Differentiable Neural Architecture Search Based on Hybrid Attention for Classification

In recent years, the rapid development of deep learning has highlighted the growing challenge of high computational costs in traditional neural architecture design. This has spurred significant research interest in efficient automated neural architecture search methods, particularly the gradientbased Differentiable Architecture Search (DARTS) known for its computational efficiency. This paper presents a novel hybrid attention-enhanced differentiable neural architecture search method by integrating channel attention and spatial attention mechanisms. The proposed approach not only improves classification accuracy but also significantly reduces computational resource requirements. Experimental results demonstrate that our method achieves architecture search within 0.28 GPU-days while attaining error rates of $\mathbf{2. 5 0 \%}$ and $\mathbf{1 7. 4 1 \%}$ on CIFAR-10 and CIFAR-100 datasets respectively, outperforming both traditional neural architecture search methods and the baseline DARTS model.