Optimizing the Architecture of Convolutional Neural Networks Using Modified Salp Swarm Algorithm

Abstract

Deep learning is highly effective in dealing with complex tasks such as image classification and recognition. However, finding the optimal architecture's hyperparameters for Convolutional Neural Networks (CNNs) to achieve the best performance and parameter regularization can be challenging. Metaheuristic optimization algorithms can be utilized to find solutions in this context. In this research, a computerized CNN was adjusted using an improved Salp Swarm Algorithm (SSA) to enhance crucial CNN settings, like dropout rate, hidden units, learning rate, and batch size. The refined design was tested on two standard datasets. MNIST and Fashion MNIST. The outcomes displayed model performance achieving accuracy levels of 99.6% for MNIST and 94.08% for Fashion MNIST. This tuned system outperformed the existing practices by 0.2% and 0.04% for each dataset while also cutting down on computational expenses. The fusion of SSA with CNNs displayed adaptability and resilience opening up possibilities, in image classification and consistently delivering outstanding outcomes.