A Hybrid Performance Estimation Strategy for Optimizing Neural Architecture Search

The emergence of neural architecture search (NAS) technology has lowered the professional threshold for optimizing model architectures. However, existing NAS methods primarily evaluate performance by fully training a network architecture, which is computationally expensive and slow. This paper proposes a hybrid performance estimation strategy search framework for neural architecture search, which can flexibly adjust the performance evaluation strategy at each stage. In the initial stage, this study uses less accurate but low-cost methods to quickly eliminate suboptimal architectures. As the search progresses, more computationally intensive but accurate evaluation strategies are employed to filter out the optimal network architectures. In the final stage, more precise verification is conducted to ensure that the selected network architecture achieves the best performance in practice. This research can adapt to different precision and speed requirements, providing flexible reduction space ratio strategies aimed at meeting accuracy requirements while maintaining efficiency. Its generalizability and flexibility help address various NAS challenges. Experimental results show that the method proposed in this study performs excellently in multiple benchmark tests, achieving a balance between performance and efficiency. Additionally, by testing on other search spaces, datasets, and tasks, this study demonstrates its good generalization ability.