On the difficulty of learning chaotic dynamics with RNNs

Code for the training of RNNs with sparsely forced BPTT.

This folder contains the python code, data files and plots from

"On the difficulty of learning chaotic dynamics with RNNs", Jonas Mikhaeil, Z. Monfared and D. Durstewitz.

@inproceedings{
mikhaeil2022on,
title={On the difficulty of learning chaotic dynamics with {RNN}s},
author={Jonas Magdy Mikhaeil and Zahra Monfared and Daniel Durstewitz},
booktitle={Advances in Neural Information Processing Systems},
editor={Alice H. Oh and Alekh Agarwal and Danielle Belgrave and Kyunghyun Cho},
year={2022},
url={https://openreview.net/forum?id=-_AMpmyV0Ll}
}

This package is distributed under the terms of the GNU GPLv3 & Creative Commons Attribution License. Please credit the source and cite the reference above when using the code in any form of publication.

---

main.py: Starts individual runs.

ubermain.py: Allows to start multiple runs simultanously. Used to sweep parameters, such as the learning interval $\tau$.

main_eval.py: Evaluates the reconstruction quality of (multiple) trained models and creates files such as klx.csv, which are used to create the plots in the paper.

CreateFigures.ipyn : code to create the figures from the csv files created in the model evaluation.

Datasets: contains all datasets.

Figures: contains all figures.