Gradient rectified parameter unit of the fully connected layer in convolutional neural networks

作者：

Highlights：

• Gradient rectified fully connected layer is proposed to remove irrelevant params.

• A simplified GRU-FC for single FC layer is introduced for improving ResNets.

• Convergence and generalization bounds of GRU-FC are discussed theoretically.

• Sufficient experiments support GRU-FC improves various CNNs’ recognition accuracy.

摘要

•Gradient rectified fully connected layer is proposed to remove irrelevant params.•A simplified GRU-FC for single FC layer is introduced for improving ResNets.•Convergence and generalization bounds of GRU-FC are discussed theoretically.•Sufficient experiments support GRU-FC improves various CNNs’ recognition accuracy.

论文关键词：Fully connected layer,Grad-CAM,Gradient,Backpropagation

论文评审过程：Received 29 January 2021, Revised 8 April 2022, Accepted 9 April 2022, Available online 18 April 2022, Version of Record 9 May 2022.

论文官网地址：https://doi.org/10.1016/j.knosys.2022.108797