psoResNet: An improved PSO-based residual network search algorithm.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Manual design of deep convolutional neural networks (DCNNs) is time-consuming and costly.
• Existing Neural Architecture Search (NAS) methods face challenges like limited architecture design, long search times, and inefficient search space utilization.

Purpose of the Study:

• To propose an optimized strategy for residual network architecture search using an enhanced Particle Swarm Optimization (PSO) algorithm.
• To address limitations of current NAS methods by improving search efficiency and network performance.

Main Methods:

• Employed low-complexity residual architecture blocks as foundational units for diverse architecture exploration with minimal parameters.
• Implemented a depth initialization strategy to effectively confine the search space.
• Introduced novel particle difference computation and velocity update mechanisms to enhance trajectory exploration and particle diversity.

Main Results:

• The proposed method significantly improved search space utilization and particle diversity.
• Developed lightweight DCNNs with enhanced classification performance.
• Validated effectiveness on benchmark datasets and a custom 13-class crime dataset.

Conclusions:

• The enhanced PSO-based NAS strategy effectively designs lightweight DCNNs with superior classification accuracy.
• The approach overcomes key limitations of existing NAS methods, offering a more efficient and effective solution for DCNN architecture optimization.