Catalyst-Neuromorphic/shd-snn-benchmark

Catalyst SHD SNN Benchmark

Spiking Neural Network trained on the Spiking Heidelberg Digits (SHD) dataset using surrogate gradient BPTT. Achieves 91.0% on SHD with adaptive LIF neurons (90.8% quantised int16).

Model Description

• Architecture (N3): 700 → 1536 (recurrent adLIF) → 20
• Architecture (N2): 700 → 512 (recurrent adLIF) → 20
• Architecture (N1): 700 → 1024 (recurrent LIF) → 20
• Neuron model: Adaptive Leaky Integrate-and-Fire (adLIF) with learnable per-neuron thresholds
• Training: Surrogate gradient BPTT, fast-sigmoid surrogate (scale=25), cosine LR scheduling
• Hardware target: Catalyst N1/N2/N3 neuromorphic processors

Results

Generation	Architecture	Float Accuracy	Params
N3	700→1536→20 (rec, adLIF)	91.0%	3.47M
N2	700→512→20 (rec, adLIF)	84.5%	759K
N1	700→1024→20 (rec, LIF)	85.9%	1.79M

Reproduce

git clone https://github.com/catalyst-neuromorphic/catalyst-benchmarks.git
cd catalyst-benchmarks
pip install -e .

# N3
python shd/train.py --neuron adlif --hidden 1536 --epochs 200 --device cuda:0 --amp

# N2
python shd/train.py --neuron adlif --hidden 512 --epochs 200 --device cuda:0

# N1
python shd/train.py --neuron lif --hidden 1024 --epochs 200 --device cuda:0

Deploy to Catalyst Hardware

python shd/deploy.py --checkpoint shd_model.pt --threshold-hw 1000

Links

• Benchmark repo: catalyst-neuromorphic/catalyst-benchmarks
• Hardware: catalyst-neuromorphic.com
• N3 paper: Zenodo DOI 10.5281/zenodo.18881283
• N2 paper: Zenodo DOI 10.5281/zenodo.18728256

Citation

@misc{catalyst-benchmarks-2026,
  author = {Shulayev Barnes, Henry},
  title = {Catalyst Neuromorphic Benchmarks},
  year = {2026},
  url = {https://github.com/catalyst-neuromorphic/catalyst-benchmarks}
}