An Evolutionary Multi-Objective Surrogate-Assisted NAS Method for Higher Efficiency and Performance in Customized Data Sets

Motivation and Expectations: EC-based NAS methods are typically slow and computationally expensive. Common NAS methods search for high-performance networks in standard datasets such as ImageNet and CIFAR-100, but these developments have not yet been widely and practically applied. Therefore, we propose an NAS method for obtaining high-performance models on non-standard customized data sets, optimizing multiple competitive objectives without the huge computational burden of existing NAS methods. Main Contribution: We create two surrogate models: one at the architecture level (upper-level objective function) to improve sampling efficiency, and another at the weight level (lower-level objective function) to enhance gradient descent training efficiency through weight sharing. Our proposed method demonstrates higher efficiency and performance than current methods with equivalent or better models on standard benchmark datasets. We prove the effectiveness and generality of our method on six different non-standard customized data sets.