NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:3328
Title:Deep Model Transferability from Attribution Maps

The contribution allows to compare heterogeneous networks by projecting them in a model space based on attribution maps (a task & model dependent attention map over inputs). The distance between the embedding of the networks in the model space is used as good metric for measuring task transferability in a very simple and cheap way. This is interesting and allows to provide good estimators of transfer capability. The justification of some choices could be improved with the help of the reviews.