Architectural Optimization of Dynamic Inception Modules in Convolutional Neural Networks using the Coral Reef Optimization Algorithm

Abstract

The rapid advancement of Deep Learning (DL) has led to increasingly complex neu- ral network architectures in Artificial Intelligence (AI), often increasing computational requirements and environmental impact. This master’s thesis presents a novel methodology for optimizing the architecture of Inception modules within Convolutional Neural Networks (CNNs) that process data at multiple scales using the Coral Reef Optimization (CRO) algorithm, a bio-inspired evolutionary approach. Aligning with Green AI principles that emphasize efficiency and sustainability, we integrate a dynamic Inception module capable of adjusting branches, depths, and filter sizes with the CRO algorithm to effectively explore and exploit the architectural search space. To promote smaller, resource-efficient architectures, we introduce a custom evaluation metric that balances accuracy and model complexity by penalizing excessive parameters. Experimental results on the MNIST dataset demonstrate that the optimized models achieve competitive performance, reducing the number of parameters by up to 40% while maintaining accuracy comparable to standard models. This work contributes to the development of sustainable AI models and provides a foundation for future research in efficient neural architecture optimization.