The author explores compressing neural networks to run on the Padauk 8-bit microcontrollers, specifically the PMS150C, which has very limited memory capacity compared to traditional microcontrollers. Despite downsizing the MNIST dataset to 8×8 pixels and making other adjustments, the author was able to achieve over 90% test accuracy even on models as small as half a kilobyte. By flattening the inference code and optimizing variable usage, the author successfully implemented an MNIST inference engine on the Padauk microcontroller. This hack proves that machine learning and edge inference can be applied even at the lowest end of microcontrollers, although the practical applications at this level remain uncertain.
https://cpldcpu.wordpress.com/2024/05/02/machine-learning-mnist-inference-on-the-3-cent-microcontroller/