A Multiobjective Evolutionary Algorithm Based on Bipopulation With Uniform Sampling for Neural Architecture Search.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Neural Architecture Search (NAS) automates neural network design but faces challenges in optimizing multiple objectives like accuracy and complexity.
• Existing NAS methods often lack population diversity and adequate search space exploration, especially in complex regions.

Purpose of the Study:

• To propose a novel multiobjective evolutionary algorithm (MOEA)-BUS for NAS that simultaneously optimizes neural network accuracy and complexity.
• To enhance the efficiency and effectiveness of NAS by improving population diversity and search space coverage.

Main Methods:

• Developed MOEA-BUS, a multiobjective evolutionary algorithm utilizing a bipopulation framework and a novel uniform sampling method for population initialization.
• Implemented synergistic evolution between two populations to ensure comprehensive search space coverage.
• Conducted experiments on CIFAR-10 and ImageNet datasets to evaluate performance.

Main Results:

• MOEA-BUS achieved top-one accuracies of 98.39% on CIFAR-10 and 80.03% on ImageNet.
• Achieved 78.28% accuracy on ImageNet with a low network complexity of 446 M MAdds.
• Ablation studies confirmed that uniform sampling and bipopulation mechanisms significantly improve population diversity and overall performance.

Conclusions:

• MOEA-BUS effectively optimizes both accuracy and network complexity in NAS, outperforming existing methods.
• The proposed uniform sampling and bipopulation strategies are crucial for enhancing diversity and performance in NAS.
• The algorithm demonstrates strong potential for designing efficient and high-performing neural networks.