Learning with Neighbor Consistency for Noisy Labels

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 thoroughly evaluate our method on datasets evaluating both synthetic (CIFAR-10, CIFAR-100) and realistic (mini-WebVision, WebVision, Clothing1M, mini-ImageNet-Red) noise, and achieve competitive or state-of-the-art accuracies across all of them.

PDF Abstract CVPR 2022 PDF CVPR 2022 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Benchmark
Image Classification mini WebVision 1.0 NCR+Mixup+DA (ResNet-50) Top-1 Accuracy 80.5 # 7
Image Classification mini WebVision 1.0 NCR+Mixup (ResNet-50) Top-1 Accuracy 79.4 # 14
Image Classification mini WebVision 1.0 NCR (ResNet-50) Top-1 Accuracy 77.1 # 30
Learning with noisy labels Red MiniImageNet 20% label noise NCR (ResNet-18) Test Accuracy 69.0 # 1
Image Classification Red MiniImageNet 20% label noise NCR (ResNet-18) Accuracy 69.0 # 1
Learning with noisy labels Red MiniImageNet 40% label noise NCR (ResNet-18) Test Accuracy 64.6 # 1
Image Classification Red MiniImageNet 40% label noise NCR (ResNet-18) Accuracy 64.6 # 1
Learning with noisy labels Red MiniImageNet 80% label noise NCR (ResNet-18) Test Accuracy 51.2 # 1
Image Classification Red MiniImageNet 80% label noise NCR (ResNet-18) Accuracy 51.2 # 1
Image Classification WebVision-1000 NCR (ResNet-50) Top-1 Accuracy 75.7% # 8
Image Classification WebVision-1000 NCR+Mixup+DA (ResNet-50) Top-1 Accuracy 76.8 # 4

Methods


No methods listed for this paper. Add relevant methods here