Show HN: Neural network digit recognition on a Game Boy

2 points

a year ago

Some weeks ago I had the silly idea of doing neural network inference on the Nintendo Game Boy. I chose the MNIST handwritten digit dataset, read up on int8 quantization and got to work.

The result is a playable ROM that fits in a single ROM memory bank (16 KB) and contains both a trained NN and inference code. The NN running on the Game Boy achieves 93.86% accuracy on the MNIST test set; although not a good result for MNIST (since it is such an easy classification task), it's quite decent for a model running on a Game Boy.

You can find the compiled ROM on the Releases page (https://github.com/jongoiko/gb-nn/releases), feel free to try it out!