GB MNIST

MNIST CNN on Game Boy

A handwritten-digit neural network running natively on original Game Boy hardware. A real quantized CNN — convolution, ReLU, fully-connected layer — all in integer math on a 4.19 MHz CPU, packed into a 32 KB cartridge. No GBDK: custom boot code and hand-tuned sm83 assembly. Watch it classify digits, draw your own, and see the inference timer count the milliseconds. Runs in your browser, but it's a real ROM tested on a GBA SP.

Controls:

-Start: go to menu

-A: select mode, draw pixel or predict 

-B: remove pixel

-Select: Predict in draw mode

Credits

• Original idea: Lucas Mohimont & Lilian Hollard
• Developed by Claude
• Web emulation: GameBoy.js by Grant Galitz
• Local emulation: mGBA with RomStation
• Hardware test: GBA SP with EZ-Flash Jr
• Title screen: ChatGPT with Game Boy Image Converter