Evolutionary bi-level neural architecture search with training: A framework for color classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Designing optimal Artificial Neural Network (ANN) architectures for classification is complex, particularly under resource constraints.
• The large number of design parameters in ANNs poses a significant challenge for efficient architecture definition.
• Existing methods often struggle to balance network complexity with predictive performance.

Purpose of the Study:

• To propose an Evolutionary Bi-Level Neural Architecture Search with Training (EB-LNAST) approach.
• To simultaneously optimize ANN architecture, weights, and biases using a bi-level strategy.
• To demonstrate EB-LNAST's effectiveness in generating compact and high-performing classification models.

Main Methods:

• Employs a bi-level optimization strategy where the upper level minimizes network complexity and the lower level optimizes training parameters.
• The lower level focuses on minimizing loss and maximizing predictive performance.
• Evaluated on real-world color classification and the WDBC dataset.

Main Results:

• EB-LNAST shows statistically significant improvements over traditional and advanced machine learning algorithms.
• Achieves superior predictive performance compared to fixed-architecture Multilayer Perceptrons (MLPs).
• Demonstrates up to a [Formula: see text] reduction in model size, creating more efficient architectures.

Conclusions:

• EB-LNAST is a reliable method for generating compact and effective neural network architectures.
• It enables efficient search space exploration while maintaining or exceeding state-of-the-art classification performance.
• Offers a viable alternative for classification tasks with resource constraints.