Adaptive Dataset Pruning Based on Spatiotemporal Characteristics

Training convolutional neural networks (CNNNs) on large-scale datasets poses significant computational challenges due to redundant information inherent in numerous training samples. While dataset pruning addresses this issue by selecting representative subsets to reduce training costs, conventional approaches relying on static sample evaluations suffer from two critical limitations: (1) failure to capture the dynamic evolution of sample importance throughout the training trajectory, and (2) sensitivity to predefined pruning configurations. To overcome these shortcomings, we propose an adaptive dataset pruning framework that continuously monitors and updates sample significance during model optimization. Our methodology integrates four pivotal components: (1) a similarity prediction network that quantifies inter-class and intra-class relationships, (2) gradient-based analysis for tracking temporal variations in sample contributions, (3) dynamic selection of samples exhibiting consistent importance across training epochs, and (4) a retention mechanism for transiently critical samples. Extensive experiments on CIFAR-10 and CIFAR-100 benchmarks demonstrate that our method achieves superior performance compared to state-of-the-art baselines, particularly at moderate pruning ratios (e.g., 1.2 % accuracy improvement on CIFAR-10 with 50 % data retention). Notably, the proposed approach maintains competitive accuracy even under aggressive pruning scenarios <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 0 - 9 0 \%}$</tex> data removal), demonstrating enhanced robustness to pruning intensity variations. This work provides a principled framework for efficient deep learning through dynamic data valuation.