CapsNet vs CNN: Analysis of the Effects of Varying Feature Spatial Arrangement

Citation

Manogaran, Ugenteraan and Wong, Ya Ping and Ng, Boon Yian (2021) CapsNet vs CNN: Analysis of the Effects of Varying Feature Spatial Arrangement. Advances in Intelligent Systems and Computing, 1251. pp. 1-9. ISSN 2194-5357

Full text not available from this repository.

Abstract

Despite the success over the recent years, convolutional neural network (CNN) has a major limitation of the inability to retain spatial relationship between learned features in deeper layers. Capsule network with dynamic routing (CapsNet) was introduced in 2017 with a speculation that CapsNet can overcome this limitation. In our research, we created a suitable collection of datasets and implemented a simple CNN model and a CapsNet model with similar complexity to test this speculation. Experimental results show that both the implemented CNN and CapsNet models have the ability to capture the spatial relationship between learned features. Counterintuitively, our experiments show that our CNN model outperforms our CapsNet model using our datasets. This implies that the speculation does not seem to be entirely correct. This might be due to the fact that our datasets are too simple, hence requiring a simple CNN model. We further recommend future research to be conducted using deeper models and more complex datasets to test the speculation.

Item Type: Article
Uncontrolled Keywords: Neural networks, CapsNet, Convolutional Neural Network, Spatial relationship
Subjects: Q Science > QA Mathematics > QA71-90 Instruments and machines > QA75.5-76.95 Electronic computers. Computer science
Divisions: Faculty of Computing and Informatics (FCI)
Depositing User: Ms Nurul Iqtiani Ahmad
Date Deposited: 08 Mar 2021 22:53
Last Modified: 08 Mar 2021 22:53
URII: http://shdl.mmu.edu.my/id/eprint/8560

Downloads

Downloads per month over past year

View ItemEdit (login required)