Show HN: MNIST64 – 99.5% accuracy MNIST BNN on a stock Commodore 64 (github.com)

A new project showcased on HN demonstrates a highly efficient 8-bit binary convolutional neural network (CNN) for handwritten digit recognition, achieving 99.42% accuracy on the combined MNIST+EMNIST dataset using a stock Commodore 64. This model operates within severe hardware constraints—64 kB RAM and a 0.985 MHz 6502 CPU without any AI accelerators or multiplication instructions. Leveraging binarized neural networks (BNNs) that replace traditional multiply operations with XNOR and popcount computations, the researchers managed to optimize their implementation to accommodate the old architecture while achieving performance close to state-of-the-art standards. Significantly, the model incorporates a novel confidence-gated early exit strategy that allows 91.99% of images to be classified faster (0.544 seconds) on a simplified inference path, with a fallback to a more comprehensive evaluation for tougher cases. By employing tactics such as hand-optimized assembly programming and adaptive test-time augmentation, the system maintains impressive accuracy and speed, cutting average inference time by around 60%. This work not only pushes the boundaries of performance on vintage hardware but also opens new dialogues around extreme quantization techniques in modern AI applications, showcasing that high efficiency can be achieved even in less powerful environments.