Neural Architecture Search for Efficient Convolutional Networks: A Multi-Objective Optimization Approach

Posted: Aug 30, 2024

• Wei Zhang
• Maria Rodriguez
• Kenji Tanaka
• Fatima Al-Mansoori

Abstract

This paper presents a novel neural architecture search (NAS) framework that optimizes convolutional neural networks for both accuracy and computational efficiency. Traditional NAS methods often prioritize accuracy while neglecting computational constraints, leading to impractical models for real-world deployment. Our approach employs a multi-objective evolutionary algorithm that simultaneously optimizes network architecture across multiple performance metrics including accuracy, parameter count, and FLOPs. We introduce a hierarchical search space that enables efficient exploration of architectural variations while maintaining structural coherence. Experimental results on CIFAR-10 and ImageNet datasets demonstrate that our method generates architectures that achieve competitive accuracy with significantly reduced computational requirements compared to hand-designed networks and existing NAS approaches. The proposed framework reduces parameter count by up to 45% and computational operations by 38% while maintaining comparable accuracy to state-of-the-art models.