Dominant Classifier-assisted Hybrid Evolutionary Multi-objective Neural Architecture Search.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Neural Architecture Search (NAS) automates deep neural network design but is computationally intensive, especially for multi-objective problems.
• Current predictor-assisted evolutionary NAS methods face challenges with slow convergence and rank disorder, impacting prediction accuracy.
• These limitations hinder the efficient discovery of optimal neural network architectures.

Purpose of the Study:

• To introduce CHENAS, a Classifier-assisted multi-objective Hybrid Evolutionary NAS framework.
• To enhance convergence speed and solution quality in multi-objective NAS.
• To address the issues of slow convergence and rank disorder in existing NAS methods.

Main Methods:

• CHENAS integrates evolutionary algorithms for global exploration and gradient-based optimization for local refinement.
• A novel dominance classifier reframes multi-objective optimization as a classification task to predict Pareto dominance relationships.
• A contrastive learning-based autoencoder creates a structured latent space for improved dominance prediction.

Main Results:

• CHENAS demonstrates superior performance compared to state-of-the-art NAS approaches on benchmark datasets.
• The framework effectively identifies high-performing architectures across multiple objectives.
• CHENAS mitigates rank disorder and improves prediction accuracy in multi-objective NAS.

Conclusions:

• CHENAS offers an efficient and effective framework for multi-objective Neural Architecture Search.
• The proposed classifier and autoencoder significantly enhance prediction accuracy and convergence.
• Future research will focus on computational efficiency and broader application domains.