Show HN: MNIST64 – 99.5% accuracy MNIST BNN on a stock Commodore 64
https://github.com/jarnoh/mnist64
I built an 8-bit binarized convolutional neural network that recognizes handwritten digits on C64 with state-of-the-art accuracy. I also designed a cascade model to get interactive response, and implemented everything with self-modifying cycle optimized 6502 assembly - resulting 0.653s average inference time.Link to the paper, the full 6502 source code, and a video demo.
https://jarnoh.github.io/mnist64/paper.pdf
https://github.com/jarnoh/mnist64
I'm interested to hear your feedback and hoping for someone to beat these numbers!
”We also challenge you to improve these results. With modern methods, we have now exceeded LeCun et al.’s 1998 accuracy (MNIST images 9,930 correct, ours 9,949), on a computer from 1982, so perhaps we can again move the needle a few years forward?”