Analysis of RNN (paper under review, preprint https://arxiv.org/abs/2005.13074): Paper: Different eigenvalue distributions encode the same temporal tasks in recurrent neural networks.

https://link.springer.com/article/10.1007/s11571-022-09802-5

Simulations were generated using python 3.6.9 Tensorflow version 2.0 and Keras 2.3.1 Following the procedure described previously in https://arxiv.org/abs/1906.01094

The trained networks are saved in hdf5 format and can be opened using.

generate_figures.py

to generate all the figures corresponding to one particular realization using the corresponding testing set generator.

For example to test the "and" task, use generate_data_set_and.py

For example #_xor_10_ran has the figures corresponding to the Network 10 trained for Xor task with random normal initial condition.