3 ms·
Show HN: MNIST64 – 99.5% accuracy MNIST BNN on a stock Commodore 64
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://jarnoh.github.io/mnist64/paper.pdf
https://github.com/jarnoh/mnist64 https://github.com/jarnoh/mnist64
https://youtu.be/Z_S0IZenW_E https://youtu.be/Z_S0IZenW_E
I'm interested to hear your feedback and hoping for someone to beat these numbers!
- nurpax 1mo agoA fun quote from the paper: ”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?”