MCRM: Mother Compact Recurrent Memory

This repository contains the experiments in MCRM: Mother Compact Recurrent Memory work by Abduallah Mohamed and Christian Claudel.

MCRM is a nested LSTM-GRU RNN architecture that has a compact memory pattern. The memory pattern combines both long and short-term behaviors.

@ARTICLE{2018arXiv180802016M,
   author = {{Mohamed}, A.~A. and {Claudel}, C.},
    title = "{MCRM: Mother Compact Recurrent Memory}",
  journal = {ArXiv e-prints},
archivePrefix = "arXiv",
   eprint = {1808.02016},
 keywords = {Computer Science - Neural and Evolutionary Computing, Computer Science - Machine Learning, Statistics - Machine Learning},
     year = 2018,
    month = aug,
   adsurl = {http://adsabs.harvard.edu/abs/2018arXiv180802016M},
  adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}

MCRM data flow diagram:

Supplied implementation

class MCRMCell which is MCRM cell .
class MCRM which is a generic class to utilizie multiple MCRM cells. It has the same Pytorch RNN modules signatures. both classes are available here

How to run the experiments

There's a total of 5 experiments in this repo.

Adding test: python ./MCRM/adding_problem/add_test.py --model MCRM
Copy memory test: python ./MCRM/copy_memory/copymem_test.py --model MCRM
Sequential MNIST test: python ./MCRM/mnist_pixel/mnist_test.py --model MCRM
Char PTB test: python ./MCRM/char_ptb/char_ptb.py --model MCRM
Word PTB test: python ./MCRM/word_ptb/word_ptb.py --model MCRM