etoxin/neuronguard-wikipedia-classifier · Hugging Face

Log In<br>Sign Up

Neuromorphic Wikipedia Domain Classifier

This repository hosts the pre-trained vocabulary and synaptic weights for NeuronGuard , a proof of concept that uses cache-aligned Spiking Neural Network (SNN) and neuromorphic event engine written in Rust and exposed to Python.

Note: This was a proof of concept to see if a event driven like neuron inspired neural model could work.

The model is trained on the 44.4 GB 'wikimedia/structured-wikipedia' dataset (10.4M articles total) to classify articles into 5 high-level domains:

0: Science & Technology

1: Geography & Places

2: Biography & People

3: History & Events

4: Arts & Culture

Using a streaming dataset pipeline (zero disk overhead) and community-curated Infobox template names for labeling, the model was trained on exactly 1,000,000 valid articles (samples) and achieved 93.14% accuracy on a 10,000-article test split, training in 15.85 seconds using local pre-filtered data on a standard Apple Silicon CPU.

Performance Metrics (Standard Apple Silicon CPU)

Training Set Size : 1,000,000 valid articles (samples)

Model Size : 320 KB (synaptic weights and vocabulary file)

Training Time : 15.85 seconds (using local pre-filtered data)

Inference Latency : Microseconds (~5–15 µs per article)

Memory Footprint : Accuracy : 93.14% (evaluated on a 10,000-article test split)

Comparison: NeuronGuard vs. PyTorch (1,000,000 Samples)

To validate the efficiency of NeuronGuard's neuromorphic architecture, we ran a head-to-head comparison against a standard PyTorch Multi-Layer Perceptron (MLP) on the full Wikipedia dataset (1,000,000 training samples, 10,000 test samples, and a 10,000-word vocabulary).

Head-to-Head Results (Apple Silicon M-Series CPU/GPU)

Metric<br>NeuronGuard (SNN)<br>PyTorch (MLP, 1 Epoch)<br>PyTorch (MLP, 10 Epochs)<br>Trade-Off / Insight

Hardware Used<br>Single CPU Core<br>Apple Silicon MPS GPU<br>Apple Silicon MPS GPU<br>NeuronGuard achieves line-rate speed without needing GPU acceleration.

Training Time<br>15.85 seconds<br>16.56 seconds<br>163.84 seconds<br>NeuronGuard is 10.3x faster than PyTorch (10 epochs) on a single CPU core.

Data Loading Overhead<br>0.00 seconds (on-the-fly)<br>16.17 seconds<br>16.17 seconds<br>NeuronGuard trains directly on the stream, bypassing memory loading overhead.

Total Pipeline Time<br>15.85 seconds<br>32.73 seconds<br>180.01 seconds<br>NeuronGuard is 2x to 11.3x faster overall from cold start to fully trained.

Model Size<br>~320 KB<br>2.44 MB<br>2.44 MB<br>NeuronGuard's model size is 7.6x smaller , making it ideal for edge devices.

Overall Accuracy<br>93.14%<br>97.98%<br>98.15%<br>PyTorch's global backpropagation yields slightly higher peak accuracy, but NeuronGuard is within 5.0%.

Because NeuronGuard trains on the fly via Hebbian plasticity, it completely bypasses the massive dataset loading and memory overhead required by traditional deep learning frameworks.

Architectural Trade-Offs

NeuronGuard (Neuromorphic SNN) :<br>Pros : Instant training (single-pass online learning), ultra-low memory footprint (Cons : Slightly lower peak accuracy due to the lack of iterative global optimization (backpropagation).

PyTorch (Traditional Deep Learning) :<br>Pros : High peak accuracy (98%+) due to multi-pass gradient descent and non-linear optimization.

Cons : Requires dedicated GPU acceleration for fast training, larger model files, and significantly higher memory and dependency overhead.

Technology Overview

NeuronGuard operates on a hardware-conscious, matrix-free neuromorphic design. Rather than relying on traditional deep learning architectures (like Transformers or Feedforward networks), it implements the following core technologies:

Spiking Neural Network (SNN) Core : Models neural computation using discrete event spikes rather than continuous floating-point activations. Stimuli are processed as temporal events that propagate through synaptic pathways.

Hebbian-Style Plasticity : Synaptic weights are updated on the fly using simple transactional increments and decrements based on co-activation, completely bypassing backpropagation and gradient storage.

Cache-Aligned Memory Layout : All neural structures are spatially aligned to exactly 64-byte boundaries (matching standard CPU cache lines). This maximizes L1/L2 cache hit rates, prevents cache thrashing, and eliminates false sharing during parallel execution.

GIL-Free Parallelism : Drops the Python Global Interpreter Lock (GIL) during stream processing to execute concurrent, lock-free evaluations across background worker threads.

Guard/Lease Transactional Pattern : Implements transactional, lock-free leases on specific memory addresses using atomic compare-and-swap (CAS) operations for safe concurrent weight mutations.

Flat, Pointerless Serialization : Synaptic weights are stored as flat, contiguous binary arrays, enabling sub-millisecond serialization and deserialization directly to and from disk.

How to Load and Use in Python

To load and run inference using these...