Towards High Performance Network Training with Noisy Label Datasets
NAVAL RESEARCH LAB WASHINGTON DC WASHINGTON United States
Pagination or Media Count:
Creating large amounts of labeled data to train neural networks is an obstacle to applying deep learning to new applications. Heuristic methodsfor labeling data typically produce a significant fraction of mislabeled samples. This report describes some methods in the literature that find thefraction of noisy labeled datasets that are probably labeled correctly and our efforts to improve on these methods. The method we describe and testis called NoisyLabel Correcting Cross Validation. The results of this method proved inferior to the INCV method in the literature but the newunderstandings learned from this effort inspired two new methods the generalized sensitivity analysis and the soft lables approaches. Our futureplans include testing these methods.