Efficient Evolutionary Neural Architecture Search Using Reliable Fitness Evaluations for Image Classification

Neural Architecture Search (NAS) has gained considerable attention for automating the design of Deep Neural Networks (DNNs) and reducing human effort. However, most NAS algorithms require substantial computational resources. One-shot NAS mitigates this by training a large supernet, allowing candidate architectures to inherit weights, but these methods remain computationally intensive and may lead to unreliable performance evaluations. To address these challenges, we propose an efficient one-shot NAS method with reliable fitness evaluations. Our approach employs a novel weight initialization strategy tailored to the supernet's architecture, significantly reducing training time. During the evolutionary search process, the supernet is fine-tuned based on the candidate architectures being evaluated, ensuring more accurate performance estimations. Furthermore, a robust population initialization method is introduced for the Genetic Algorithm (GA), leveraging insights from supernet training to guide the search process effectively. Experimental results demonstrate that our method achieves error rates of 2.59%, 16.67%, and 23.9% on CIFAR10, CIFAR100, and ImageNet, respectively, while requiring only 0.31, 0.28, and 4.1 GPU-days. Compared with those competitors, our approach offers a favorable trade-off between performance and computational cost. Additional analyses confirm the robustness and efficiency of the proposed supernet training, population initialization, and fine-tuning techniques.