Efficient Network Architecture Search Using Hybrid Optimizer.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Manual design of Convolutional Neural Networks (CNNs) for image classification is time-consuming and resource-intensive.
• The increasing demand for efficient Neural Architecture Search (NAS) methods necessitates innovative solutions.
• Existing NAS methods often struggle with computational cost and design complexity.

Purpose of the Study:

• To propose a novel deep architecture generation model for CNNs.
• To enhance the efficiency and accuracy of the CNN architecture search process.
• To combine evolutionary computing algorithms with CNN design for automated structure generation.

Main Methods:

• A new encoding strategy for CNN structures to integrate with evolutionary algorithms.
• A location update mechanism incorporating genetic algorithm (GA) operators for optimized search.
• Handling variable-length CNN structures using skip connections.
• Integrating traditional CNN layers, residual blocks, and a grouping strategy for flexible architecture exploration.

Main Results:

• The proposed model demonstrates good performance in terms of search accuracy.
• The method achieves favorable results in reducing the time required for CNN architecture search.
• Experimental validation on MNIST and CIFAR-10 datasets confirms the model's effectiveness.

Conclusions:

• The novel deep architecture generation model effectively addresses the limitations of manual CNN design.
• The integration of Aquila Optimization (AO) and GA offers a promising approach for efficient NAS.
• The proposed method provides a flexible and powerful tool for discovering optimal CNN architectures.