Wasserstein Adversarial Regularization for Learning With Label Noise

Fatras, Kilian; Damodaran, Bharath Bhushan; Lobry, Sylvain; Flamary, Rémi; Tuia, Devis; Courty, Nicolas

doi:10.1109/tpami.2021.3094662

Cited by 19 publications

(18 citation statements)

References 34 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…R KL forces the local Lipschitz constant to be small with respect to the KL divergence. The adversarial regularization was first designed to solve semi-supervised learning problem (Miyato et al, 2019), but it has been proven to be efficient against label noise as the classifier learns over an interpolation of the label and the prediction (Fatras et al, 2021a). The adversarial direction is approximated using the power iteration algorithm (Golub & van der Vorst, 2000).…”

Section: Proposed Methodsmentioning

confidence: 99%

“…Our classifier trained with adversarial regularization is able to detect the outliers in the measure α c . The hyperparameters used in the adversarial regularization algorithms are similar to one used in (Miyato et al, 2019;Fatras et al, 2021a) at the notable exception of η which is set to 10 like in the GAN experiment.…”

Section: B22 Architectures and Training Detailsmentioning

confidence: 99%

“…In this scenario, second type outliers can be seen as noisy label samples and we show that classifiers overfit over them. Thus, we want our classifier to be robust against noisy labels and to do this, we use the virtual adversarial training algorithm (Miyato et al, 2019) which was shown to be robust against noisy labels (Fatras et al, 2021a). When the classification of a sample does not match its assigned label, we consider the sample as an outlier.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

On making optimal transport robust to all outliers

Fatras¹

2022

Preprint

Self Cite

View full text Add to dashboard Cite

Optimal transport (OT) is known to be sensitive against outliers because of its marginal constraints. Outlier robust OT variants have been proposed based on the definition that outliers are samples which are expensive to move. In this paper, we show that this definition is restricted by considering the case where outliers are closer to the target measure than clean samples. We show that outlier robust OT fully transports these outliers leading to poor performances in practice. To tackle these outliers, we propose to detect them by relying on a classifier trained with adversarial training to classify source and target samples. A sample is then considered as an outlier if the prediction from the classifier is different from its assigned label. To decrease the influence of these outliers in the transport problem, we propose to either remove them from the problem or to increase the cost of moving them by using the classifier prediction. We show that we successfully detect these outliers and that they do not influence the transport problem on several experiments such as gradient flows, generative models and label propagation.Preprint. Under review.

show abstract

Section: Proposed Methodsmentioning

confidence: 99%

Section: B22 Architectures and Training Detailsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

On making optimal transport robust to all outliers

Fatras¹

2022

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…When a sample has noise or label error [22], the feature vector it maps on the highdimensional feature space will be far away from the feature group of the label class, which we call abnormal feature points. Under the interference of noise samples, the above method may be less effective.…”

Section: Anomaly Feature Detection Strategymentioning

confidence: 99%

Efficient Data-Driven Crop Pest Identification Based on Edge Distance-Entropy for Sustainable Agriculture

Yang

et al. 2022

Sustainability

View full text Add to dashboard Cite

Human agricultural activities are always accompanied by pests and diseases, which have brought great losses to the production of crops. Intelligent algorithms based on deep learning have achieved some achievements in the field of pest control, but relying on a large amount of data to drive consumes a lot of resources, which is not conducive to the sustainable development of smart agriculture. The research in this paper starts with data, and is committed to finding efficient data, solving the data dilemma, and helping sustainable agricultural development. Starting from the data, this paper proposed an Edge Distance-Entropy data evaluation method, which can be used to obtain efficient crop pests, and the data consumption is reduced by 5% to 15% compared with the existing methods. The experimental results demonstrate that this method can obtain efficient crop pest data, and only use about 60% of the data to achieve 100% effect. Compared with other data evaluation methods, the method proposed in this paper achieve state-of-the-art results. The work conducted in this paper solves the dilemma of the existing intelligent algorithms for pest control relying on a large amount of data, and has important practical significance for realizing the sustainable development of modern smart agriculture.

show abstract

“…This choice is key when using OT as a loss in learning [67] and inference [68]. Henceforth we describe the mini-batch framework of [69], for using OT as a loss.…”

Section: Mini-batch Optimal Transportmentioning

confidence: 99%