CAM-NAS: A Fast Model for Neural Architecture Search based on Class Activation Map
Zhiyuan Zhang, Inwhee Joe · Research Square · 2023
Abstract Artificial intelligence has developed rapidly in recent years. However, with the increasing complexity of AI models and the realistic requirements of AI multi-platform deployment, it is difficult for experts to design a specific network model structure for a specific platform. Neural architecture search (NAS) can be used to help experts discover new network structures. However, traditional NAS algorithms need to consume a lot of computing resources and computing time. In this paper, we propose a class activation map-based neural architecture search method (CAM-NAS) by avoiding the step of training sub-models from traditional NAS algorithms, thus greatly speeding up the search efficiency. On a single NVIDIA RTX3090 graphics card, CAM-NAS takes only 0.08 seconds to evaluate a single sub-model. This is orders of magnitude faster than traditional NAS algorithms. We conducted experiments on the cifar-10 and cifar-100 datasets and achieved 95.68\% and 75.94\% accuracy, respectively. CAM-NAS is the first algorithm to combine the interpretability method with NAS, which greatly inspires further exploration of the inner principles of deep learning models in the future.