language_bottleneck

Here you can find a set of the games/tasks that are used the following paper:

• Entropy Minimization In Emergent Languages. Eugene Kharitonov, Rahma Chaabouni, Diane Bouchacourt, Marco Baroni. ICML 2020. arxiv.

To get a glimpse of some of the interesting things that can be done with Information Bottlenecks, see this tiny MNIST style transfer notebook: notebook / colab.

Structure

Each game directory contains the definition of the agents' architecture, training logic, hyperparameter grids compatible with the nest tool, and notebooks that were used to obtain the figures in the text.

guess_number/ # Guess Number task
    ./hyperparam_grid/ # json files with hyperparameter grids
mnist_classification/ # Image Classification task
mnist_overfit/ # MNIST overfitting experiments
mnist_adv/ # adversarial robustness experiments

Running

Launching each game is as simple as regular EGG games, e.g.

python -m egg.zoo.language_bottleneck.mnist_classification.train

For the set of game-specific parameters, please check each game's train.py script.

Reproducibility

If you want to recover results maximally close to those reported in the paper, please use EGG v1.0. This can be done by running the following command:

git checkout v1.0

In later versions of EGG, some metrics are aggregated differently, which might lead to small discrepancies.