soroban

[!NOTE] The GPU lessons in this repo are developed and tested on a real RTX 4090 machine generously provided by longkt90. None of the hardware measurements would exist without it. Thank you.

Soroban (算盤) is the Japanese abacus, the tool for doing real arithmetic by hand. That is the whole idea of this repo: learn to train neural networks by computing the training runs yourself, on paper, with a four-function calculator, and only then letting the machines confirm your numbers.

I started this because every explanation of backpropagation I read either hid the arithmetic behind matrix notation or hid it behind a framework. Neither ever made it stick for me. What made it stick was sitting down and computing three steps of gradient descent by hand, then writing code whose only job was to agree with my paper. So that is the format of every lesson here.

Every lesson follows the same discipline. You predict what will happen, you compute a small training run entirely by hand, then code reproduces your arithmetic with assert statements standing guard over every number, and finally you break something on purpose and study the wreckage. If the asserts pass, the machine has agreed with your paper. If they fail, you get to find out who is wrong, and learning to find out who is wrong is the actual curriculum.

There is no required background beyond arithmetic. The one calculus idea we need (a derivative is a slope you can measure by nudging) is built from a numerical experiment inside lesson 0001, not imported from a course you were supposed to have taken. And when notation gets in the way, the maths shelf has one plain-language page per symbol and idea, each worked out on the same numbers the lessons use, so the lessons can stay focused on training while the shelf carries the reference material.

The lessons