Self-Filtering: A Noise-Aware Sample Selection for Label Noise with Confidence Penalization

Wei, Qi; Sun, Haoliang; Lu, Xiankai; Yin, Yilong

doi:10.1007/978-3-031-20056-4_30

Cited by 26 publications

(11 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Previous literatures leverage the small-loss criterion that selects the examples with small empirical loss as the clean one [7,42]. Recently, the works [1,24,43] represented by SELF [24] pay more attention to history prediction results, providing selection with more information and thus promoting the selection results. Besides, sample reweight methods [29,31] give examples with different weights, which can be regarded as a special form of sample selection.…”

Section: Related Workmentioning

confidence: 99%

“…Forward [27] 69.84 SFT+ [43] 75.08 JoCoR [42] 70.30 CE 64.54 Joint Optim [35] 72. 23 CE + SNSCL 73.49 SL [41] 71.02 DivideMix [14] 74.76 ELR+ [18] 74.81 DivideMix + SNSCL 75.31…”

Section: Comparison With State-of-the-artsmentioning

confidence: 99%

“…Learning from noisy labels [2,7,14,18,20,23,33,43,46] poses great challenges for training deep models, whose performance heavily relies on large-scaled labeled datasets. Annotating training data with high confidence would be resource-intensive, especially for some domains, such as medical and remote sensing images [16,34].…”

Section: Introductionmentioning

confidence: 99%

“…Previous methods [1,2,7,14,20,31,42] in LNL always focus on generic classification (e.g. CIFAR-10 & 100) and artificially construct random label noise [18,20,35,36] and dependent label noise [7,14,31,42,43] to evaluate the performance of their algorithms. In this work, we extend LNL to fine-grained classification, which is a rarely studied task.…”

Section: Introductionmentioning

confidence: 99%

“…The number of total training epochs is both 100, and the learning rate is decayed with the factor 10 by 20 and 40 epoch. For Clothing-1M, refers to [43], we train the classifier network for 15 epochs and use SGD with 0.9 momentum, weight decay of 5e-4. The learning rate is set as 0.002 and decayed with the factor of 10 after 10 epochs, while warm up stage is one epoch.…”

mentioning

confidence: 99%

See 4 more Smart Citations

Fine-Grained Classification with Noisy Labels

Qi¹,

Liu²,

Sun³

et al. 2023

Preprint

View full text Add to dashboard Cite

Learning with noisy labels (LNL) aims to ensure model generalization given a label-corrupted training set. In this work, we investigate a rarely studied scenario of LNL on fine-grained datasets (LNL-FG), which is more practical and challenging as large inter-class ambiguities among fine-grained classes cause more noisy labels. We empirically show that existing methods that work well for LNL fail to achieve satisfying performance for LNL-FG, arising the practical need of effective solutions for LNL-FG. To this end, we propose a novel framework called stochastic noise-tolerated supervised contrastive learning (SNSCL) that confronts label noise by encouraging distinguishable representation. Specifically, we design a noise-tolerated supervised contrastive learning loss that incorporates a weight-aware mechanism for noisy label correction and selectively updating momentum queue lists. By this mechanism, we mitigate the effects of noisy anchors and avoid inserting noisy labels into the momentum-updated queue. Besides, to avoid manually-defined augmentation strategies in contrastive learning, we propose an efficient stochastic module that samples feature embeddings from a generated distribution, which can also enhance the representation ability of deep models. SNSCL is general and compatible with prevailing robust LNL strategies to improve their performance for LNL-FG. Extensive experiments demonstrate the effectiveness of SNSCL.

show abstract

Section: Related Workmentioning

confidence: 99%

“…Forward [27] 69.84 SFT+ [43] 75.08 JoCoR [42] 70.30 CE 64.54 Joint Optim [35] 72. 23 CE + SNSCL 73.49 SL [41] 71.02 DivideMix [14] 74.76 ELR+ [18] 74.81 DivideMix + SNSCL 75.31…”

Section: Comparison With State-of-the-artsmentioning

confidence: 99%