Transfer learning with fewer ImageNet classes

Michal Kucer, Diane Oyen

Abstract: Though much previous work tried to uncover the best practices for transfer learning, much is left unexplored. Our preliminary work explores the effect of removing a portion of the ImageNet classes with low per-class validation accuracy on the accuracy of the remaining classes. Furthermore, we explore if models trained with a reduced number of classes are suitable for transfer learning.

Submission Track: Extended abstract track, 3 pages max

Reviewer Emails: [e-mail]