Repro_ICLR18 : reproduction of Dynamically Expandable Networks

This project aim is to reproduce the DEN model submitted at ICLR18.

It tries to reconstruct the perfomance plots presented in the original article on a variation of the MNIST dataset.

To run code:

1. unzip data archive
2. run main.py using python 3.5

This produces a figures/ directory and saves three plots :

• Online AUROC per task.
• Offline AUROC per task.
• Average AUROC after each task with timestamped inference (evaluation used in the original paper).

Poster :

The poster/ directory contains the pdf and the original latex files of the poster that illustrated this work.