DK-CNNs: Dynamic kernel convolutional neural networks

Jialin Liu, Fei Chao*, Chih Min Lin, Changle Zhou, Changjing Shang

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

114 Downloads (Pure)

Abstract

This paper introduces dynamic kernel convolutional neural networks (DK-CNNs), an enhanced type of CNN, by performing line-by-line scanning regular convolution to generate a latent dimension of kernel weights. The proposed DK-CNN applies regular convolution to the DK weights, which rely on a latent variable, and discretizes the space of the latent variable to extend a new dimension; this process is named “DK convolution”. DK convolution increases the expressive capacity of the convolution operation without increasing the number of parameters by searching for useful patterns within the new extended dimension. In contrast to conventional convolution, which applies a fixed kernel to analyse the changed features, DK convolution employs a DK to analyse fixed features. In addition, DK convolution can replace a standard convolution layer in any CNN network structure. The proposed DK-CNNs were compared with different network structures with and without a latent dimension on the CIFAR and FashionMNIST datasets. The experimental results show that DK-CNNs can achieve better performance than regular CNNs.

Original languageEnglish
Pages (from-to)95-108
Number of pages14
JournalNeurocomputing
Volume422
Early online date12 Sept 2020
DOIs
Publication statusPublished - 21 Jan 2021

Keywords

  • Convolution kernel
  • Convolutional neural networks
  • Deep neural networks

Fingerprint

Dive into the research topics of 'DK-CNNs: Dynamic kernel convolutional neural networks'. Together they form a unique fingerprint.

Cite this