---
license: apache-2.0
tags:
  - spiking-neural-network
  - neuromorphic
  - loihi
  - temporal-classification
  - audio
datasets:
  - zenke-lab/spiking-heidelberg-digits
metrics:
  - accuracy
model-index:
  - name: Catalyst SHD SNN
    results:
      - task:
          type: audio-classification
          name: Spoken Digit Recognition
        dataset:
          name: Spiking Heidelberg Digits (SHD)
          type: zenke-lab/spiking-heidelberg-digits
        metrics:
          - name: Test Accuracy (float32)
            type: accuracy
            value: 85.9
          - name: Test Accuracy (int16 quantized)
            type: accuracy
            value: 85.4
---

# Catalyst SHD Spiking Neural Network

A recurrent spiking neural network trained on the [Spiking Heidelberg Digits (SHD)](https://zenkelab.org/resources/spiking-heidelberg-datasets-shd/) benchmark, achieving **85.9% test accuracy** (float) / **85.4% quantized (int16)**.

Trained using the [Catalyst Neuromorphic](https://catalyst-neuromorphic.com) SDK (`neurocore`) with surrogate gradient descent, and deployable directly to Loihi-compatible neuromorphic hardware via the Catalyst Cloud API.

## Model Details

| Parameter | Value |
|-----------|-------|
| **Architecture** | Recurrent LIF (Leaky Integrate-and-Fire) |
| **Input neurons** | 700 (cochlea spike channels) |
| **Hidden neurons** | 512 (recurrent) |
| **Output classes** | 20 (digits 0-9, German + English) |
| **Total parameters** | ~1.49M |
| **Time bins** | 250 (dt = 4ms, 1s duration) |

## Training Configuration

| Hyperparameter | Value |
|----------------|-------|
| Epochs | 200 |
| Batch size | 128 |
| Optimizer | AdamW |
| Learning rate | 1e-3 (cosine annealing) |
| Weight decay | 1e-4 |
| Dropout | 0.3 |
| Gradient clipping | 1.0 norm |
| Surrogate gradient | Fast sigmoid (scale=25.0) |
| Membrane decay (hidden) | 0.95 (learnable) |
| Membrane decay (output) | 0.9 (learnable) |
| Threshold | 1.0 |
| Reset mechanism | Hard reset: v = v * (1 - spike) |

## Hardware Deployment

The model is quantized to **int16** for direct deployment on Catalyst neuromorphic hardware (FPGA) and the [Catalyst Cloud API](https://catalyst-neuromorphic.com/cloud):

- Threshold: 1000 (int16)
- Decay mapping: `decay_v = round(beta * 4096)` (CUBA neurons)
- Quantization accuracy loss: only **0.4%** (85.9% -> 85.4%)

## Benchmark Context

| Method | SHD Accuracy |
|--------|-------------|
| Cramer et al. (2020) | 83.2% |
| Zenke & Vogels (2021) | 83.4% |
| **Catalyst (this model)** | **85.9%** |

## Checkpoint Format

PyTorch `.pt` file containing:
```python
{
    'epoch': int,
    'model_state_dict': OrderedDict,  # PyTorch model weights
    'test_acc': 0.859,
    'args': dict  # Complete training arguments
}
```

## Usage

```python
import torch

checkpoint = torch.load("shd_model.pt", map_location="cpu")
print(f"Test accuracy: {checkpoint['test_acc']:.1%}")

# Load weights into your model
model.load_state_dict(checkpoint['model_state_dict'])
```

Or use the Catalyst Cloud API:
```bash
pip install catalyst-cloud
```

```python
from catalyst_cloud import CatalystClient
client = CatalystClient(api_key="your-key")
result = client.simulate(network_config={...})
```

## Citation

If you use this model, please cite:
```bibtex
@misc{shulayevbarnes2026catalyst,
    title={Catalyst N1: An Open-Design Neuromorphic Processor with Full Loihi Parity},
    author={Shulayev Barnes, Henry},
    year={2026},
    publisher={Zenodo}
}
```

## Links

- [Website](https://catalyst-neuromorphic.com)
- [Cloud API](https://catalyst-neuromorphic.com/cloud)
- [PyPI: catalyst-cloud](https://pypi.org/project/catalyst-cloud/)
- [Paper (Zenodo)](https://zenodo.org/records/14868368)