Learning with Neighbor Consistency for Noisy Labels

Ahmet Iscen, Jack Valmadre, Anurag Arnab, Cordelia Schmid

Abstract: Recent advances in deep learning have relied on large, labelled datasets to train high-capacity models. However, collecting large datasets in a time- and cost-efficient manner often results in label noise. We present a method for learning from noisy labels that leverages similarities between training examples in feature space, encouraging the prediction of each example to be similar to its nearest neighbours. Compared to training algorithms that use multiple models or distinct stages, our approach takes the form of a simple, additional regularization term. It can be interpreted as an inductive version of the classical, transductive label propagation algorithm. We compare our approach to relevant baselines under both synthetic and realistic noise, and demonstrate that our simple approach achieves state-of-the-art accuracy under the realistic conditions of mini-ImageNet-Red, mini-WebVision and Clothing1M.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2202.02200/code)