Surrogate-Assisted Hybrid Multiobjective Evolutionary Neural Architecture Search

Existing Neural Architecture Search (NAS) techniques typically emphasize accuracy while neglecting the real-time inference performance. To address this challenge, we propose a surrogate-assisted adaptive multiobjective evolutionary hybrid neural architecture search algorithm that aims to simultaneously optimize accuracy and inference speed. By incorporating multi-scale hybrid convolutions and lightweight Transformer operators, we enhance the expressiveness and computational efficiency of the supernet. Additionally, we design a performance-aware hybrid sampling strategy and a hybrid offline/online surrogate model to accelerate the search process and optimize the performance of candidate architectures. Experimental results demonstrate that our method achieves an effective balance between accuracy and inference speed, showing promising potential for deployment in sensing-computing integrated chips and systems.