cgn-mnist

CGN (Confluence Gate Network) — MNIST

A new standard architecture built on Gate Neurons, trained without backpropagation.

What is CGN?

The Confluence Gate Network (CGN) is a network architecture built on a single primitive: the Gate Neuron (GN).

h_j = max(0, Σ x_i · W_ij − θ_j)

Multiple signals converge, sum, and fire above threshold — the same operation as a biological neuron. No filter size, no stride, no pooling, no weight sharing. Zero architectural hyperparameters.

Key Results on MNIST

• No backpropagation — forward-only learning (River Learning)
• No GPU — single CPU core, 35 seconds
• No optimization tricks — no batch normalization, no data augmentation, no momentum
• Self-compressing — 256 gates automatically prune to 96 (62% removed)

CNN vs CGN

What's in this repo

• checkpoint/ — Trained weights (W1, W2) for h=128 configuration
• scripts/verify_mnist.py — Inference-only verification script
• scripts/visualize_gates.py — Gate receptive field and vote visualization
• scripts/compare_resolution.py — CNN vs CGN resolution comparison
• figures/ — Pre-generated visualizations
• results/ — Training logs

Verification

pip install numpy
python scripts/verify_mnist.py

Expected output: ~89.3% on the full 10K test set.

Note: The checkpoint was saved at a different epoch than the best test accuracy (90.4% at epoch 82).

Visualizations

Gate Receptive Fields

Each gate discovers its own spatial pattern from data — no filter shape prescribed.

CNN vs CGN: What Each Architecture Sees

CNN reduces 28×28 to 5×5 (97% information loss). CGN sees the full image.

CGN Architecture

Gate Neuron Detail

Paper Series

1. Forward-Only Path Carving Without Backpropagation (Zenodo, 2026)
2. Inference Is Learning: No Phase Separation (Zenodo, 2026)
3. One Gate, One Hundred Thousand Edges: Scaling to MNIST (Zenodo, 2026)
4. The Converged Structure Is the Explanation (Zenodo, 2026)
5. Confluence Gate Networks: From Biological Neuron to Standard Architecture (Zenodo, 2026)
6. Template Sharing and Network Design from Learning (upcoming)

Patent

Korean Patent Application 10-2026-0052624 (filed 2026). PCT filing planned.

License

The checkpoint and inference scripts are provided for verification and research purposes only. The training algorithm (River Learning) is proprietary and not included in this repository.

Contact

Yeonseong Cynn — [email protected]