Skip to content
Conference Open access

Neural Architecture Search: Combines Quantum-Inspired PSO with Gradient Descent for Automated Deep Learning Design

Sep 2026 · 2026 Ninth International Conference on Vocational Education and Electrical Engineering (ICVEE) · pp. 87-92 · 0 citations · 19 references

Abstract

The manual design of deep neural network architectures remains a time-consuming and expertise-dependent process, often requiring extensive trial-and-error experimentation. This paper proposes a novel Neural Architecture Search (NAS) framework that synergistically combines quantum-inspired Particle Swarm Optimization (QPSO) with gradient descent for automated deep learning design. We evaluate our approach on 45 benchmark datasets spanning computer vision, natural language processing, and time-series forecasting tasks, including validation on three real-world industrial applications. The methodology employs quantum rotation gates to enhance exploration capabilities of traditional PSO, while gradient descent refines promising architectural candidates through continuous optimization of network weights and structural parameters. Our hybrid search strategy achieves an average accuracy improvement of 3.7% over manually designed architectures and 2.4% over baseline NAS methods, with 40% reduction in computational cost compared to evolutionary and reinforcement learning-based approaches. On ImageNet classification, the discovered architecture attains 82.3% top-1 accuracy with only 4.2M parameters, demonstrating competitive performance against state-of-the-art efficient networks. For industrial deployment scenarios involving limited computational budgets, the framework successfully identifies optimal architectures within 12 GPU-hours, compared to 450+ hours required by conventional NAS methods. This research advances the field of automated machine learning by demonstrating that quantum-inspired metaheuristics, when properly integrated with gradient-based optimization, can significantly accelerate neural architecture discovery while maintaining superior model quality.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.