From Bayes to Deep Networks

The kiosk and the trainee. After the new year, two arrivals change campus life: a photo-recognizing bento kiosk in Cafeteria A, and a robot mascot trainee assigned to shadow Chibany. Neither can be reasoned with — and yet both clearly learn. This Part opens the black boxes: how a network sees, why its training is a likelihood in disguise, what attention actually attends to, and why a large language model doing in-context learning looks suspiciously like the hierarchical Bayes Chibany already knows.

graph LR
    V[From Bentos<br>to Vectors] --> A[From Rules<br>to Weights 🤖]
    A --> B[Neural Net<br>Fundamentals 🤖]
    B --> C[Transformers<br>& Attention 🤖]
    C --> D[LLMs & In-Context<br>Learning]
    D --> E[World Models<br>& Imagination]

Start with From Bentos to Vectors — the linear-algebra-from-zero on-ramp (vectors, dot products, matrices as transformations of space), written for readers with no math background, with two interactive widgets. Everything after it assumes it.

Part in preparation

The remaining chapters accompany Weeks 11–12 of the Human and Machine Learning course and are being written now. The stubs below sketch each chapter’s premise.

Chapters

This project is generously funded by the Japanese Probabilistic Computing Consortium Association (JPCCA).