I made a CPU only spiking neuron network lib that comes pretty close to PyTorch
5 hours ago
- #Wikipedia Classification
- #Spiking Neural Networks
- #Neuromorphic Computing
- Neuromorphic classifier called NeuronGuard, built using Spiking Neural Network and neuromorphic event engine in Rust, exposed to Python.
- Trained on 44.4 GB Wikipedia dataset to classify articles into 5 domains with 93.14% accuracy on 10,000-article test split.
- Training completed in 15.85 seconds on Apple Silicon CPU using streaming dataset pipeline with zero disk overhead.
- Model size is 320 KB, with microsecond inference latency and memory footprint under 50 MB.
- Compared to PyTorch MLP, NeuronGuard is 10.3x faster in training on CPU, with 7.6x smaller model size, but slightly lower accuracy within 5%.
- Architecture features include Hebbian-style plasticity, cache-aligned memory layout, GIL-free parallelism, and transactional patterns.
- Can be used in Python via pip install neuronguard, with example code provided for loading pre-trained weights and running inference.
- Open-source and available on GitHub and PyPI for community use.