Reproducibility Challenge @ NeurIPS 2019 Learning Robust Global Representations by Penalizing Local Predictive Power (Replication Track)

Guanqun Yang

Abstract:

This work focuses on the replication of some of the results presented in from scratch. In particular, we implement the patch-wise adversarial regularization (PAR) and its variants and apply them to two-layer CNN and ResNet for domain adaptation (DA) and domain generalization (DG) tasks, respectively. The comprehensive experiments confirm the claims made in the paper, showing that

- The PAR and its variants could improve out-of-domain performance than baseline models.
- The variants of the PAR do not consistently improve upon the vanilla PAR across architectures and datasets.

Besides, we make following extension

- A review of the variants and application of the adversarial training approach, which is the origin of the PAR proposed in the paper.
- A dissection of the approach described in the paper by providing additional implementation and experiment details.

Eventually, the code for this work is hosted on GitHub
URL: https://github.com/guanqun-yang/reproducibility_challenge_neurips_2019

Track: Replicability

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