Mashinini, Thabang L2022-12-212022-12-212022https://hdl.handle.net/10539/33910A dissertation submitted to the School of Computer Science And Applied Mathematics, Faculty of Science, University of the Witwatersrand, in fulfillment of the requirements for the degree of Master of Science, 2022The back-propagation of Recurrent Neural Networks (RNNs) is computationally costly andunstable, resulting in vanishing and exploding gradients. Echo State Networks (ESNs) are a modern method of training RNNs that is computationally efficient. Unlike RNNs, the ESN has an untrained RNN called a reservoir, and the reservoir to output connection are the onlycomponents of the ESN that are trained using a linear approach (i.e. linear regression) without using back-propagation. However, setting up the parameters of the reservoir in order to achieve optimal results is a challenge. This research aims to investigate the merits of using ESNs as an alternative approach to training RNNs, through a comparative study, applied to iterative segmentation problems. Under this iterative segmentation, we propose a novel approach to the current deep learning Variational Level Sets (VLS). The variational level set is formulated as a learnable spatiotemporal data-driven approach under ESNs. We investigate five approaches for learning the VLS Convolutional ESN (CESN), RNN (CRNN), Gated Recurrent Network (CGRU), Long Short-Term Memory (CLSTM), and a 3D Convolutional Neural Network (3DCNN). The CGRU and CLSTM are shown to be the best architectures for learning data-driven spatiotemporal VLS on four different databases in our experiments. Significant improvements in the segmentation results are achieved with an increase in data size. The memory of the ESNis related to the leaking rate and spectral radius of the reservoir. These parameters are majorc ontributors to the low performance of the CESN in learning the VLS. Contributions of this research include a novel formulation to learnable deep learning VLS as a spatiotemporal data-driven method; the investigation of the ESNs hyper-parameters and how they relate to the memory and performance of ESNs.enThesis