Bayesian DivideMix++ for enhanced learning with noisy labels
Leveraging inexpensive and human intervention-based annotating methodologies, such as crowdsourcing and web crawling, often leads to datasets with noisy labels. Noisy labels can have a detrimental impact on the performance and generalization of deep neural networks. Robust models that are able to ha...
| Authors: | ; ; ; |
|---|---|
| Format: | Article |
| Status: | Published version |
| Publication date: | 2024 |
| Country: | España |
| Institution: | Universitat Pompeu Fabra |
| Repository: | Repositorio Digital de la UPF |
| OAI Identifier: | oai:repositori.upf.edu:10230/71905 |
| Online Access: | http://hdl.handle.net/10230/71905 http://dx.doi.org/10.1016/j.neunet.2024.106122 |
| Access Level: | Open access |
| Keyword: | Learning with noisy labels Neural network memorization Data augmentation Self-supervised pre-training Label uncertainty Monte-Carlo dropouts |
| _version_ | 1878738976352763905 |
|---|---|
| author | Nagarajan, Bhalaji |
| author2 | Marques, Ricardo Aguilar, Eduardo J. Radeva, Petia |
| author2_role | author author author |
| author_browse | Aguilar, Eduardo J. Marques, Ricardo Nagarajan, Bhalaji Radeva, Petia |
| author_facet | Nagarajan, Bhalaji Marques, Ricardo Aguilar, Eduardo J. Radeva, Petia |
| author_role | author |
| collection | Repositorio Digital de la UPF |
| description | Leveraging inexpensive and human intervention-based annotating methodologies, such as crowdsourcing and web crawling, often leads to datasets with noisy labels. Noisy labels can have a detrimental impact on the performance and generalization of deep neural networks. Robust models that are able to handle and mitigate the effect of these noisy labels are thus essential. In this work, we explore the open challenges of neural network memorization and uncertainty in creating robust learning algorithms with noisy labels. To overcome them, we propose a novel framework called “Bayesian DivideMix++” with two critical components: (i) DivideMix++, to enhance the robustness against memorization and (ii) Monte-Carlo MixMatch, which focuses on improving the effectiveness towards label uncertainty. DivideMix++ improves the pipeline by integrating the warm-up and augmentation pipeline with self-supervised pre-training and dedicated different data augmentations for loss analysis and backpropagation. Monte-Carlo MixMatch leverages uncertainty measurements to mitigate the influence of uncertain samples by reducing their weight in the data augmentation MixMatch step. We validate our proposed pipeline using four datasets encompassing various synthetic and real-world noise settings. We demonstrate the effectiveness and merits of our proposed pipeline using extensive experiments. Bayesian DivideMix++ outperforms the state-of-the-art models by considerable differences in all experiments. Our findings underscore the potential of leveraging these modifications to enhance the performance and generalization of deep neural networks in practical scenarios. |
| eu_rights_str_mv | openAccess |
| format | article |
| id | ES_b8bdf4d7c2fbf1f8e5a2cbd1a2dca026 |
| institution | Universitat Pompeu Fabra |
| instname_str | Universitat Pompeu Fabra |
| network_acronym_str | ES |
| network_name_str | España |
| oai_identifier_str | oai:repositori.upf.edu:10230/71905 |
| publishDate | 2024 |
| publishDateSort | 2024 |
| publisher | Elsevier |
| publisherStr | Elsevier |
| reponame_str | Repositorio Digital de la UPF |
| spelling | Bayesian DivideMix++ for enhanced learning with noisy labels Nagarajan, Bhalaji Marques, Ricardo Aguilar, Eduardo J. Radeva, Petia Learning with noisy labels Neural network memorization Data augmentation Self-supervised pre-training Label uncertainty Monte-Carlo dropouts Leveraging inexpensive and human intervention-based annotating methodologies, such as crowdsourcing and web crawling, often leads to datasets with noisy labels. Noisy labels can have a detrimental impact on the performance and generalization of deep neural networks. Robust models that are able to handle and mitigate the effect of these noisy labels are thus essential. In this work, we explore the open challenges of neural network memorization and uncertainty in creating robust learning algorithms with noisy labels. To overcome them, we propose a novel framework called “Bayesian DivideMix++” with two critical components: (i) DivideMix++, to enhance the robustness against memorization and (ii) Monte-Carlo MixMatch, which focuses on improving the effectiveness towards label uncertainty. DivideMix++ improves the pipeline by integrating the warm-up and augmentation pipeline with self-supervised pre-training and dedicated different data augmentations for loss analysis and backpropagation. Monte-Carlo MixMatch leverages uncertainty measurements to mitigate the influence of uncertain samples by reducing their weight in the data augmentation MixMatch step. We validate our proposed pipeline using four datasets encompassing various synthetic and real-world noise settings. We demonstrate the effectiveness and merits of our proposed pipeline using extensive experiments. Bayesian DivideMix++ outperforms the state-of-the-art models by considerable differences in all experiments. Our findings underscore the potential of leveraging these modifications to enhance the performance and generalization of deep neural networks in practical scenarios. This work was partially funded by the Horizon EU project MUSAE (No. 01070421), 2021-SGR-01094 (AGAUR), Icrea Academia’2022 (Generalitat de Catalunya), Robo STEAM (2022-1-BG01-KA220-VET-000089434, Erasmus＋ EU), DeepSense (ACE053/22/000029, ACCIÓ), DeepFoodVol (AEI-MICINN, PDC2022-133642-I00), IDEATE (AEI-MICINN, PID2022-141566NB-I00), A-BMC (AEI-MICINN, CNS2022-135480), CERCA Programme/Generalitat de Catalunya, and Agencia Nacional de Investigación y Desarrollo de Chile (ANID) (Grant No. FONDECYT INICIACIÓN 11230262). Ricardo Marques acknowledges the support of the Serra Húnter Programme. B. Nagarajan acknowledges the support of FPI Becas, MICINN, Spain . The authors thankfully acknowledge the computer resources at FinisTerrae III and the technical support provided by the Galician Supercomputing Center (CESGA) (RES-IM-2023-2-0025). Elsevier http://hdl.handle.net/10230/71905 http://dx.doi.org/10.1016/j.neunet.2024.106122 |
| spellingShingle | Bayesian DivideMix++ for enhanced learning with noisy labels Nagarajan, Bhalaji Learning with noisy labels Neural network memorization Data augmentation Self-supervised pre-training Label uncertainty Monte-Carlo dropouts |
| status_str | publishedVersion |
| title | Bayesian DivideMix++ for enhanced learning with noisy labels |
| title_full | Bayesian DivideMix++ for enhanced learning with noisy labels |
| title_fullStr | Bayesian DivideMix++ for enhanced learning with noisy labels |
| title_full_unstemmed | Bayesian DivideMix++ for enhanced learning with noisy labels |
| title_short | Bayesian DivideMix++ for enhanced learning with noisy labels |
| title_sort | Bayesian DivideMix++ for enhanced learning with noisy labels |
| topic | Learning with noisy labels Neural network memorization Data augmentation Self-supervised pre-training Label uncertainty Monte-Carlo dropouts |
| topic_facet | Learning with noisy labels Neural network memorization Data augmentation Self-supervised pre-training Label uncertainty Monte-Carlo dropouts |
| url | http://hdl.handle.net/10230/71905 http://dx.doi.org/10.1016/j.neunet.2024.106122 |