A JAX re-implementation of e-prop algorithm. eprop-jax is written to be simple, clean and fast.
I tried to replicate the pattern generation task in eprop/dataset.py, as described in the paper. And the implementation eprop/model.py includes only Leaky Integrate-and-Fire (LIF) neurons.
Since the implementation is JIT compiled, I suggest doing experimentation/modifications on a Jupyter notebook e.g. regression_task.ipynb, to avoid XLA compilation at every run.
It's just enough to run python eprop/train.py to test the network with default hyperparameters.
$ python eprop/train.py
Epoch: [0/100] - MSE Loss: 0.3905
Epoch: [10/100] - MSE Loss: 0.2088
Epoch: [20/100] - MSE Loss: 0.0821
Epoch: [30/100] - MSE Loss: 0.0395
Epoch: [40/100] - MSE Loss: 0.0307
Epoch: [50/100] - MSE Loss: 0.0265
Epoch: [60/100] - MSE Loss: 0.0187
Epoch: [70/100] - MSE Loss: 0.0221
Epoch: [80/100] - MSE Loss: 0.0173
Epoch: [90/100] - MSE Loss: 0.0217
This code is inspired from the PyTorch implementation of e-prop by Charlotte Frenkel.
MIT