Noise-Aware Dynamic Data Pruning
Year: 2026
Type: article
Abstract: Dynamic data pruning speeds up neural network training by skipping easy examples. When labels are incorrect, loss-based pruning can mistake mislabeled examples for informative ones. This report reproduces the ICLR 2024 InfoBatch method on CIFAR-10 and CIFAR-100 and evaluates it on CIFAR-100 with 20% symmetric, 40% symmetric, and 40% asymmetric synthetic label noise. AUROC measurements across training epochs showed that maximum softmax probability and energy scores had limited ability to detect mislabeled examples, especially among high-loss samples. The study evaluates an iterative filter that uses ResNet-18 features to flag likely mislabeled images and refine class centroids. With 40% symmetric label noise, adding the filter to InfoBatch raised mean test accuracy from 49.68% to 69.49% across three matched training seeds, a gain of 19.81 percentage points. In that setting, the combined method used 53.2% less measured time for training and filtering than the single full-training run. With 40% asymmetric label noise, iterative refinement improved test accuracy by 12.3 percentage points over plain centroid filtering in the matched seed-0 comparison. Full experimental code, regression tests, and reproducible figures are available at https://github.com/prithvihouse/noise-aware-data-pruning.
Author: Kurmi, Prithvi
Institution: Georgia State University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236861779
