University of Illinois Urbana-Champaign

An exploration into the effect of thresholding feature maps of convolutional neural network in frequency domain

Yu, Mang

This item is only available for download by members of the University of Illinois community. Students, faculty, and staff at the U of I may log in with your NetID and password to view the item. If you are trying to access an Illinois-restricted dissertation or thesis, you can request a copy through your library's Inter-Library Loan office or purchase a copy directly from ProQuest.

Permalink

https://hdl.handle.net/2142/100060

Description

Title
An exploration into the effect of thresholding feature maps of convolutional neural network in frequency domain
Author(s)
Yu, Mang
Contributor(s)
Chen, Deming
Issue Date
2018-05
Keyword(s)
  • CNN
  • Sparse Fourier Transform
Abstract
While convolutional neural networks (CNNs) are very successful in many areas, the state-of-the-art multi-layer CNNs usually require a large amount of computation, which limits their application in scenarios where the computation capability is limited. Since the convolution operation can be done efficiently in the frequency domain, researchers have successfully reduced the amount of computation by applying the fast Fourier transform (FFT) and its inverse to the CNN. Furthermore, the sparse Fourier transform (SFT) algorithm can further reduce the amount of computation by only extracting the salient points in the frequency domain. However, due to this feature, it requires the inputs to be sparse or approximately sparse in the frequency domain. To explore the possibility of applying the SFT to CNN, we simulate the effect of SFT by removing the frequencies with lower power. We refer to this operation as Thresholding. In the experiments, we first inspect the effect of removing low-power frequencies for a sample feature map extracted from intermediate outputs. The result shows that most features are still identifiable to human eyes when 90% of the frequencies are removed; thus, it is possible that CNN can still recognize the features. We then apply the thresholding to each individual layer of VGG-16 and test the accuracy over the ILSVRC2012 dataset. The result shows that thresholding each layer only slightly reduced the accuracy and the reductions are smaller for the top layers (layer close to the output). However, thresholding uniformly on every layer of the network significantly reduced the accuracy. Therefore, we conclude that we should apply SFT to different layers with different configurations to achieve the optimal balance between performance and accuracy. In addition, layer-by-layer fine-tuning and image processing techniques might also help reducing the accuracy loss.
Type of Resource
text
Language
en
Permalink
http://hdl.handle.net/2142/100060

Owning Collections

Senior Theses - Electrical and Computer Engineering PRIMARY
The best of ECE undergraduate research