# Random noise promotes slow heterogeneous synaptic dynamics important for robust working memory computation This repository provides all scripts for training recurrent neural networks (RNNs) with random noise, as well as the trained RNNs themselves used in the analyses presented in our paper. For more details, see our preprint on bioRxiv (https://www.biorxiv.org/content/10.1101/2022.10.14.512301v4). ## Code for training RNNs with random noise The code used to train RNNs with random noise has been adapted from existing code in this repository (https://github.com/rkim35/spikeRNN). The main script (`main_noise.py`) uses a similar set of input parameters as the original code (https://github.com/rkim35/spikeRNN), and includes the following additional parameters: - `osc`: Specifies the type of noise used during training. If set to `none`, no noise will be included during training. If set to `noise`, random Gaussian noise is added. - `mod_n`: Indicates the number of noise channels to include. If `osc` (see above) is set to `none`, then this input argument is ignored. Please see `example.sh` for an example on how to train RNNs with ten channels of random Gaussian noise. ## Data folder structure - `no_noise/`: contains 50 RNNs trained to perform the delayed match-to-sample (DMS) task without any noise - `noise_1/`: contains 50 RNNs trained to perform the delayed match-to-sample (DMS) task with one channel of Gaussian noise - `noise_5/`: contains 50 RNNs trained to perform the delayed match-to-sample (DMS) task with five channels of Gaussian noise - `noise_10/`: contains 50 RNNs trained to perform the delayed match-to-sample (DMS) task with ten channels of Guassian noise - `noise_20/`: contains 50 RNNs trained to perform the delayed match-to-sample (DMS) task with 20 channels of Guassian noise - `noise_50/`: contains 50 RNNs trained to perform the delayed match-to-sample (DMS) task with 50 channels of Guassian noise