Proceedings of the
The Nineteenth International Conference on Computational Intelligence and Security (CIS 2023)
December 1 – 4, 2023, Haikou, China
A Rapid Network Structure Analysis Method for Low-cost Multi-objective Neural Architecture Search
School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, China.
ABSTRACT
Evolutionary algorithms have attracted great attention in neural architecture search (NAS) community, especially for solving multi-objective NAS problems, since the multi-objective evolutionary algorithm (MOEA) can automatically
search for appropriate network architectures for a given task with conflicting objectives. The relationship between micro-structures and final network performance is an important part in MOEA-based NAS algorithms. However, most of them suffer from
the traditional traversal-based network analysis techniques since they are too timeconsuming to afford. To overcome this barrier, a method named Text Process Network Analysis (TextNet) is proposed in this paper. TextNet
encodes the candidate network individuals in different network architecture levels of the population as binary strings, and uses a fast hierarchical string matching algorithm to analyze the network architecture to generate a dictionary of relationships
between different network module architectures and overall network performance. This method significantly reduces the time and computational complexity of network analysis compared to traditional methods, and provides a more rich search
information for existing MOEA-based NAS methods, in order to improve their search efficiency and performance. Using the representative MOEAs on the tailored multi-objective NAS test suites for image classification, the TextNet shows
significant improvement effects on existing MOEA-based NAS methods.
Keywords: Neural architecture search, Multi-objective evolutionary algorithm, Network structure analysis, Image classification.
Download PDF
School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, China.
ABSTRACT
Evolutionary algorithms have attracted great attention in neural architecture search (NAS) community, especially for solving multi-objective NAS problems, since the multi-objective evolutionary algorithm (MOEA) can automatically search for appropriate network architectures for a given task with conflicting objectives. The relationship between micro-structures and final network performance is an important part in MOEA-based NAS algorithms. However, most of them suffer from the traditional traversal-based network analysis techniques since they are too timeconsuming to afford. To overcome this barrier, a method named Text Process Network Analysis (TextNet) is proposed in this paper. TextNet encodes the candidate network individuals in different network architecture levels of the population as binary strings, and uses a fast hierarchical string matching algorithm to analyze the network architecture to generate a dictionary of relationships between different network module architectures and overall network performance. This method significantly reduces the time and computational complexity of network analysis compared to traditional methods, and provides a more rich search information for existing MOEA-based NAS methods, in order to improve their search efficiency and performance. Using the representative MOEAs on the tailored multi-objective NAS test suites for image classification, the TextNet shows significant improvement effects on existing MOEA-based NAS methods.
Keywords: Neural architecture search, Multi-objective evolutionary algorithm, Network structure analysis, Image classification.
Download PDF