Optimizing convolutional neural networks architecture using a modified particle swarm optimization for image classification

Although Convolutional Neural Networks (CNNs) have been shown to be highly effective in image classification tasks, designing their architecture to achieve optimal results is often challenging. This process is time consuming, requires significant effort and expertise, and is complicated by the large number of hyperparameters. To address this problem, in this work we propose an approach that reduces human intervention and automatically generates the best CNN design. Our approach uses a variant of Particle Swarm Optimization (PSO), called Particle Swarm Optimization without Velocity (PSWV), to speed up convergence and reduce the number of iterations required to determine the optimal CNN hyperparameters. We developed a novel strategy to determine the updated position of each particle using a linear combination of the best position of the particle and the best position of the swarm without relying on the velocity equation. Our algorithm harnesses the power of the variable-length encoding strategy to represent particles within the population, thereby providing swift convergence towards the best architecture. We evaluate our proposed algorithm against several recent algorithms in the literature by using nine benchmark datasets for classification tasks and comparing it to 27 other algorithms, including state-of-the-art ones. Our experimental results show that our proposed method, pswvCNN, is able to quickly find effective CNN architectures that provide comparable performance to the best currently available designs, indicating its significant potential.