The Capsule Neural Network (CapsNet) architecture addresses several limitations that exist in Deep Learning (DL) architectures such as sub-standard training/learning due to limited availability of training data, data complexity, inter-class similarities, and invariance generated by mean/max poolings. This thesis investigates the benefits of the CapsNet approach to enhance performance and extract knowledge across a range of datasets. Contributions to the advancement of the CapsNets and its deployment in this thesis are categorised into three main areas: architecture development, routing reconstruction, and evaluation on multiple data types/applications. We first introduce the Hybrid Capsule Network (HCapsNet), an end-to-end DL architecture specifically designed for Hyperspectral Image (HIS) classification in remote sensing. By applying 2D- and 3D- Convolutional Neural Network (CNN) to extract higher-level spatial and spectral features, HCapsNet achieves superior performance compared to state-of-the-art methods in terms of overall classification accuracy on three widely used hyperspectral datasets, Indian Pines dataset, the University of Pavia dataset, the Salinas Valley dataset when using only 1% of the data for training. Second, we present the Speech-CapsNet (SCapsNet), a CapsNet-based architecture tailored for decoding inner/imagined speech from electroencephalography (EEG). SCapsNet incorporates multi-level feature maps and multiple capsule layers to enhance the routing reconstruction and standard CapsNet architecture. The models' performance was evaluated using two datasets: 1) an Imagined-Speech dataset, and 2) Overt- and Imagined-Speech dataset. In both datasets, SCapsNet significantly outperformed a Deep-CNN, and Shallow-CNN approaches. In the third contribution, to further thoroughly evaluate the potential for CapsNets, we applied CapsNet to include the analysis of infant intrinsic movement recorded through Motion Capture (MoCap) techniques. The data were obtained from a mobile conjugate reinforcement (MCR) experiment which explored mechanisms underlying agentive discovery through analysis of infant motion, mobile motion and their coordination dynamics. Movement data were collected using a 3D MoCap system with motion markers placed on the infant’s body and the mobile. To analyse the data, the 1D-CapsNet and 2D-CapsNet architectures are employed to classify the experiment's assigned stages and identify the most active section within a sliding window scenario. The results suggest that the 2D-CapsNet architecture achieved the highest accuracy for the fused Feet feature set (p < 0.0005) and was effective in capturing and analysing the movement patterns of infants during the contingency feedback experiment.
- Capsule network
- Deep learning
Refined capsule network architectures for data classification and knowledge discovery
Khodadadzadeh, M. (Author). Jan 2024
Student thesis: Doctoral Thesis