Overview
A project-driven companion to Mathematical Awakening. Where the first book builds the math, this one uses it: 25 worked machine-learning projects across three domains — healthcare and medical AI (10), bioinformatics and genomic AI (8), and computer vision and robotics (7). Each project is a self-contained read with a problem framing, a pointer back to the underlying math, an architecture choice, an implementation sketch, and a short list of failure modes to watch.
The book assumes you can read Python and recognize the standard ML stack (NumPy, Pandas, PyTorch / TensorFlow, scikit-learn). It does not re-teach calculus, linear algebra, or probability; the math is invoked where it does work, with pointers back to the relevant chapters of Mathematical Awakening for refreshers.
Scope — what's here, what isn't
The original outline was broader (five domains, fifty projects). Three domains and twenty-five projects were actually written:
- Chapter 1 — Healthcare & Medical AI (10 projects): imaging with ViT, predictive diagnosis, clinical note summarization, vital-sign anomaly detection, treatment recommendation, radiology report generation, outbreak prediction, medical segmentation, drug-drug interaction.
- Chapter 2 — Bioinformatics & Genomic AI (8 projects): gene-expression classification, protein folding, CRISPR efficiency, disease-risk modeling, single-cell RNA-seq, pathway prediction, drug discovery, variant classification.
- Chapter 3 — Computer Vision & Robotics (7 projects): RL for robotic control, vision-based grasping, autonomous navigation, human-robot interaction, object detection and tracking, facial emotion recognition, image captioning.
NLP and Finance / Environmental / Autonomous-Systems domains were planned and not written. This edition doesn't pretend they exist.
Vintage note
These are 2024 architectures. Several sections are already aging:
- The NLP-flavored projects (clinical note summarization, radiology report generation, healthcare chatbot, image captioning) describe pre-LLM patterns — custom transformer stacks and BART / T5 fine-tuning. Today you'd reach for a foundation model + retrieval or light fine-tuning first.
- The drug-discovery and protein-folding projects train AlphaFold-inspired and chem-language architectures from scratch; in practice the modern path is to build on top of ESM / AlphaFold2/3 / DiffDock checkpoints.
- Vision Transformers for medical imaging and robotic manipulation are still correct building blocks, but the surrounding stack (distillation, LoRA, quantization, multimodal adapters) has moved forward.
What does still hold up, and why the book is on the shelf: the mapping from problem class to architecture family is stable, the loss and metric choices are stable, and the failure modes called out in each project are the kind of thing that doesn't age with the framework version. Read it as a structured walk through the classical ML stack applied to real domains, circa 2024 — not as current-frontier guidance.
What's reproducible, what isn't
- Code listings are static. The Quarto source uses
execute: eval: false, so every cell renders as a readable listing rather than a freshly executed run. That keeps the book buildable in seconds and honest about what it is: a structured pattern catalogue, not a reproducible experiment harness. - Datasets are referenced, not vendored. Projects pull from public sources (Kaggle, Hugging Face, MIMIC, public genomic databases, OpenStreetMap). Every project cites where the data comes from. Reproducing a project is a fork-and-run exercise on your own hardware.
- What you can rely on: the architectures, the loss / metric choices, the math-to-code mapping, the failure modes.
How to read
- Skim: read this preview + the preface, then open one chapter matching a domain you care about. Each project's first four subsections (Problem Statement, Mathematical Foundation, Implementation walkthrough, Extensions) give you the shape without the full listing.
- Study: pick one project and finish it end-to-end on real data before moving on. Substitute your own dataset, change one architectural choice, and watch how the metrics move. That's where the value is.
- Reference: search by problem class (classification, retrieval, segmentation, sequence modeling, control). Implementation patterns repeat across domains.
Read online vs. PDF
The web reader (below) splits the book into four chapters — preface, healthcare, bioinformatics, vision/robotics — and renders the code listings and inline math directly in the browser. The PDF is the same content in a single downloadable file, ~700 pages, 1.7 MB.