mnist-convnet

A convolutional spiking network trained on MNIST with NengoDL, deployed on Loihi neuromorphic hardware.

Description

This submission demonstrates the canonical Nengo "train in TensorFlow, deploy on Loihi" workflow:

Define a convolutional architecture using NengoDL's TensorFlow integration.
Train it on MNIST with backprop (TensorFlow handles the gradient flow).
Convert the resulting weights into a spiking Nengo network.
Run inference on Loihi (or a Loihi simulator if you don't have hardware).

The full walkthrough — architecture, training loop, weight analysis, and Loihi deployment — lives in mnist-convnet.ipynb. Open it in Jupyter.

Run it

This submission requires a fairly heavy stack. We strongly recommend a fresh venv:

python -m venv .venv
source .venv/bin/activate
pip install nengo nengo-dl tensorflow nengo-loihi matplotlib requests
jupyter notebook mnist-convnet.ipynb

For inference-only (no training, faster), the notebook has a do_training = False toggle that loads pre-trained weights from the Nengo examples bucket.

A note on CI

This submission is ci_runnable: false — the Zoo's CI only verifies metadata and structure, not runtime, because:

Training requires TensorFlow + GPU (or a long CPU pass) to be meaningful.
The Loihi deployment step requires either Intel's nengo_loihi simulator (~slow) or actual Loihi hardware (no GitHub Actions runners).
The combined install (nengo_dl + tensorflow + nengo_loihi) is multi-hundred-MB and slow to provision per CI run.

The nengo_dl and nengo_loihi backends in metadata.yaml are author-claimed, not CI-verified. Future zoo infrastructure could add an optional heavy-CI lane that runs against a Docker image with the full stack pre-installed; that's tracked as a design question.

Citation

@article{hunsberger2016training,
  author  = {Hunsberger, Eric and Eliasmith, Chris},
  title   = {Training spiking deep networks for neuromorphic hardware},
  journal = {arXiv preprint arXiv:1611.05141},
  year    = {2016}
}