Abstract
Feature selection is essential in multimodal learning to mitigate noise, redundancy, and the curse of dimensionality inherent in fused data sources. This study evaluates the efficacy of Genetic Algorithms (GA) and Particle Swarm Optimization (PSO) for multimodal feature selection using synthetic datasets with controlled noise and dimensionality. A Convolutional Neural Network (CNN) serves as the fitness function to guide the bio-inspired search process. Experimental results demonstrate that while bio-inspired algorithms effectively identify optimal feature subsets, their performance is highly sensitive to problem complexity and dimensionality. These findings provide insights into the scalability of metaheuristics in increasingly complex multimodal environments.