Reproducibility Challenge: Analysis of robust classifiers and handling image synthesis tasks using representations learned from robust models

Tian Bai, Qiang Xu, Kevin Chen

Abstract:

Neural Information Processing Systems
(NeurIPS) holds a challenge to ensure that published articles
are reliable and reproducible. The goal of this report is
to study and reproduce experiment described in ”Image
Synthesis with a Single (Robust) Classifier” published by
Shibani Santurkar in 2019, where a basic classification
framework was used to tackle challenging tasks in image
synthesis such as image generation, inpainting, superresolution, etc. The CIFAR-10 dataset is chosen for this
experiment to compare the results with the original paper
on the image generation task. We also discovered a set
of parameters which the results might be more plausible

Track: Baseline

NeurIPS Paper Id: https://openreview.net/forum?id=Hkeer4Bx8S&noteId=HkxuZHGqcB