Sample Prior Guided Robust Model Learning to Suppress Noisy Labels

2 Dec 2021  ยท  Wenkai Chen, Chuang Zhu, Yi Chen, Mengting Li, Tiejun Huang ยท

Imperfect labels are ubiquitous in real-world datasets and seriously harm the model performance. Several recent effective methods for handling noisy labels have two key steps: 1) dividing samples into cleanly labeled and wrongly labeled sets by training loss, 2) using semi-supervised methods to generate pseudo-labels for samples in the wrongly labeled set. However, current methods always hurt the informative hard samples due to the similar loss distribution between the hard samples and the noisy ones. In this paper, we proposed PGDF (Prior Guided Denoising Framework), a novel framework to learn a deep model to suppress noise by generating the samples' prior knowledge, which is integrated into both dividing samples step and semi-supervised step. Our framework can save more informative hard clean samples into the cleanly labeled set. Besides, our framework also promotes the quality of pseudo-labels during the semi-supervised step by suppressing the noise in the current pseudo-labels generating scheme. To further enhance the hard samples, we reweight the samples in the cleanly labeled set during training. We evaluated our method using synthetic datasets based on CIFAR-10 and CIFAR-100, as well as on the real-world datasets WebVision and Clothing1M. The results demonstrate substantial improvements over state-of-the-art methods.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Learning with noisy labels CIFAR-100N PGDF Accuracy (mean) 74.08 # 1
Learning with noisy labels CIFAR-10N-Aggregate PGDF Accuracy (mean) 96.11 # 2
Learning with noisy labels CIFAR-10N-Random1 PGDF Accuracy (mean) 96.01 # 2
Learning with noisy labels CIFAR-10N-Worst PGDF Accuracy (mean) 93.65 # 2
Image Classification CIFAR-10 (with noisy labels) PGDF (ResNet-18) Accuracy (under 20% Sym. label noise) 96.7% # 3
Accuracy (under 50% Sym. label noise) 96.3% # 1
Accuracy (under 80% Sym. label noise) 94.7% # 2
Accuracy (under 90% Sym. label noise) 84.0% # 4
Image Classification Clothing1M PGDF Accuracy 75.19% # 8
Image Classification mini WebVision 1.0 PGDF (Inception-ResNet-v2) Top-1 Accuracy 81.47 # 3
Top-5 Accuracy 94.03 # 2
ImageNet Top-1 Accuracy 75.45 # 16
ImageNet Top-5 Accuracy 93.11 # 6

Methods


No methods listed for this paper. Add relevant methods here