A Novel Approach for Feature Optimization in Deep Learning Models

Chathurangi Kumari Atugoda, Subha Fernando
dx.doi.org/10.17576/apjitm-2026-1501-16

Abstract

Deep Neural Networks (DNNs) are widely used in image recognition and classification tasks, where performance strongly depends on effective optimization. Conventional gradient-based optimizers such as Stochastic Gradient Descent (SGD), Adam, and AdaDelta are effective, but they may suffer from slow convergence, sensitivity to initialization, and the risk of getting trapped in local minima, particularly in deep architectures. To address these limitations, this study proposes an Enhanced Particle Swarm Optimization integrated Convolutional Neural Network (EPSO-CNN). The novelty of the proposed method lies in embedding a PSO-operated layer after the group convolution stage to optimize intermediate feature maps before classification. The proposed model was evaluated on MNIST, CIFAR-10, and CIFAR-100 datasets and compared with a baseline LeNet-5 model trained using SGD, Adam, and AdaDelta. Experimental results show that the proposed EPSO-CNN achieves improved classification accuracy, faster convergence, and better training stability, especially on more complex datasets. These findings demonstrate that PSO-based feature map optimization can serve as an effective complementary strategy to conventional gradient-based learning in CNNs.

keyword

Deep neural network; Feature optimization; Particle swarm optimization; Swarm intelligence; Weight values

Area

Data Mining and Optimization