Making Deep Neural Networks Robust to Label Noise: A Loss Correction Approach

Patrini, Giorgio; Rozza, Alessandro; Menon, Aditya Krishna; Nock, Richard; Qu, Lizhen

doi:10.1109/cvpr.2017.240

Cited by 1,161 publications

(1,388 citation statements)

References 21 publications

Supporting

Mentioning

1,309

Contrasting

Unclassified

Order By: Relevance

“…The result shows that, the performance of these models is not much affected when small parts of the dataset were not precisely labeled [54]. Aiming at an automatic labelling procedure, we identified that historical weather data are suitable indicators for identifying whether a tweet is relevant to rainfall events, as pluvial floods are directly caused by heavy rainfalls and fast storms.…”

Section: Pre-processing and Training Preparationmentioning

confidence: 95%

Extraction of Pluvial Flood Relevant Volunteered Geographic Information (VGI) by Deep Learning from User Generated Texts and Photos

Feng

Sester

2018

IJGI

View full text Add to dashboard Cite

Abstract:In recent years, pluvial floods caused by extreme rainfall events have occurred frequently. Especially in urban areas, they lead to serious damages and endanger the citizens' safety. Therefore, real-time information about such events is desirable. With the increasing popularity of social media platforms, such as Twitter or Instagram, information provided by voluntary users becomes a valuable source for emergency response. Many applications have been built for disaster detection and flood mapping using crowdsourcing. Most of the applications so far have merely used keyword filtering or classical language processing methods to identify disaster relevant documents based on user generated texts. As the reliability of social media information is often under criticism, the precision of information retrieval plays a significant role for further analyses. Thus, in this paper, high quality eyewitnesses of rainfall and flooding events are retrieved from social media by applying deep learning approaches on user generated texts and photos. Subsequently, events are detected through spatiotemporal clustering and visualized together with these high quality eyewitnesses in a web map application. Analyses and case studies are conducted during flooding events in Paris, London and Berlin.

show abstract

Section: Pre-processing and Training Preparationmentioning

confidence: 95%

Extraction of Pluvial Flood Relevant Volunteered Geographic Information (VGI) by Deep Learning from User Generated Texts and Photos

Feng

Sester

2018

IJGI

View full text Add to dashboard Cite

show abstract

“…For example, this can be represented as a noise or confusion matrix between the clean and the noisy labels, as explained in Section 3. Having its roots in statistics (Dawid and Skene, 1979), this or similar ideas have been recently studied in NLP (Fang and Cohn, 2016;Hedderich and Klakow, 2018;Paul et al, 2019), image classification (Mnih and Hinton, 2012;Sukhbaatar et al, 2015;Dgani et al, 2018) and general machine learning settings (Bekker and Goldberger, 2016;Patrini et al, 2017;Hendrycks et al, 2018). All of these methods, however, do not take the features into account that are used to represent the instances during classification.…”

Section: Related Workmentioning

confidence: 99%

Feature-Dependent Confusion Matrices for Low-Resource NER Labeling with Noisy Labels

Lange

Hedderich

Klakow

2019

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conferen

View full text Add to dashboard Cite

In low-resource settings, the performance of supervised labeling models can be improved with automatically annotated or distantly supervised data, which is cheap to create but often noisy. Previous works have shown that significant improvements can be reached by injecting information about the confusion between clean and noisy labels in this additional training data into the classifier training. However, for noise estimation, these approaches either do not take the input features (in our case word embeddings) into account, or they need to learn the noise modeling from scratch which can be difficult in a low-resource setting. We propose to cluster the training data using the input features and then compute different confusion matrices for each cluster. To the best of our knowledge, our approach is the first to leverage feature-dependent noise modeling with pre-initialized confusion matrices. We evaluate on low-resource named entity recognition settings in several languages, showing that our methods improve upon other confusion-matrix based methods by up to 9%.

show abstract

“…In general, the existing methods for handling noisy labels can be divided into four categories: noise transition estimation, noise‐robust loss design, label adjustment, and sample selection. Specifically, The methods from the first category mainly focus on estimating the noise transition matrix . In particular, Goldberger et al proposed using a noise adaptation layer to model the noise transition matrix.…”

Section: Introductionmentioning

confidence: 99%

“…Specifically, The methods from the first category mainly focus on estimating the noise transition matrix. [29][30][31][32] In particular, Goldberger et al 29 proposed using a noise adaptation layer to model the noise transition matrix. Patrini et al 30 proposed estimating the noise transition matrix through a two-step strategy.…”

Section: Introductionmentioning

confidence: 99%

Breast tumor classification through learning from noisy labeled ultrasound images

et al. 2019

View full text Add to dashboard Cite

Purpose To train deep learning models to differentiate benign and malignant breast tumors in ultrasound images, we need to collect many training samples with clear labels. In general, biopsy results can be used as benign/malignant labels. However, most clinical samples generally do not have biopsy results. Previous works have proposed generating benign/malignant labels according to Breast Imaging, Reporting and Data System (BI‐RADS) ratings. However, this approach will cause noisy labels, which means that the benign/malignant labels produced from BI‐RADS diagnoses may be inconsistent with the ground truths. Consequently, deep models will overfit the noisy labels and hence obtain poor generalization performance. In this work, we mainly focus on how to reduce the negative effect of noisy labels when they are used to train breast tumor classification models. Methods We propose an effective approach called noise filter network (NF‐Net) to address the problem of noisy labels when training breast tumor classification models. Specifically, to prevent deep models from overfitting the noisy labels, we propose incorporating two softmax layers for classification. Additionally, to strengthen the effect of clean labels, we design a teacher–student module for distilling the knowledge of clean labels. Results We conduct extensive comparisons with the existing works on addressing noisy labels. Our method achieves a classification accuracy of 73%, with a precision of 69%, recall of 80%, and F1‐score of 0.74. This result is significantly better than those of the existing state‐of‐the‐art works on addressing noisy labels. Conclusions This work provides a means to overcome the label shortage problem in training breast tumor classification models. Specifically, we can generate benign/malignant labels according to the BI‐RADS ratings. Although this approach will cause noisy labels, the design of NF‐Net can effectively reduce the negative effect of such labels.

show abstract

Making Deep Neural Networks Robust to Label Noise: A Loss Correction Approach

Cited by 1,161 publications

References 21 publications

Extraction of Pluvial Flood Relevant Volunteered Geographic Information (VGI) by Deep Learning from User Generated Texts and Photos

Extraction of Pluvial Flood Relevant Volunteered Geographic Information (VGI) by Deep Learning from User Generated Texts and Photos

Feature-Dependent Confusion Matrices for Low-Resource NER Labeling with Noisy Labels

Breast tumor classification through learning from noisy labeled ultrasound images

Contact Info

Product

Resources

About