INVESTIGATING THE IMPACT OF FEATURE SELECTION METHODS ON THE PERFORMANCE OF DEEP NEURAL NETWORKS FOR IMAGE RECOGNITION TASKS

Feature selection has been a fundamental step in traditional machine learning pipelines, yet its role in deep learning remains underexplored. In this study, we investigate how different feature selection methods—both filter-based and embedded—affect the performance of deep neural networks (DNNs) on image recognition tasks. Using benchmark datasets such as CIFAR-10 and Fashion-MNIST, we apply popular feature selection techniques prior to training convolutional neural networks (CNNs). Our results show that while traditional feature selection often degrades performance when applied naively to image data, hybrid methods integrating domain knowledge and structural pruning yield notable improvements in training efficiency and generalization. This paper provides a nuanced view of how classical concepts like feature relevance intersect with the end-to-end learning paradigm of deep networks.