Transfer learning in computer vision tasks: Remember where you come from

作者：

Highlights：

• We replicated experiments on three state-of-the-art approaches to compare regularizers for fine-tuning.

• Our protocol ensures that there is no experimental bias.

• The regularizer that uses the pre-trained weights as a reference consistently outperforms weight decay in all experiments.

摘要

•We replicated experiments on three state-of-the-art approaches to compare regularizers for fine-tuning.•Our protocol ensures that there is no experimental bias.•The regularizer that uses the pre-trained weights as a reference consistently outperforms weight decay in all experiments.

论文关键词：Transfer learning,Parameter regularization,Computer vision

论文评审过程：Received 24 October 2019, Accepted 19 November 2019, Available online 29 November 2019, Version of Record 13 December 2019.

论文官网地址：https://doi.org/10.1016/j.imavis.2019.103853