Sparse-SNN — connectome-inspired sparse spiking networks

Sparse spiking neural networks whose structural masks are derived from the statistical laws of the real Drosophila connectome — matching dense MLP accuracy at ~2 orders of magnitude lower energy.

Two trained classifiers: MNIST and Fashion-MNIST. Architecture: 784 → 800 → 10 with 5% connection density, LIF neurons and surrogate-gradient training (straight-through estimator with sigmoid gradient).

Files

File	Description
sparse_snn_mnist.pt	MNIST checkpoint — 96.83% test accuracy (firing rate 7.9%)
sparse_snn_fashion.pt	Fashion-MNIST checkpoint — 87.07% test accuracy (firing rate 8.2%)
train_results.json	Raw metrics from the training runs

Each checkpoint contains the state_dict (weights + mask buffer), the E/I sign matrix, the full config, and metrics.

Measured results (these checkpoints)

Task	Dense MLP (documented)	Sparse SNN (this checkpoint)	Gap	Energy saving*
MNIST	98.32%	96.83%	1.5 pts	~112×
Fashion-MNIST	87.56%	87.07%	0.5 pts	~107×

* Energy estimate using the repository's 45nm CMOS model (Horowitz 2014: MAC = 3.7 pJ, addition = 0.9 pJ) with this checkpoint's measured firing rate — event-driven spike ops are additions only. This is a model-based estimate, not measured on hardware. The project's documented headline range across runs is 105–106×.

What the research found (and disproved)

1. Disproved: the MaleCNS connectome topology as a static structure carries no measurable advantage over random / structured-sparse graphs (5 experiments: classification, temporal, robustness, sample-efficiency, plasticity — all indistinguishable).
2. Extracted — the topology's statistical laws that DO matter: sparsity 0.09%, long-tail degree distribution (scale-free, max/mean ≈ 75), strong small-worldness (clustering 6.65× random, path length 2.39), E/I ratio 60/40.
3. Built: a trainable sparse SNN from these laws → near-lossless accuracy at ~105–112× energy savings.

Usage

Requires the model code from the repository (the class is small and self-contained):

import torch
from sparse_snn import SparseSNN   # github.com/AwareLiquid/human-brain-simulation

ckpt = torch.load("sparse_snn_mnist.pt", map_location="cpu", weights_only=False)
cfg = ckpt["config"]
model = SparseSNN(cfg["in_dim"], cfg["hid_dim"], cfg["out_dim"],
                  ckpt["state_dict"]["mask"], ckpt["sign"],
                  T=cfg["T"], decay=cfg["decay"], threshold=cfg["threshold"])
model.load_state_dict(ckpt["state_dict"])
model.eval()
# inputs: flattened 28x28 grayscale in [0, 1]

Loading was verified to reproduce the saved accuracy bit-exactly (acc=0.9683 MNIST / acc=0.8707 Fashion on reload).

Honest boundaries

• The accuracy gap vs a dense MLP is small but real (1.5 pts MNIST, 0.5 pts Fashion) — "near-lossless", not lossless.
• The connectome's static topology gave no advantage; only its statistical laws transferred. Do not attribute the result to "copying the fly brain".
• Energy figures are analytic (45nm CMOS), not measured on neuromorphic hardware.
• Recurrent spiking training remains an open problem in this line (sMNIST from scratch reaches only ~70%; conversion methods do better but lose efficiency).
• Masks are generated with a fixed seed (mask_seed=0) — rerun the code in the repository to regenerate the exact structure.

Related

• Code & full experiment report: AwareLiquid/human-brain-simulation
• AwareLiquid/M1 — MT-LNN liquid architecture
• awareliquid.ai — benchmarks and retractions

Try it

• 🎮 Interactive demo (runs in your browser, ONNX Runtime Web): https://huggingface.co/spaces/AwareLiquid/Sparse-SNN-demo

License

MIT. MaleCNS connectome data is CC-BY 4.0 (HHMI Janelia / Cambridge / Google Research, male-cns.janelia.org) — not redistributed here.