Reproducibility Challenge: Meta-LearningRepresentations for Continual Learning

Mihaela Georgieva Stoycheva, Sergio Liberman Bronfman, Konstantinos Saitas Zarkias

Abstract: In this report we briefly introduce the MRCL method presented in Meta-Learning Representations for Continual Learning by Javed and White [2019] for learning representations which promote fast model adaptability to new tasks while trying to minimise catastrophic forgetting, which is the problem of continual learning.Additionally, we report results of our implementation based on the presented approach. Based on our level of knowledge in this topic, we found that the article is not reproducible on its own because several details essential for its implementation are not explicitly stated in the paper or are misleading. Finally, we discuss the issues and differences between expected and obtained results. Our implementation code can be found here: https://github.com/sergiolib/reproduce_oml using TensorFlow 2, while the original paper’s code is written using PyTorch.

Track: Replicability

NeurIPS Paper Id: https://openreview.net/forum?id=HJGqrEBgIr