Evaluating Lottery Tickets Under Distributional Shifts

TL;DR

This paper investigates whether sparse subnetworks identified by the Lottery Ticket Hypothesis maintain their effectiveness under distributional shifts across different domains, highlighting their potential for transferability.

Contribution

It evaluates the transferability and initialization strategies of lottery ticket subnetworks under domain shifts, revealing their generalization capabilities.

Findings

Sparse subnetworks reflect an inductive bias of neural networks.

Lottery tickets can be re-trained in dissimilar domains.

Initialization strategies impact transfer performance.

Abstract

The Lottery Ticket Hypothesis suggests large, over-parameterized neural networks consist of small, sparse subnetworks that can be trained in isolation to reach a similar (or better) test accuracy. However, the initialization and generalizability of the obtained sparse subnetworks have been recently called into question. Our work focuses on evaluating the initialization of sparse subnetworks under distributional shifts. Specifically, we investigate the extent to which a sparse subnetwork obtained in a source domain can be re-trained in isolation in a dissimilar, target domain. In addition, we examine the effects of different initialization strategies at transfer-time. Our experiments show that sparse subnetworks obtained through lottery ticket training do not simply overfit to particular domains, but rather reflect an inductive bias of deep neural networks that can be exploited in multiple domains.