Skip to content

Source code for our paper "Learning to Communicate with Strangers via Channel Randomisation Methods"

License

Notifications You must be signed in to change notification settings

DylanCope/zero-shot-comm

Repository files navigation

Learning to Communicate with Strangers via Channel Randomisation Methods

This is the source code for our paper "Learning to Communicate with Strangers via Channel Randomisation Methods" (Cope and Shoots, 2020) presented at the 4th NeurIPS Workshop on Emergent Communication.

Check out our poster for a quick overview of the results!

Installation

There are two ways of installing and running this codebase.

  • One way is to use the environment.yml to recreate the conda environment.

  • Another way is to create a Docker container using the provided Dockerfile and scripts. The image inherits from tensorflow/tensorflow:latest-gpu-jupyter. If you do not have gpu access then you will need to adapt the Dockerfile and change the run script (remove the --runtime nvidia parameter).

Repository Structure

  • notebooks: this folder contains the Jupyter Notebooks that contain our experiments.
  • zscomm: this is a Python package written to house the core code for our project.
  • experiments: this contains all the data produced during our experiments: i.e. network weights and train/test metrics.
  • figures: figures generated by our experiments

Citing this work

Cope, Dylan R. & Schoots, Nandi (2020), 'Learning to Communicate with Strangers via Channel Randomisation Methods', Emergent Communication Workshop at the 34th Conference on Neural Information Processing Systems

Acknowledgements

This work was done at the UKRI Centre for Doctoral Studies in Safe and Trusted Artificial Intelligence.

About

Source code for our paper "Learning to Communicate with Strangers via Channel Randomisation Methods"

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published