Automatic Design of Convolutional Neural Network Architectures Under Resource Constraints.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• High-accuracy Convolutional Neural Network (CNN) models face deployment challenges on resource-constrained smart electronics.
• Existing lightweight networks and Neural Architecture Search (NAS) methods have limitations in optimizing for multiple objectives simultaneously.

Purpose of the Study:

• To propose an automatic method for designing CNN architectures that meet predefined resource constraints.
• To address the trade-off between model size and accuracy in CNN design for edge devices.

Main Methods:

• Developed an automatic CNN architecture design method incorporating constraint handling.
• Utilized an adaptive penalty algorithm for fitness evaluation and a selective repair operation for infeasible individuals.
• Set model complexity (number of parameters) as a resource constraint for optimization.

Main Results:

• Demonstrated the effectiveness of the proposed method on CIFAR-10 and CIFAR-100 datasets.
• The automatically designed CNN models satisfied predefined resource constraints while achieving superior accuracy compared to manual designs and NSGA-Net.
• Successfully addressed the challenge of uncertain optimal CNN model size under random search.

Conclusions:

• The proposed method effectively designs CNN architectures that balance accuracy and resource efficiency for smart electronics.
• This approach offers a viable solution for deploying advanced AI models on edge devices with limited computational power.