Efficient Training of ResNet-50 on Large Scale Datasets Using Enhanced Particle Swarm Optimization

Deep Neural Networks (DNNs) such as ResNet-50 have achieved state-of-the-art results in large-scale image classification. However, their training performance depends strongly on optimization efficiency. Conventional methods like Stochastic Gradient Descent (SGD) often struggle with slow convergence, learning-rate sensitivity, and local minima entrapment. To overcome these challenges, this study introduces an Enhanced Particle Swarm Optimization (PSO) technique that adaptively adjusts particle dynamics to maintain a better exploration-exploitation balance during training. The method was evaluated on the CIFAR-10 dataset and compared against Standard PSO and SGD under identical experimental conditions. The results demonstrate that Enhanced PSO achieves superior validation accuracy (up to 95.7%), faster convergence, and a more stable weight distribution centered near zero. These characteristics reflect a well-regularized learning process with improved generalization. Overall, the Enhanced PSO framework provides a robust and scalable optimization approach for deep neural networks, offering a viable alternative to conventional gradient-based training algorithms.