An Efficient Evolutionary Neural Architecture Search Algorithm Without Training.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Neural Architecture Search (NAS) automates the design of high-performance neural networks.
• Existing NAS methods face challenges due to extensive time and computational resources required for performance evaluation.

Purpose of the Study:

• To propose an efficient Evolutionary Neural Architecture Search (ENAS) method.
• To address the time and computational cost challenges in NAS.
• To accelerate the convergence speed and shorten search time in NAS algorithms.

Main Methods:

• Redesigned evolutionary algorithm interactions based on biometrics principles to enhance information exchange and optimization.
• Introduced a multi-metric, training-free evaluator to assess network performance, bypassing resource-intensive training.
• Utilized NAS-Bench-101 and NAS-Bench-201 benchmarks for evaluation.

Main Results:

• The proposed ENAS method demonstrates improved local and global search capabilities.
• The multi-metric training-free evaluator effectively assesses performance and mitigates ranking offset issues.
• Identified network architectures with comparable or superior performance compared to state-of-the-art methods.

Conclusions:

• The ENAS method significantly reduces the time and computational resources needed for NAS.
• The approach offers a more efficient and effective solution for automated neural network architecture design.