayesian-cognitive-science material here covers the same ground — generalization, sampling as a process model, structured inference — but with runnable GenJAX (rather than WebPPL), interactive widgets, and a from-zero on-ramp. The CogSci track is the closest map.
Feedback and reuse
The book is CC-BY-4.0 — reuse,
remix, and adapt with attribution; a CITATION.cff
is included. If you adopt it, find an error, or want a topic covered, please
reach out to Joe — feedback from other instructors is exactly what the next
revision is for.
This project is generously funded by the Japanese Probabilistic Computing Consortium Association (JPCCA).