Deconstructing Lottery Tickets: A Reproducibility Study

Numa Karolinski, Bartosz Miselis, Monisha Shcherbakova

Abstract:

The main aim of this project was to conduct a reproducibility study on a paper that
was presented at the NeurIPS 2019 Conference. The paper chosen was Deconstructing
Lottery Tickets: Zeros, Signs, and the Supermask, by Hattie Zhou, Janice Lan,
Rosanne Liu and Jason Yosinski [1]. An ablation track was followed for this study
to determine whether the conclusions of the paper are still consistent when certain
modifications to the original experiments are introduced; these include the effect of
the sign with magnitude increase mask criterion. Additionally, the statement that
"masking is training" was tested by applying the masks on the models that did not
fully converged. Due to limited computational resources, the study was narrowed
down to reproducing only the fully connected neural network, excluding all the
convolutional architectures. The code for all the experiments and figures presented
in this report is available on GitHub

Track: Ablation

NeurIPS Paper Id: https://openreview.net/forum?id=r1GnvVHeLB