Learning with Noisy Labels by Adaptive Gradient-Based Outlier Removal

An accurate and substantial dataset is necessary to train a reliable and well-performing model. However, even manually labeled datasets contain errors, not to mention automatically labeled ones. The problem of data denoising was addressed in different existing research, most of which focuses on the detection of outliers and their permanent removal - a process that is likely to over- or underfilter the dataset. In this work, we propose AGRA: a new method for Adaptive GRAdient-based outlier removal. Instead of cleaning the dataset prior to model training, the dataset is adjusted during the training process. By comparing the aggregated gradient of a batch of samples and an individual example gradient, our method dynamically decides whether a corresponding example is helpful for the model at this point or is counter-productive and should be left out for the current update. Extensive evaluation on several datasets demonstrates the AGRA effectiveness, while comprehensive results analysis supports our initial hypothesis: permanent hard outlier removal is not always what model benefits the most from.