Dr N T Pisharam / Mr R N Pisharody
Librería Samer Atenea
Kálamo Books
Librería Elías (Asturias)
Librería Kolima (Madrid)
Librería Proteo (Málaga)
The machine tells this one itself. A neural network is not a brain. This book is the machine explaining what it actually is: an arrangement of numbers, a way of being corrected, and a set of choices somebody made on its behalf. Thirty chapters take it from a single bent line to the machines behind today’s language models, and every one of them is built from nothing. No framework, no automatic differentiation, no accelerator. Every backward pass is written out by hand and checked against finite differences first. What makes this book unusual is that every number in it was measured before it was printed. Each chapter has a verification script beside it that states the chapter’s claims in plain words, runs the experiment, and tests each claim in turn. If a claim fails, the script stops. The script writes its results into a file the chapter prints from, so the sentence you read and the measurement behind it cannot come apart. Wherever the experiment disagreed with the draft, the words were changed and not the experiment. That discipline produced results the usual account leaves out. Dropout made no measurable difference on the problem it was tested on. Weight decay bought a little and no more. A network built to ignore where things are was caught reading position off the edge of the picture. Those are in the book because they are what happened. Forty-two full-page plates. A toolkit chapter that builds the whole thing again in one place, and a glossary that points every term back to the chapter that earned it. The verification scripts are available to any reader who wants to run them. Written by a fifteen-year-old and his father, for anyone who would rather understand the machine than be impressed by it. Book Two of Know the Machine, following *The Mathematics Behind Machine Learning* (978-1-911866-04-6).
