A Pairwise Comparison Relation-assisted Multi-objective Evolutionary Neural Architecture Search Method with Multi-population Mechanism

TL;DR

This work tackles the NAS efficiency bottleneck while addressing multiple deployment metrics by introducing SMEMNAS, which combines a surrogate model based on pairwise architecture comparisons with a multi-population MOEA. The surrogate predicts architecture rankings rather than absolute performance, enabling data-efficient environment selection, while the main/vice population framework maintains diversity and accelerates convergence. validated on CIFAR-10, CIFAR-100, and ImageNet, SMEMNAS delivers competitive architectures with significantly reduced search cost (e.g., 0.17 GPU days on ImageNet in the abstract, ~4 hours in experiments) and favorable FLOPs/MAdds trade-offs. The results highlight practical gains in search efficiency and solution diversity, with ablations confirming the effectiveness of the surrogate and multi-population components and identifying avenues for future multi-objective surrogates and cross-domain transferability.

Abstract

Neural architecture search (NAS) has emerged as a powerful paradigm that enables researchers to automatically explore vast search spaces and discover efficient neural networks. However, NAS suffers from a critical bottleneck, i.e. the evaluation of numerous architectures during the search process demands substantial computing resources and time. In order to improve the efficiency of NAS, a series of methods have been proposed to reduce the evaluation time of neural architectures. However, they are not efficient enough and still only focus on the accuracy of architectures. Beyond classification accuracy, real-world applications increasingly demand more efficient and compact network architectures that balance multiple performance criteria. To address these challenges, we propose the SMEMNAS, a pairwise comparison relation-assisted multi-objective evolutionary algorithm based on a multi-population mechanism. In the SMEMNAS, a surrogate model is constructed based on pairwise comparison relations to predict the accuracy ranking of architectures, rather than the absolute accuracy. Moreover, two populations cooperate with each other in the search process, i.e. a main population that guides the evolutionary process, while a vice population that enhances search diversity. Our method aims to discover high-performance models that simultaneously optimize multiple objectives. We conduct comprehensive experiments on CIFAR-10, CIFAR-100 and ImageNet datasets to validate the effectiveness of our approach. With only a single GPU searching for 0.17 days, competitive architectures can be found by SMEMNAS which achieves 78.91% accuracy with the MAdds of 570M on the ImageNet. This work makes a significant advancement in the field of NAS.

Figures (9)

• Figure 1: Overall framework of the multi-objective algorithm based on a multi-population mechanism and the surrogate model assisting the search process (SMEMNAS). (a) Offspring are generated with a multi-population mechanism. (b) The illustration of the evaluation assisted by the surrogate model. (c) The flowchart of the proposed algorithm. A set of elite architectures are retained in each generation and are added into the archive $\mathcal{A}$, which contains architectures for evolution.
• Figure 2: A candidate architecture encoding example. The encoding is divided into five parts by blocks. The parameters we search include image resolution, the number of layers in each block, the expansion rate and the kernel size in each layer.
• Figure 3: Search space: The left part shows a complete network stacked by five blocks. The right part represents a Block and its internal structure, which consists of multiple layers.
• Figure 4: An illustration of the surrogate model based on pairwise comparison relation: the input contains the concatenation of two architectures, and the output indicates which architecture is better.
• Figure 5: Accuracy and number of multi-adds in millions on CIFAR-10. Models from multi-objective approaches are joined with lines.