Ordered sample weighting: Improving neural network training for complex image classification
Yu Wang, Weipeng Wu · Discrete and Continuous Dynamical Systems - S · 2025
The growing abundance of labeled image data has propelled significant advancements in neural networks, particularly in image classification tasks, where automatic feature extraction plays a crucial role. However, the inherent imbalance in data distribution and the varying complexity of samples in real-world applications hinder consistent classification performance across different classes. This paper addresses the challenges by proposing a classification loss-weighted neural network training and optimization method. The method begins by selecting a batch of samples and ranking them based on classification loss values. Hard-to-classify samples with larger loss values are prioritized, reducing the repetitive training of easy samples. The proposed Top-q Sample Weighting (TSW) strategy estimates sample weights by maximizing the reduction in classification loss, focusing training on the most informative samples. The weighted sample subset is used to update the network parameters, improving both convergence speed and classification performance. Comparative experiments on five datasets and two network architectures validate the effectiveness of the proposed method, demonstrating significant improvements in neural network generalization and classification accuracy.