Multi-objective optimization based robust automatic CNN model design

This dissertation presents an innovative method for the design of robust and automated convolutional neural network (CNN) models using multi-objective optimization techniques. While the design of deep learning models often requires expert knowledge and time-intensive manual processes, the proposed CESA method automates this process, optimizing both model accuracy and response time. The study leverages various multi-objective optimization algorithms (e.g., NSGA-II, FDB-NSGA-II) to generate Pareto-optimal solutions, achieving a successful balance between accuracy and speed. Experiments conducted on four different datasets demonstrate that the CESA method delivers 1.93% higher accuracy and 69.85% faster test times compared to existing CNN architectures. The ability to produce lightweight and fast models, particularly for low-capacity devices, validates the method's applicability to mobile and embedded systems. Furthermore, by fully automating hyper-parameter optimization, CESA enables the generation of flexible and effective models independent of user intervention, broadening its range of applications. This study highlights the power of multi-objective optimization approaches in CNN model design, serving as a significant guide for both academic research and industrial applications.