Quantization and Adaptivity of Wavelet Scattering Networks for Classification Applications
Open Access
- Author:
- Fox, Maxine Rebecca
- Graduate Program:
- Electrical Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- February 17, 2020
- Committee Members:
- Ram Mohan Narayanan, Thesis Advisor/Co-Advisor
Timothy Joseph Kane, Committee Member
Dr. Raghu G. Raj, Special Signatory
Kultegin Aydin, Program Head/Chair - Keywords:
- wavelet scattering networks
quantization
machine learning
backpropagation
convolutional neural networks - Abstract:
- GPR has been proposed as a solution for mine clearance and the removal of unexploded ordnance. To detect targets, convolutional neural networks (CNNs) may be used. However, the training of CNNs for this purpose is hampered by the lack of extensive GPR data on buried targets; therefore, this thesis investigates the use and application of CNNs on the freely available MSTAR dataset for target classification. Although CNNs improve upon the required memory of preceding neural networks, large CNNs still need large amounts of memory. Furthermore, despite the promising results of quantization methods, extensions to other datasets and architectures are not easily proven. To explore the effect of quantization, this thesis implements several quantization schemes on the wavelet scattering network (WSN), due to its shared properties with CNNs. Results indicate that the WSN is robust to few quantization levels; future work should focus on the application of these quantization schemes to CNNs, as well as the exploration of PDF-based quantization schemes. In addition, an adaptive WSN (AWSN) is constructed using the backpropagation algorithm inherent in CNNs. Using particle swarm optimization (PSO), the parameters of the mother and father wavelets are adjusted. Multiple AWSN constructions are compared to static WSN networks and a fully-connected layer network. Though the validation accuracy never exceeded 0.75, the results were within 0.05 for all AWSN implementations using PSO. Future work should explore modifications to parameter updates, as well as the implementation of existing adaptive wavelet methods.