Efficient network architecture search via multiobjective particle swarm optimization based on decomposition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Manual design of convolutional neural networks (CNNs) is time-consuming and requires significant expertise.
• Existing neural architecture search (NAS) methods often neglect network complexity, optimizing solely for prediction error.
• There is a growing need for automated NAS methods that consider both accuracy and efficiency.

Purpose of the Study:

• To propose MOPSO/D-Net, a CNN architecture search method using multiobjective particle swarm optimization based on decomposition (MOPSO/D).
• To address the limitations of current NAS approaches by formulating the problem as a multiobjective optimization task.
• To minimize both classification error rate and the number of network parameters.

Main Methods:

• Reformulating NAS as a multiobjective evolutionary optimization problem.
• Employing an improved MOPSO/D algorithm with hybrid binary encoding and adaptive penalty-based boundary intersection.
• Solving the multiobjective NAS problem to provide diverse trade-off solutions between error rate and parameter count.

Main Results:

• MOPSO/D-Net demonstrates effectiveness compared to manual and automated CNN generation methods.
• Achieved 0.4% error rate with 0.16M parameters on the MNIST dataset.
• Achieved 5.88% error rate with 8.1M parameters on the CIFAR-10 dataset, showcasing strong classification performance with reduced complexity.

Conclusions:

• MOPSO/D-Net successfully balances classification accuracy and network size in CNN architecture search.
• The proposed method offers a viable solution for developing efficient and high-performing CNNs.
• Experimental results validate the effectiveness of MOPSO/D-Net for automated CNN design.