I made a CPU only spiking neuron network lib that comes pretty close to PyTorch (huggingface.co)

A new library called NeuronGuard has been created, implementing a spiking neural network (SNN) framework that closely rivals the functionality of PyTorch, but with a focus on CPU-only execution. This proof-of-concept leverages a cache-aligned architecture implemented in Rust, allowing it to process data directly from streams without traditional disk overhead. The model was trained on 1,000,000 articles from a substantial 44.4 GB Wikipedia dataset, achieving an impressive 93.14% accuracy in just 15.85 seconds, demonstrating both speed and efficiency. The significance of NeuronGuard lies in its ability to operate with minimal overhead, making it suitable for edge devices. By employing techniques like Hebbian-style plasticity for on-the-fly learning without backpropagation, it not only reduces memory usage to less than 50 MB but also aligns well with neuromorphic computing principles. While it slightly lags in accuracy compared to conventional deep learning models, the framework showcases potential for real-time applications where quick processing and low resource consumption are essential, setting a new benchmark for CPU-based machine learning in the AI/ML community.