Pruned Variant
Minimal subnetwork extracted from the full model via neuron ablation analysis.
Architecture
| Layer | Full | Pruned |
|---|---|---|
| Hidden 1 | 32 | 11 |
| Hidden 2 | 16 | 3 |
| Parameters | 833 | 139 |
Pruning Method
Systematic single-neuron ablation identified which neurons are critical (their removal degrades accuracy) versus redundant (their removal has no effect). Only 11 of 32 Layer-1 neurons and 3 of 16 Layer-2 neurons were critical. The minimal subnetwork retains only these neurons.
Verification
Correctness is proven in Coq for all 256 inputs, identical to the full network. See PrunedThresholdLogic.v in the GitHub repository.
Usage
from model import PrunedThresholdNetwork
model = PrunedThresholdNetwork.from_safetensors('model.safetensors')