Identifying the Training Stop Point with Noisy Labeled Data

Finding an early stopping point at maximum obtainable test accuracy (MOTA) is a challenging problem when training deep neural networks (DNNs) with noisy labeled data. Recent studies assume either that i) a clean validation set is available or ii) the noise ratio is known, or, both. However, often a clean validation set is unavailable, and the noise estimation can be inaccurate. To overcome these issues, we provide a novel training solution, free of these conditions. We analyze the rate of change of the training accuracy under different conditions to identify a training stop region. We further develop a heuristic algorithm (AutoTSP) to find a training stop point (TSP) at or close to MOTA. We validate the robustness of AutoTSP through several experiments on various datasets, noise ratios and architectures.