Multi-objective CNN optimization: A robust framework for automated model design

Convolutional neural networks (CNNs) have demonstrated high performance in classifying image data. However, CNNs require expert-level hyperparameter tuning and involve substantial computational complexity, which hinders their effective deployment in real-time and IoT systems. Research on CNNs indicates that optimal hyperparameter configurations can enhance inference speed while improving classification accuracy. We developed a novel approach that uses multi-objective optimization to design CNN models automatically. Our method tunes hyperparameters to balance classification accuracy and inference speed. We define classification performance and inference speed as objectives and balance them using a Pareto-optimal strategy. Unlike traditional approaches, MoCNN systematically explores Pareto-optimal trade-offs between classification performance and computational efficiency, enabling a fully automated architecture search without manual intervention. In this study, we select NSGA-II as our preferred MOEA while ensuring the framework remains flexible enough to accommodate other evolutionary strategies. Experimental evaluations on benchmark datasets (CIFAR-10, CIFAR-100, and FRUITS-360) demonstrate that MoCNN reduces inference time by up to 72.02% on average and improves classification accuracy by 6.72% compared to manually tuned CNN architectures. By eliminating the need for heuristic hyperparameter selection, MoCNN enhances scalability and is particularly well suited for real-time, mobile AI, and edge-computing applications. Our results show that MoCNN outperforms state-of-the-art optimization frameworks in both computational efficiency and predictive performance, highlighting its potential for deployment in scenarios where accuracy and speed are critical.