Self-Attention in Convolutional Neural Networks: a Potentially Fundamental Improvement for Image Classi cation

Keywords

Loading...
Thumbnail Image

Issue Date

2021-06-18

Language

en

Document type

Journal Title

Journal ISSN

Volume Title

Publisher

Title

ISSN

Volume

Issue

Startpage

Endpage

DOI

Abstract

Conventional (feedforward) Deep Neural Networks (DNNs) have some in- herent limitations and weaknesses that humans do not have. In the domain of computer vision, this becomes clear when a CNN is confronted with ad- versarial attacks. An explanation for this di erence in performance and robustness between a CNN and the human brain could be the usage of recurrence. The visual cortex of the human brain amply uses feedback con- nections when processing visual input. Most conventional CNNs, however, do not make use of this type of connections. Therefore, we believe that recurrence might help improve performance of CNNs, at least in certain challenging tasks. In this paper, we propose a kind of recurrence that is based on top-down attention. For our experiments, we used a feedforward CNN as a benchmark network and compared it to an equivalent network but with top-down attention. We compared the performance of both models in a digit classi cation task that was complicated by partial occlusion. In this task, the recurrent network signi cantly outperformed the benchmark net- work. Although further research is needed to make de nitive conclusions, the results provide promising initial evidence that the use of recurrence in a CNN can lead to fundamental improvements in image classi cation.

Description

Citation

Faculty

Faculteit der Sociale Wetenschappen