Optimization of the Neural Network Structure of the Bilevel Multi-Objective Generalized Quantum PSO Algorithm for Image Recognition

• Published in: Chinese Control Conference pp. 8541 - 8546
• Main Authors: Xu, Wenxue, Zhu, Qidan, Mou, Jinyou
• Format: Conference Proceeding
• Language: English
• Published: Technical Committee on Control Theory, Chinese Association of Automation 28.07.2025
• Subjects: Accuracy; Bilevel optimization problem; Computational complexity; Computer architecture; Gaussian distribution; Image classification; Model compression; multi-objective optimization; Neural architecture search; Neural networks; Optimization; Particle swarm optimization; Quantum computation; Quantum computing
• ISSN: 1934-1768
• DOI: 10.23919/CCC64809.2025.11179020

With the development of deep learning, the traditional approach of manually designing network archi-tectures and determining parameters inevitably loses its reliance on expert systems, which has brought significant challenges to our parameter configuration. To address this issue, we adopted an automated search architecture, reducing manpower while enhancing efficiency. Based on this, this paper proposes the Multi-Objective Generalized Quantum Particle Swarm Optimization (MOGQPSO) algorithm, which combines quantum behavior and Gaussian distribution characteristics, introducing a brand-new mode for the movement and update of particles in the search space. Simultaneously, a bilevel optimization algorithm is designed. A particle encoding scheme is utilized to effectively represent the CNN architecture to assist in optimization. Through multiple iterations of MOGQPSO, the optimal solutions for the upper and lower levels are sought. Superior upper-level solutions are selected for lower-level optimization, and the optimized parameters of the lower level are fed back to the upper level. The upper and lower levels cooperate and iterate continuously. Moreover, combined with the decoding method in NSGA-Net, the entire CNN architecture is jointly designed. Experimental results demonstrate that the multi-objective optimal solution set obtained by this algorithm has been significantly improved compared to traditional algorithms, achieving a reduction of nearly one million parameters while enhancing accuracy.