ACD-U: Asymmetric co-teaching with machine unlearning for robust learning with noisy labels

Dimensions overall score 8.0

• ACD-U utilizes asymmetric co-teaching with different model architectures, specifically a CLIP-pretrained vision Transformer and a CNN, to mitigate confirmation bias.method 70%
  Evidencepartial

  First, its asymmetric co-teaching pairs a contrastive language-image pretraining (CLIP)-pretrained vision Transformer with a convolutional neural network (CNN), leveraging their complementary learning behaviors: the pretrained model provides stable predictions, whereas the CNN adapts throughout training. This asymmetry, where the vision Transformer is trained only on clean samples and the CNN is trained through SSL, effectively mitigates confirmation bias.

  Implicationpartial

  This is a core mechanism of the proposed method, explicitly described in the abstract.

  Verificationpartial

  partial

• ACD-U incorporates selective unlearning for post-hoc error correction by identifying and removing the influence of incorrectly memorized samples.method 70%
  Evidencepartial

  Second, selective unlearning enables post-hoc error correction by identifying incorrectly memorized samples through loss trajectory analysis and CLIP consistency checks, and then removing their influence via Kullback--Leibler divergence-based forgetting.

  Implicationpartial

  This is the second core mechanism of the proposed method, explicitly described in the abstract.

  Verificationpartial

  partial

• ACD-U achieves state-of-the-art performance on various noisy datasets, including CIFAR-10/100, CIFAR-N, WebVision, Clothing1M, and Red Mini-ImageNet.result 70%
  Evidencepartial

  Experiments on synthetic and real-world noisy datasets, including CIFAR-10/100, CIFAR-N, WebVision, Clothing1M, and Red Mini-ImageNet, demonstrate state-of-the-art performance, particularly in high-noise regimes and under instance-dependent noise.

  Implicationpartial

  The abstract explicitly states state-of-the-art performance on these datasets.

  Verificationpartial

  partial

• ACD-U demonstrates superior performance in high-noise regimes and under instance-dependent noise.result 70%
  Evidencepartial

  Experiments on synthetic and real-world noisy datasets, including CIFAR-10/100, CIFAR-N, WebVision, Clothing1M, and Red Mini-ImageNet, demonstrate state-of-the-art performance, particularly in high-noise regimes and under instance-dependent noise.

  Implicationpartial

  The abstract specifically highlights these conditions as areas where the method excels.

  Verificationpartial

  partial

• The proposed method shifts the learning paradigm from passive error avoidance to active error correction.technical 70%
  Evidencepartial

  This approach shifts the learning paradigm from passive error avoidance to active error correction.

  Implicationpartial

  This is a conceptual claim about the impact of the method, directly stated in the abstract.

  Verificationpartial

  partial

• The code for ACD-U is publicly available.technical 70%
  Evidencepartial

  The code is publicly available at https://github.com/meruemon/ACD-U.

  Implicationpartial

  The abstract provides a direct link to the public code repository.

  Verificationpartial

  partial

• Traditional semi-supervised learning methods combined with sample selection cannot correct selection errors once a sample is misclassified.limitation 70%
  Evidencepartial

  Although recent methods have combined sample selection with semi-supervised learning (SSL) to exploit the memorization effect -- where networks learn from clean data before noisy data -- they cannot correct selection errors once a sample is misclassified.

  Implicationpartial

  This is presented as a limitation of existing methods that ACD-U aims to overcome.

  Verificationpartial

  partial

Competitive landscape

ACD-U is a robust noisy label learning framework that leverages asymmetric co-teaching and machine unlearning to achieve state-of-the-art performance, offering a potential solution for improving model accuracy in real-world datasets with noisy labels.