A baseline regularization scheme for transfer learning with convolutional neural networks

作者：

Highlights：

• The standard L2 regularization is not adequate for transfer learning problems.

• We recommend regularizers that drive parameters towards the pre-trained model.

• Experimental results in image classification and segmentation favor this scheme.

• Analyses and some theoretical insights are proposed.

摘要

•The standard L2 regularization is not adequate for transfer learning problems.•We recommend regularizers that drive parameters towards the pre-trained model.•Experimental results in image classification and segmentation favor this scheme.•Analyses and some theoretical insights are proposed.

论文关键词：Transfer learning,Regularization,Convolutional networks

论文评审过程：Received 3 August 2018, Revised 10 July 2019, Accepted 10 September 2019, Available online 11 September 2019, Version of Record 14 September 2019.

论文官网地址：https://doi.org/10.1016/j.patcog.2019.107049