Training deep learning models on large-scale datasets is often constrained by computational cost, time, and energy requirements. Coreset selection offers a promising solution by constructing compact and representative data subsets that preserve the full dataset’s essential characteristics. This paper proposes a novel coreset selection method based on class-wise Neural Principal Component Analysis (NPCA). The method identifies dominant components for each class and selects samples that most strongly express these components. We evaluate the approach on a subset of the ImageNet dataset and on the CIFAR-10 dataset and compare it against established coreset selection methods, Herding, and K-Center. Experimental results demonstrate that the NPCA-based method performs competitively, especially in low-data regimes, while offering improved interpretability through visual analysis of selected samples. Our findings suggest that the proposed method is a viable and efficient strategy for training under resource constraints.