v0.1.1: Full MNIST Benchmark Results

v0.1.1: Full MNIST Benchmark Results

This release updates the project with a full Stage 0-4 MNIST benchmark using the release-oriented configuration.

Benchmark Settings

• hidden_dim=128
• epochs=100
• full MNIST train/test batches
• seed 35
• 4 time steps for temporal stages

Environment

• Python 3.9.13
• PyTorch 1.12.0
• torchvision 0.13.0
• CUDA available
• GPU: Quadro P620
• Runtime: about 1 h 41 min

Results

Stage	Train accuracy	Test accuracy	Spike rate	Sparsity	Event ops proxy
0	99.80%	97.91%	-	-	-
1	99.24%	97.09%	-	-	-
2	99.22%	97.27%	34.83%	65.17%	17,833,800
3	99.06%	97.41%	40.00%	60.00%	20,481,800
4	99.25%	97.24%	25.79%	74.21%	13,206,980

Stage 4 keeps high MNIST accuracy while producing the sparsest event activity among the temporal stages.

Reference Weights

Stage 4 reference weights are available as release assets:

• stage4_mnist_hidden128_seed35_v0.1.1.pt
• stage4_mnist_hidden128_seed35_v0.1.1_metrics.csv
• stage4_mnist_hidden128_seed35_v0.1.1.sha256

The standalone Stage 4 checkpoint reaches 97.75% best test accuracy.

v0.1.0: Reproducible ANN-to-SNN Structural Evolution Toolkit

v0.1.0: Reproducible ANN-to-SNN Structural Evolution Toolkit

Initial reproducible release.

Features

• Five-stage ANN-to-SNN structural evolution
• MNIST dataloader
• CPU/CUDA automatic device selection
• Config-based training
• Metrics CSV output
• Evidence figure generation
• Preliminary Stage 0-4 MNIST sanity benchmark table
• Unit tests and GitHub Actions CI
• Legacy script preserved for reference

Recommended Follow-Up

• Run full Stage 0-4 benchmarks for release-quality claims
• Replace the preliminary README table with full-run accuracy, spike rate, sparsity, and event-operation proxy