How stable are Transferability Metrics evaluations?

Andrea Agostinelli
Michal Pandy
Vittorio Ferrari
ECCV (2022)

Abstract

Transferability metrics is a maturing field with increasing interest, which aims at providing heuristics for selecting the most suitable source models to transfer to a given target dataset, without finetuning them all. However, existing works rely on custom experimental setups which differ across papers, leading to inconsistent conclusions about which transferability metrics work best. In this paper we conduct
a large-scale study by systematically constructing a broad range of 715k experimental setup variations. We discover that even small variations to an experimental setup lead to different conclusions about the superiority of a transferability metric over another. Then we propose better evaluations by aggregating across many experiments, enabling to reach more stable conclusions. As a result, we reveal the superiority of LogME at selecting good source datasets to transfer from in a semantic segmentation scenario, and N LEEP at selecting good source architectures in an image classification scenario. However, no single transferability metric works best in all scenarios.

Research Areas