Knowledge transfer evolutionary search for lightweight neural architecture with dynamic inference

Xiaoxue Qian,Fang Liu,Licheng Jiao,Xiangrong Zhang,Xinyan Huang,Shuo Li,Puhua Chen,Xu Liu
DOI: https://doi.org/10.1016/j.patcog.2023.109790
IF: 8
2023-01-01
Pattern Recognition
Abstract:Relying on the availability of massive labeled samples, most neural architecture search (NAS) methods focus on searching large and complex models; and adopt fixed structures and parameters at the infer-ence stage. Few approaches automatically design lightweight networks for label-limited tasks and fur-ther consider the inference differences between inputs. To address these issues, we introduce evolution-ary computation (EC) and attention mechanism and propose a knowledge transfer evolutionary search for lightweight neural architecture with dynamic inference, then verify it using synthetic aperture radar (SAR) images. SAR image classification is a typical label-limited task due to the inherent imaging mecha-nism of SAR. We design the EC-based architecture search and attention-based dynamic inference for SAR image scene classification. Specifically, we build a SAR-tailored search space, explore topology pruning -based mutation operators to search lightweight architectures, and further design a dynamic Ridgelet con-volution capable of adaptive reasoning to enhance the representation ability of searched lightweight net-works. Moreover, we propose a knowledge transfer training strategy and hybrid evaluation criteria to ensure searching quickly and robustly. Experimental results show that the proposed method can search for superior neural architectures, thus improving the classification performance of SAR images.& COPY; 2023 Elsevier Ltd. All rights reserved.
What problem does this paper attempt to address?