Automatic Labeling of Data for Transfer Learning

Parijat Dube, Bishwaranjan Bhattacharjee, Siyu Huo, Patrick Watson, John Kender, Brian Belgodere

Keywords: transfer learning, fine-tuning, divergence, pseudo labeling, automated labeling, experiments

TL;DR: A technique for automatically labeling large unlabeled datasets so that they can train source models for transfer learning and its experimental evaluation.

Abstract: Transfer learning uses trained weights from a source model as the initial weightsfor the training of a target dataset. A well chosen source with a large numberof labeled data leads to significant improvement in accuracy. We demonstrate atechnique that automatically labels large unlabeled datasets so that they can trainsource models for transfer learning. We experimentally evaluate this method, usinga baseline dataset of human-annotated ImageNet1K labels, against five variationsof this technique. We show that the performance of these automatically trainedmodels come within 17% of baseline on average.