Joint Negative and Positive Learning for Noisy Labelsharmonylab
紹介論文
Joint Negative and Positive Learning for Noisy Labels
Youngdong Kim Juseung Yun Hyounguk Shon Junmo KimSchool of Electrical Engineering, KAIST, South Korea
概要:Noisy Labelsに対する従来手法のNLNLを改善したJNPLを提案した.新たな損失関数NL+とPL+を用いた単一の学習アルゴリズムを用いることで単純化し学習コストの削減と精度向上を狙い,SOTAを達成した.
Joint Negative and Positive Learning for Noisy Labelsharmonylab
紹介論文
Joint Negative and Positive Learning for Noisy Labels
Youngdong Kim Juseung Yun Hyounguk Shon Junmo KimSchool of Electrical Engineering, KAIST, South Korea
概要:Noisy Labelsに対する従来手法のNLNLを改善したJNPLを提案した.新たな損失関数NL+とPL+を用いた単一の学習アルゴリズムを用いることで単純化し学習コストの削減と精度向上を狙い,SOTAを達成した.
ArcFace: Additive Angular Margin Loss for Deep Face Recognitionharmonylab
出典: Jiankang Deng, Jia Guo, Niannan Xue, Stefanos Zafeiriou : ArcFace: Additive Angular Margin Loss for Deep Face Recognition, Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019)
公開URL:https://arxiv.org/abs/1801.07698
概要 : 顔認識のための畳み込みニューラルネットワーク(DCNN)の課題は識別力を高める適切な損失関数を設計することです。本論文では、顔認識のための識別性の高い特徴量を得るために、Additive Angular Margin Loss (ArcFace)を提案します。一般的な顔認識ベンチマークから1兆ペアの大規模データセットなどを用いて、最先端顔認識技術との比較実験を行いました。結果は、従来手法を凌駕する精度を持つことが明らかになりました。
This document summarizes a research paper that proposes a technique called Stable Rank Normalization to improve generalization in neural networks and GANs. The technique aims to reduce the Lipschitz constant of neural networks by normalizing the stable rank of the weight matrices. The stable rank is a measure of how many effective dimensions a matrix has. Normalizing it makes networks less sensitive to certain parameter settings. The paper shows experimentally that stable rank normalization improves generalization on image recognition and GAN tasks without affecting performance.
The document discusses the paper "t-vMF Similarity for Regularizing Intra-Class Feature Distribution" presented at CVPR2021. The paper proposes a new similarity measure called t-vMF similarity that can control the width of the peak and skirt of the cosine similarity. This allows intra-class variance to be reduced while preventing gradient vanishing, especially for imbalanced or small-scale datasets where maximizing discrimination is more important than minimizing intra-class variance. The t-vMF similarity is implemented by considering the von Mises-Fisher distribution in the process of the softmax cross-entropy loss, making it simple to implement.
Distribution Aligning Refinery of Pseudo-label for Imbalanced Semi-supervised...ALINLAB
Official slides for the NeurIPS 2020 paper "Distribution Aligning Refinery of Pseudo-label for Imbalanced Semi-supervised Learning," by Jaehyung Kim, Youngbum Hur, Sejun Park, Eunho Yang, Sung Ju Hwang, and Jinwoo Shin.
Machine translation (MT) is one of the earliest and most successful applications of natural language processing. Many MT services have been deployed via web and smartphone apps, enabling communication and information access across the globe by bypassing language barriers. However, MT is not yet a solved problem. MT services that cover the most languages cover only about a hundred; thousands more are currently unsupported. Even for the currently supported languages, the translation quality is far from perfect.
A key obstacle in our way to achieving usable MT models for any language is data imbalance. On the one hand, machine learning techniques perform subpar on rare categories, having only a few to no training examples. On the other hand, natural language datasets are inevitably imbalanced with a long tail of rare types. The rare types carry more information content, and hence correctly translating them is crucial. In addition to the rare word types, rare phenomena also manifest in other forms as rare languages and rare linguistic styles.
Our contributions towards advancing rare phenomena learning in MT are four-fold: (1) We show that MT models have much in common with classification models, especially regarding the data imbalance and frequency-based biases. We describe a way to reduce the imbalance severity during the model training. (2) We show that the currently used automatic evaluation metrics overlook the importance of rare words. We describe an interpretable evaluation metric that treats important words as important. (3) We propose methods to evaluate and improve translation robustness to rare linguistic styles such as partial translations and language alternations in inputs. (4) Lastly, we present a set of tools intended to advance MT research across a wider range of languages. Using these tools, we demonstrate 600 languages to English translation, thus supporting 500 more rare languages currently unsupported by others.
ArcFace: Additive Angular Margin Loss for Deep Face Recognitionharmonylab
出典: Jiankang Deng, Jia Guo, Niannan Xue, Stefanos Zafeiriou : ArcFace: Additive Angular Margin Loss for Deep Face Recognition, Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019)
公開URL:https://arxiv.org/abs/1801.07698
概要 : 顔認識のための畳み込みニューラルネットワーク(DCNN)の課題は識別力を高める適切な損失関数を設計することです。本論文では、顔認識のための識別性の高い特徴量を得るために、Additive Angular Margin Loss (ArcFace)を提案します。一般的な顔認識ベンチマークから1兆ペアの大規模データセットなどを用いて、最先端顔認識技術との比較実験を行いました。結果は、従来手法を凌駕する精度を持つことが明らかになりました。
This document summarizes a research paper that proposes a technique called Stable Rank Normalization to improve generalization in neural networks and GANs. The technique aims to reduce the Lipschitz constant of neural networks by normalizing the stable rank of the weight matrices. The stable rank is a measure of how many effective dimensions a matrix has. Normalizing it makes networks less sensitive to certain parameter settings. The paper shows experimentally that stable rank normalization improves generalization on image recognition and GAN tasks without affecting performance.
The document discusses the paper "t-vMF Similarity for Regularizing Intra-Class Feature Distribution" presented at CVPR2021. The paper proposes a new similarity measure called t-vMF similarity that can control the width of the peak and skirt of the cosine similarity. This allows intra-class variance to be reduced while preventing gradient vanishing, especially for imbalanced or small-scale datasets where maximizing discrimination is more important than minimizing intra-class variance. The t-vMF similarity is implemented by considering the von Mises-Fisher distribution in the process of the softmax cross-entropy loss, making it simple to implement.
Distribution Aligning Refinery of Pseudo-label for Imbalanced Semi-supervised...ALINLAB
Official slides for the NeurIPS 2020 paper "Distribution Aligning Refinery of Pseudo-label for Imbalanced Semi-supervised Learning," by Jaehyung Kim, Youngbum Hur, Sejun Park, Eunho Yang, Sung Ju Hwang, and Jinwoo Shin.
Machine translation (MT) is one of the earliest and most successful applications of natural language processing. Many MT services have been deployed via web and smartphone apps, enabling communication and information access across the globe by bypassing language barriers. However, MT is not yet a solved problem. MT services that cover the most languages cover only about a hundred; thousands more are currently unsupported. Even for the currently supported languages, the translation quality is far from perfect.
A key obstacle in our way to achieving usable MT models for any language is data imbalance. On the one hand, machine learning techniques perform subpar on rare categories, having only a few to no training examples. On the other hand, natural language datasets are inevitably imbalanced with a long tail of rare types. The rare types carry more information content, and hence correctly translating them is crucial. In addition to the rare word types, rare phenomena also manifest in other forms as rare languages and rare linguistic styles.
Our contributions towards advancing rare phenomena learning in MT are four-fold: (1) We show that MT models have much in common with classification models, especially regarding the data imbalance and frequency-based biases. We describe a way to reduce the imbalance severity during the model training. (2) We show that the currently used automatic evaluation metrics overlook the importance of rare words. We describe an interpretable evaluation metric that treats important words as important. (3) We propose methods to evaluate and improve translation robustness to rare linguistic styles such as partial translations and language alternations in inputs. (4) Lastly, we present a set of tools intended to advance MT research across a wider range of languages. Using these tools, we demonstrate 600 languages to English translation, thus supporting 500 more rare languages currently unsupported by others.
Grammatical Error Correction with Improved Real-world ApplicabilityMasato Mita
The document discusses grammatical error correction with improved real-world applicability. It discusses three main issues with current grammatical error correction:
[1] Evaluation - Current single corpus evaluations are not reliable and do not reflect real-world scenarios with varying proficiency levels.
[2] Data noise - There is noise in existing grammatical error correction training data that can negatively impact model performance.
[3] Low resource - Current approaches require large amounts of data and model parameters, which has low real-world applicability.
The document proposes approaches to address these issues, including cross-sectional evaluation using multiple test sets, a self-refinement strategy to reduce data noise, and analyzing grammatical generalization
Neural Semi-supervised Learning under Domain ShiftSebastian Ruder
This document discusses neural semi-supervised learning under domain shift. It presents three research areas:
1) Learning across domains by selecting relevant source domain data for transfer learning using Bayesian optimization. Experimental results on sentiment analysis, POS tagging, and dependency parsing show this approach outperforms baselines.
2) Revisiting classic semi-supervised learning techniques like self-training, tri-training, and comparing them to recent advances. Experiments on sentiment analysis and POS tagging find tri-training works best.
3) The possibility of leveraging pre-trained language models for semi-supervised learning when the target task differs from the source task.
Learning to Balance: Bayesian Meta-Learning for Imbalanced and Out-of-distrib...MLAI2
While tasks could come with varying the number of instances and classes in realistic settings, the existing meta-learning approaches for few-shot classification assume that number of instances per task and class is fixed. Due to such restriction, they learn to equally utilize the meta-knowledge across all the tasks, even when the number of instances per task and class largely varies. Moreover, they do not consider distributional difference in unseen tasks, on which the meta-knowledge may have less usefulness depending on the task relatedness. To overcome these limitations, we propose a novel meta-learning model that adaptively balances the effect of the meta-learning and task-specific learning within each task. Through the learning of the balancing variables, we can decide whether to obtain a solution by relying on the meta-knowledge or task-specific learning. We formulate this objective into a Bayesian inference framework and tackle it using variational inference. We validate our Bayesian Task-Adaptive Meta-Learning (Bayesian TAML) on two realistic task- and class-imbalanced datasets, on which it significantly outperforms existing meta-learning approaches. Further ablation study confirms the effectiveness of each balancing component and the Bayesian learning framework.
This document introduces a method called "co-curricular learning" that dynamically combines clean-data selection and domain-data selection for neural machine translation. It applies an EM-style optimization procedure to refine the "co-curriculum." Experimental results on two domains demonstrate the effectiveness of the method and properties of the data scheduled by the co-curriculum.
Présentation d'une communication acceptée dans Iceri2019Thouraya Daouas
This document discusses using machine learning to predict students' skill levels based on an online collaborative learning course. It describes the course, which focused on remote group work and using ICT tools. Student data was collected through questionnaires and used to train a random forest model in Microsoft Azure Machine Learning. The model aimed to classify students as having acquired or not acquired teamwork skills based on their course experiences and evaluations. Evaluation results showed the model could accurately predict students' skill levels with over 80% accuracy, demonstrating the potential of using machine learning to help improve students' skills.
Wharton People Analytics Conference 2017Janmejay Dave
Presented university scoring using Multi Criteria Decision Analysis (MCDA) & clustering to improve efficiency of Teach for America recruitment efforts.
Research on multi-class imbalance from a number of researchers faces
obstacles in the form of poor data diversity and a large number of classifiers.
The Hybrid Approach Redefinition-Multiclass Imbalance (HAR-MI) method
is a Hybrid Ensembles method which is the development of the Hybrid
Approach Redefinion (HAR) method. This study has compared the results
obtained with the Dynamic Ensemble Selection-Multiclass Imbalance
(DES-MI) method in handling multiclass imbalance. In the HAR-MI
Method, the preprocessing stage was carried out using the random balance
ensembles method and dynamic ensemble selection to produce a candidate
ensemble and the processing stages was carried out using different
contribution sampling and dynamic ensemble selection to produce
a candidate ensemble. This research has been conducted by using multi-class
imbalance datasets sourced from the KEEL Repository. The results show that
the HAR-MI method can overcome multi-class imbalance with better data
diversity, smaller number of classifiers, and better classifier performance
compared to a DES-MI method. These results were tested with a Wilcoxon
signed-rank statistical test which showed that the superiority of the HAR-MI
method with respect to DES-MI method.
The document summarizes recent work in natural language generation (NLG), including common training and evaluation practices as well as efforts to address limitations. It discusses how teacher forcing can lead to exposure bias during inference and explores alternatives like reinforcement learning and generative adversarial networks. It also reviews work on multilingual datasets and metrics as well as efforts to develop more accurate evaluation methods for NLG like question-based metrics and SAFEval. The document concludes by discussing promising directions for future work such as leveraging discriminators during training and generating questions to evaluate NLG models.
[CVPR 22] Context-rich Minority Oversampling for Long-tailed ClassificationSeulki Park
The document proposes a new method called Context-rich Minority Oversampling (CMO) to address the problem of long-tailed classification. CMO leverages the rich context of majority class samples as backgrounds to generate new augmented minority class samples. This requires little computational cost compared to previous methods. Experiments on long-tailed benchmarks show CMO achieves state-of-the-art performance, outperforming other oversampling baselines. Analysis demonstrates the effectiveness of using different distributions for the background and foreground samples.
The document provides an overview of the TESTA (Transforming the Experience of Students Through Assessment) programme approach. It outlines the rationale for taking a whole-programme approach to assessment, highlighting issues with modular degrees and how they can lead to student alienation. It then describes the different research tools used in the TESTA approach, including programme audits, student questionnaires, and focus groups, and provides examples of how the data from these tools can be analyzed and discussed with programme teams. The goal is to help teams improve their assessment practices in a way that enhances student learning and experience.
This document summarizes several papers on semi-supervised learning methods published between 2017-2019. It describes techniques such as consistency regularization, which encourages consistent predictions from unlabeled data after augmentation; entropy minimization, which reduces the entropy of label distributions; and adversarial training, which adds random perturbations to make models smooth. Papers discussed include MixMatch, Mean Teachers, Virtual Adversarial Training, and S4L.
CVPR2022 paper reading - Balanced multimodal learning - All Japan Computer Vi...Antonio Tejero de Pablos
Introduction of the CVPR2022 paper: Balanced multimodal learning via on-the-fly gradient modulation @ The All Japan Computer Vision Study Group (2022/08/07)
Interpreting complex machine learning models can be difficult. Given an interpretation, its meaningfulness and reliability are hard to evaluate. Even more, depending on the purpose (debugging, ...), a technique in the literature may be more appropriate than others. How to choose the best approach in the landscape of the existing techniques?
This talk is organized as a virtual "walk" through different techniques for interpreting machine learning, and particularly deep learning. Moving from the inside out, we will first cover techniques (such as gradient ascent and deconvolution) for interpreting the internal state of the model, namely its neurons, channels and layer activations. We will then focus on the model behavior from the outside. The model output, for instance, can be explained by attributing the final decisions to subsets of input pixels (as in saliency, occlusion and class activation maps) or to higher-level concepts, such as object size, scale and texture. Concept-based attribution, in particular, has been our research focus over the last years, allowing us to explain deep learning in simple terms to clinicians. For this, digital pathology and retinopathy were our main application domains. In addition, concept-based interpretability helped us explain internal CNN mechanisms such as the encoding of scale and memorization of input-label pairs.
Similar to M2m: Imbalanced Classification via Major-to-minor Translation (CVPR 2020) (15)
Trajectory-wise Multiple Choice Learning for Dynamics Generalization in Reinf...ALINLAB
This document proposes a method called Trajectory-wise Multiple Choice Learning (MCL) to improve generalization in model-based reinforcement learning. The method uses a multi-headed dynamics model to approximate the multi-modal distribution of transition dynamics. Trajectory-wise MCL updates the prediction head that is most accurate over an entire trajectory segment, allowing each head to specialize. An adaptive planning method then uses the most accurate head based on recent experience. Evaluation shows the approach achieves superior generalization to new environments compared to baseline methods.
Polynomial Tensor Sketch for Element-wise Matrix Function (ICML 2020)ALINLAB
1) The document proposes a polynomial tensor sketch method to approximate element-wise matrix functions in linear time. It combines tensor sketching, which can approximate matrix monomials fast, with polynomial approximation of the target function.
2) Coreset-based regression is used to efficiently compute optimal polynomial coefficients by selecting a small subset of rows.
3) Experiments show the method outperforms alternatives like random Fourier features for applications like kernel approximation, kernel SVM, and Sinkhorn algorithm, providing speedups of up to 49x.
Context-aware Dynamics Model for Generalization in Model-Based Reinforcement ...ALINLAB
This document summarizes a research paper on developing a context-aware dynamics model (CaDM) to improve generalization in model-based reinforcement learning. The CaDM uses a context encoder to separately learn context from past observations and condition a dynamics model on the learned context vector. This allows the model to generalize better to new environments. The CaDM achieves improved generalization over vanilla dynamics models in simulation experiments, and can also help model-free RL methods generalize better by conditioning policies on the context vector.
Self-supervised Label Augmentation via Input Transformations (ICML 2020)ALINLAB
This document discusses a new method called Self-supervised Label Augmentation (SLA) for utilizing self-supervised learning in fully-supervised settings. SLA constructs joint labels combining original labels and self-supervised labels from input transformations. It trains a single joint-label classifier, eliminating the need to learn invariance to transformations. This outperforms baselines that separately optimize original and self-supervised tasks. At test time, SLA aggregates predictions across transformations using conditional probabilities, requiring only a single inference pass. Experiments show SLA improves performance over baselines in standard, few-shot and imbalanced classification settings.
For the full video of this presentation, please visit: https://www.edge-ai-vision.com/2024/06/building-and-scaling-ai-applications-with-the-nx-ai-manager-a-presentation-from-network-optix/
Robin van Emden, Senior Director of Data Science at Network Optix, presents the “Building and Scaling AI Applications with the Nx AI Manager,” tutorial at the May 2024 Embedded Vision Summit.
In this presentation, van Emden covers the basics of scaling edge AI solutions using the Nx tool kit. He emphasizes the process of developing AI models and deploying them globally. He also showcases the conversion of AI models and the creation of effective edge AI pipelines, with a focus on pre-processing, model conversion, selecting the appropriate inference engine for the target hardware and post-processing.
van Emden shows how Nx can simplify the developer’s life and facilitate a rapid transition from concept to production-ready applications.He provides valuable insights into developing scalable and efficient edge AI solutions, with a strong focus on practical implementation.
AI-Powered Food Delivery Transforming App Development in Saudi Arabia.pdfTechgropse Pvt.Ltd.
In this blog post, we'll delve into the intersection of AI and app development in Saudi Arabia, focusing on the food delivery sector. We'll explore how AI is revolutionizing the way Saudi consumers order food, how restaurants manage their operations, and how delivery partners navigate the bustling streets of cities like Riyadh, Jeddah, and Dammam. Through real-world case studies, we'll showcase how leading Saudi food delivery apps are leveraging AI to redefine convenience, personalization, and efficiency.
Programming Foundation Models with DSPy - Meetup SlidesZilliz
Prompting language models is hard, while programming language models is easy. In this talk, I will discuss the state-of-the-art framework DSPy for programming foundation models with its powerful optimizers and runtime constraint system.
Unlock the Future of Search with MongoDB Atlas_ Vector Search Unleashed.pdfMalak Abu Hammad
Discover how MongoDB Atlas and vector search technology can revolutionize your application's search capabilities. This comprehensive presentation covers:
* What is Vector Search?
* Importance and benefits of vector search
* Practical use cases across various industries
* Step-by-step implementation guide
* Live demos with code snippets
* Enhancing LLM capabilities with vector search
* Best practices and optimization strategies
Perfect for developers, AI enthusiasts, and tech leaders. Learn how to leverage MongoDB Atlas to deliver highly relevant, context-aware search results, transforming your data retrieval process. Stay ahead in tech innovation and maximize the potential of your applications.
#MongoDB #VectorSearch #AI #SemanticSearch #TechInnovation #DataScience #LLM #MachineLearning #SearchTechnology
Taking AI to the Next Level in Manufacturing.pdfssuserfac0301
Read Taking AI to the Next Level in Manufacturing to gain insights on AI adoption in the manufacturing industry, such as:
1. How quickly AI is being implemented in manufacturing.
2. Which barriers stand in the way of AI adoption.
3. How data quality and governance form the backbone of AI.
4. Organizational processes and structures that may inhibit effective AI adoption.
6. Ideas and approaches to help build your organization's AI strategy.
“An Outlook of the Ongoing and Future Relationship between Blockchain Technologies and Process-aware Information Systems.” Invited talk at the joint workshop on Blockchain for Information Systems (BC4IS) and Blockchain for Trusted Data Sharing (B4TDS), co-located with with the 36th International Conference on Advanced Information Systems Engineering (CAiSE), 3 June 2024, Limassol, Cyprus.
Cosa hanno in comune un mattoncino Lego e la backdoor XZ?Speck&Tech
ABSTRACT: A prima vista, un mattoncino Lego e la backdoor XZ potrebbero avere in comune il fatto di essere entrambi blocchi di costruzione, o dipendenze di progetti creativi e software. La realtà è che un mattoncino Lego e il caso della backdoor XZ hanno molto di più di tutto ciò in comune.
Partecipate alla presentazione per immergervi in una storia di interoperabilità, standard e formati aperti, per poi discutere del ruolo importante che i contributori hanno in una comunità open source sostenibile.
BIO: Sostenitrice del software libero e dei formati standard e aperti. È stata un membro attivo dei progetti Fedora e openSUSE e ha co-fondato l'Associazione LibreItalia dove è stata coinvolta in diversi eventi, migrazioni e formazione relativi a LibreOffice. In precedenza ha lavorato a migrazioni e corsi di formazione su LibreOffice per diverse amministrazioni pubbliche e privati. Da gennaio 2020 lavora in SUSE come Software Release Engineer per Uyuni e SUSE Manager e quando non segue la sua passione per i computer e per Geeko coltiva la sua curiosità per l'astronomia (da cui deriva il suo nickname deneb_alpha).
Building Production Ready Search Pipelines with Spark and MilvusZilliz
Spark is the widely used ETL tool for processing, indexing and ingesting data to serving stack for search. Milvus is the production-ready open-source vector database. In this talk we will show how to use Spark to process unstructured data to extract vector representations, and push the vectors to Milvus vector database for search serving.
Let's Integrate MuleSoft RPA, COMPOSER, APM with AWS IDP along with Slackshyamraj55
Discover the seamless integration of RPA (Robotic Process Automation), COMPOSER, and APM with AWS IDP enhanced with Slack notifications. Explore how these technologies converge to streamline workflows, optimize performance, and ensure secure access, all while leveraging the power of AWS IDP and real-time communication via Slack notifications.
In his public lecture, Christian Timmerer provides insights into the fascinating history of video streaming, starting from its humble beginnings before YouTube to the groundbreaking technologies that now dominate platforms like Netflix and ORF ON. Timmerer also presents provocative contributions of his own that have significantly influenced the industry. He concludes by looking at future challenges and invites the audience to join in a discussion.
Best 20 SEO Techniques To Improve Website Visibility In SERPPixlogix Infotech
Boost your website's visibility with proven SEO techniques! Our latest blog dives into essential strategies to enhance your online presence, increase traffic, and rank higher on search engines. From keyword optimization to quality content creation, learn how to make your site stand out in the crowded digital landscape. Discover actionable tips and expert insights to elevate your SEO game.
Climate Impact of Software Testing at Nordic Testing DaysKari Kakkonen
My slides at Nordic Testing Days 6.6.2024
Climate impact / sustainability of software testing discussed on the talk. ICT and testing must carry their part of global responsibility to help with the climat warming. We can minimize the carbon footprint but we can also have a carbon handprint, a positive impact on the climate. Quality characteristics can be added with sustainability, and then measured continuously. Test environments can be used less, and in smaller scale and on demand. Test techniques can be used in optimizing or minimizing number of tests. Test automation can be used to speed up testing.
Ocean lotus Threat actors project by John Sitima 2024 (1).pptxSitimaJohn
Ocean Lotus cyber threat actors represent a sophisticated, persistent, and politically motivated group that poses a significant risk to organizations and individuals in the Southeast Asian region. Their continuous evolution and adaptability underscore the need for robust cybersecurity measures and international cooperation to identify and mitigate the threats posed by such advanced persistent threat groups.
Your One-Stop Shop for Python Success: Top 10 US Python Development Providersakankshawande
Simplify your search for a reliable Python development partner! This list presents the top 10 trusted US providers offering comprehensive Python development services, ensuring your project's success from conception to completion.
Things to Consider When Choosing a Website Developer for your Website | FODUUFODUU
Choosing the right website developer is crucial for your business. This article covers essential factors to consider, including experience, portfolio, technical skills, communication, pricing, reputation & reviews, cost and budget considerations and post-launch support. Make an informed decision to ensure your website meets your business goals.
Things to Consider When Choosing a Website Developer for your Website | FODUU
M2m: Imbalanced Classification via Major-to-minor Translation (CVPR 2020)
1. M2m: Imbalanced Classification via
Major-to-minor Translation
Jaehyung Kim* Jongheon Jeong* Jinwoo Shin
*Equal contribution
Korea Advanced Institute of Science and Technology (KAIST)
2. • Many real-world datasets have imbalanced class distributions
• Standard training (e.g. ERM) often fails to generalize at the “tail” classes
[Wang et al. 2017; Cui et al. 2019; Cao et al. 2019]
Class Imbalance in Training Data
Species [Van Horn et al. 2019] Places [Wang et al. 2017]
[Wang et al. 2017] Learning to Model the Tail. NeurIPS, 2017
[Van Horn et al. 2019] The iNaturalist Species Classification and Detection Dataset. In CVPR, 2018
[Cao et al. 2019] Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss. In NeurIPS 2019
[Cui et al. 2019] Class-balanced Loss Based on Effective Number of Samples. In CVPR, 2019
3. • Many real-world datasets have imbalanced class distributions
• Standard training (e.g. ERM) often fails to generalize at the “tail” classes
[Wang et al. 2017; Cui et al. 2019; Cao et al. 2019]
• Several training strategies have been investigated
Class Imbalance in Training Data
Re-balancing methods
• Re-sampling
[Japkowicz et al. 2000; Chawla et al. 2002]
• Re-weighting
[Khan et al. 2017; Cui et al. 2019]
Regularization methods
• Margin-based method
[Dong et al. 2017; Cao et al. 2019]
• Minority focused loss
[Lin et al. 2017]
Fundamental problem: Limited information of minority classes
[Japkowicz et al. 2000] The Class Imbalance Problem: Significance and Strategies. In ICAI 2000
[Chawla et al. 2002] SMOTE: Synthetic Minority Oversampling Technique. In JAIR 2002
[Khan et al. 2017] Cost-sensitive Learning of Deep Feature Representations from Imbalanced Data. In TNNLS 2017
[Dong et al. 2018] Imbalanced Deep Learning by Minority Class Incremental Rectification., TPAMI 2018
[Lin et al. 2017] Focal Loss for Dense Object Detection., CVPR 2017
4. • M2m augments minority classes using information of majority samples
• Over-fitting on minority classes is prevented by utilizing the diversity of majority
• A simple optimization using a pre-trained classifier surprisingly works well
M2m: Major-to-minor Translation
5. • M2m augments minority classes via translating from majority samples
• Each translation is done by solving the following optimization:
M2m: Major-to-minor Translation
Optimization objective of M2m
(a) Translation to minority class
- Using a pre-trained classifier
(b) Regularization for reducing a risk as majority class
- On the logit of training classifier
is used as a minority sample for training f<latexit sha1_base64="dC5uJX4s+XLUhctJUEqp/+YNQhw=">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</latexit><latexit sha1_base64="dC5uJX4s+XLUhctJUEqp/+YNQhw=">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</latexit><latexit sha1_base64="dC5uJX4s+XLUhctJUEqp/+YNQhw=">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</latexit><latexit sha1_base64="dC5uJX4s+XLUhctJUEqp/+YNQhw=">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</latexit>
x⇤
<latexit sha1_base64="xiLUe47ZeHY4KDuiiyZk9HZurI4=">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</latexit><latexit sha1_base64="xiLUe47ZeHY4KDuiiyZk9HZurI4=">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</latexit><latexit sha1_base64="xiLUe47ZeHY4KDuiiyZk9HZurI4=">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</latexit><latexit sha1_base64="xiLUe47ZeHY4KDuiiyZk9HZurI4=">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</latexit>
k<latexit sha1_base64="OSVpthl7bZLrnH3UTqOtdanyPNQ=">AAAB6HicdVDLSsNAFL3xWeur6tLNYBFchaSmD3cFNy5bsA9oQ5lMp+3YySTMTIQS+gVuXCji1k9y5984aSuo6IELh3Pu5d57gpgzpR3nw1pb39jc2s7t5Hf39g8OC0fHbRUlktAWiXgkuwFWlDNBW5ppTruxpDgMOO0E0+vM79xTqVgkbvUspn6Ix4KNGMHaSM3poFB07ErF86pXyLHdS8+tlTPilMulGnJtZ4EirNAYFN77w4gkIRWacKxUz3Vi7adYakY4nef7iaIxJlM8pj1DBQ6p8tPFoXN0bpQhGkXSlNBooX6fSHGo1CwMTGeI9UT99jLxL6+X6FHNT5mIE00FWS4aJRzpCGVfoyGTlGg+MwQTycytiEywxESbbPImhK9P0f+kXTKx2G7TK9a9VRw5OIUzuAAXqlCHG2hACwhQeIAneLburEfrxXpdtq5Zq5kT+AHr7RNBDY0w</latexit><latexit sha1_base64="OSVpthl7bZLrnH3UTqOtdanyPNQ=">AAAB6HicdVDLSsNAFL3xWeur6tLNYBFchaSmD3cFNy5bsA9oQ5lMp+3YySTMTIQS+gVuXCji1k9y5984aSuo6IELh3Pu5d57gpgzpR3nw1pb39jc2s7t5Hf39g8OC0fHbRUlktAWiXgkuwFWlDNBW5ppTruxpDgMOO0E0+vM79xTqVgkbvUspn6Ix4KNGMHaSM3poFB07ErF86pXyLHdS8+tlTPilMulGnJtZ4EirNAYFN77w4gkIRWacKxUz3Vi7adYakY4nef7iaIxJlM8pj1DBQ6p8tPFoXN0bpQhGkXSlNBooX6fSHGo1CwMTGeI9UT99jLxL6+X6FHNT5mIE00FWS4aJRzpCGVfoyGTlGg+MwQTycytiEywxESbbPImhK9P0f+kXTKx2G7TK9a9VRw5OIUzuAAXqlCHG2hACwhQeIAneLburEfrxXpdtq5Zq5kT+AHr7RNBDY0w</latexit><latexit sha1_base64="OSVpthl7bZLrnH3UTqOtdanyPNQ=">AAAB6HicdVDLSsNAFL3xWeur6tLNYBFchaSmD3cFNy5bsA9oQ5lMp+3YySTMTIQS+gVuXCji1k9y5984aSuo6IELh3Pu5d57gpgzpR3nw1pb39jc2s7t5Hf39g8OC0fHbRUlktAWiXgkuwFWlDNBW5ppTruxpDgMOO0E0+vM79xTqVgkbvUspn6Ix4KNGMHaSM3poFB07ErF86pXyLHdS8+tlTPilMulGnJtZ4EirNAYFN77w4gkIRWacKxUz3Vi7adYakY4nef7iaIxJlM8pj1DBQ6p8tPFoXN0bpQhGkXSlNBooX6fSHGo1CwMTGeI9UT99jLxL6+X6FHNT5mIE00FWS4aJRzpCGVfoyGTlGg+MwQTycytiEywxESbbPImhK9P0f+kXTKx2G7TK9a9VRw5OIUzuAAXqlCHG2hACwhQeIAneLburEfrxXpdtq5Zq5kT+AHr7RNBDY0w</latexit><latexit sha1_base64="OSVpthl7bZLrnH3UTqOtdanyPNQ=">AAAB6HicdVDLSsNAFL3xWeur6tLNYBFchaSmD3cFNy5bsA9oQ5lMp+3YySTMTIQS+gVuXCji1k9y5984aSuo6IELh3Pu5d57gpgzpR3nw1pb39jc2s7t5Hf39g8OC0fHbRUlktAWiXgkuwFWlDNBW5ppTruxpDgMOO0E0+vM79xTqVgkbvUspn6Ix4KNGMHaSM3poFB07ErF86pXyLHdS8+tlTPilMulGnJtZ4EirNAYFN77w4gkIRWacKxUz3Vi7adYakY4nef7iaIxJlM8pj1DBQ6p8tPFoXN0bpQhGkXSlNBooX6fSHGo1CwMTGeI9UT99jLxL6+X6FHNT5mIE00FWS4aJRzpCGVfoyGTlGg+MwQTycytiEywxESbbPImhK9P0f+kXTKx2G7TK9a9VRw5OIUzuAAXqlCHG2hACwhQeIAneLburEfrxXpdtq5Zq5kT+AHr7RNBDY0w</latexit>
k0<latexit sha1_base64="yC+mOBKc1K9Cosnmd5PnqsiH0PU=">AAAB6nicdVDLSgMxFL3js9ZX1aWbYBFcDUlpte4KblxWtA9oh5JJ0zY0kxmSjFCGfoIbF4q49Yvc+TemD0FFD1w4nHMv994TJlIYi/GHt7K6tr6xmdvKb+/s7u0XDg6bJk414w0Wy1i3Q2q4FIo3rLCStxPNaRRK3grHVzO/dc+1EbG6s5OEBxEdKjEQjFon3Y57uFcoYh9XcIkQ5Ej1nFxiRyqkXClVEfHxHEVYot4rvHf7MUsjriyT1JgOwYkNMqqtYJJP893U8ISyMR3yjqOKRtwE2fzUKTp1Sh8NYu1KWTRXv09kNDJmEoWuM6J2ZH57M/Evr5PaQTXIhEpSyxVbLBqkEtkYzf5GfaE5s3LiCGVauFsRG1FNmXXp5F0IX5+i/0mz5BPsk5tysVZexpGDYziBMyBwATW4hjo0gMEQHuAJnj3pPXov3uuidcVbzhzBD3hvn0WDjb4=</latexit><latexit sha1_base64="yC+mOBKc1K9Cosnmd5PnqsiH0PU=">AAAB6nicdVDLSgMxFL3js9ZX1aWbYBFcDUlpte4KblxWtA9oh5JJ0zY0kxmSjFCGfoIbF4q49Yvc+TemD0FFD1w4nHMv994TJlIYi/GHt7K6tr6xmdvKb+/s7u0XDg6bJk414w0Wy1i3Q2q4FIo3rLCStxPNaRRK3grHVzO/dc+1EbG6s5OEBxEdKjEQjFon3Y57uFcoYh9XcIkQ5Ej1nFxiRyqkXClVEfHxHEVYot4rvHf7MUsjriyT1JgOwYkNMqqtYJJP893U8ISyMR3yjqOKRtwE2fzUKTp1Sh8NYu1KWTRXv09kNDJmEoWuM6J2ZH57M/Evr5PaQTXIhEpSyxVbLBqkEtkYzf5GfaE5s3LiCGVauFsRG1FNmXXp5F0IX5+i/0mz5BPsk5tysVZexpGDYziBMyBwATW4hjo0gMEQHuAJnj3pPXov3uuidcVbzhzBD3hvn0WDjb4=</latexit><latexit sha1_base64="yC+mOBKc1K9Cosnmd5PnqsiH0PU=">AAAB6nicdVDLSgMxFL3js9ZX1aWbYBFcDUlpte4KblxWtA9oh5JJ0zY0kxmSjFCGfoIbF4q49Yvc+TemD0FFD1w4nHMv994TJlIYi/GHt7K6tr6xmdvKb+/s7u0XDg6bJk414w0Wy1i3Q2q4FIo3rLCStxPNaRRK3grHVzO/dc+1EbG6s5OEBxEdKjEQjFon3Y57uFcoYh9XcIkQ5Ej1nFxiRyqkXClVEfHxHEVYot4rvHf7MUsjriyT1JgOwYkNMqqtYJJP893U8ISyMR3yjqOKRtwE2fzUKTp1Sh8NYu1KWTRXv09kNDJmEoWuM6J2ZH57M/Evr5PaQTXIhEpSyxVbLBqkEtkYzf5GfaE5s3LiCGVauFsRG1FNmXXp5F0IX5+i/0mz5BPsk5tysVZexpGDYziBMyBwATW4hjo0gMEQHuAJnj3pPXov3uuidcVbzhzBD3hvn0WDjb4=</latexit><latexit sha1_base64="yC+mOBKc1K9Cosnmd5PnqsiH0PU=">AAAB6nicdVDLSgMxFL3js9ZX1aWbYBFcDUlpte4KblxWtA9oh5JJ0zY0kxmSjFCGfoIbF4q49Yvc+TemD0FFD1w4nHMv994TJlIYi/GHt7K6tr6xmdvKb+/s7u0XDg6bJk414w0Wy1i3Q2q4FIo3rLCStxPNaRRK3grHVzO/dc+1EbG6s5OEBxEdKjEQjFon3Y57uFcoYh9XcIkQ5Ej1nFxiRyqkXClVEfHxHEVYot4rvHf7MUsjriyT1JgOwYkNMqqtYJJP893U8ISyMR3yjqOKRtwE2fzUKTp1Sh8NYu1KWTRXv09kNDJmEoWuM6J2ZH57M/Evr5PaQTXIhEpSyxVbLBqkEtkYzf5GfaE5s3LiCGVauFsRG1FNmXXp5F0IX5+i/0mz5BPsk5tysVZexpGDYziBMyBwATW4hjo0gMEQHuAJnj3pPXov3uuidcVbzhzBD3hvn0WDjb4=</latexit>
6. • Rejection criterion: selective use of translated sample
• This reduces a risk of unreliable generation when is small
• Optimal sampling : better choice of majority seed
• Majority classes to sample are selected by considering two aspects:
(a) Maximizes the acceptance probability (b) Chooses diverse classes
M2m: Techniques for Better Efficiency
x⇤
<latexit sha1_base64="F1KMxJXvEGNv3MpOrIRy7Tp7EXs=">AAAB7HicbVA9SwNBEJ2LXzF+RS1tFoMgFuFOBC0sAjaWEbwkkJxhbzNJluztHbt7YjjyG2wsFLH1B9n5b9wkV2jig4HHezPMzAsTwbVx3W+nsLK6tr5R3Cxtbe/s7pX3Dxo6ThVDn8UiVq2QahRcom+4EdhKFNIoFNgMRzdTv/mISvNY3ptxgkFEB5L3OaPGSv7TQ3Y26ZYrbtWdgSwTLycVyFHvlr86vZilEUrDBNW67bmJCTKqDGcCJ6VOqjGhbEQH2LZU0gh1kM2OnZATq/RIP1a2pCEz9fdERiOtx1FoOyNqhnrRm4r/ee3U9K+CjMskNSjZfFE/FcTEZPo56XGFzIixJZQpbm8lbEgVZcbmU7IheIsvL5PGedVzq97dRaV2ncdRhCM4hlPw4BJqcAt18IEBh2d4hTdHOi/Ou/Mxby04+cwh/IHz+QPCTo6e</latexit><latexit sha1_base64="F1KMxJXvEGNv3MpOrIRy7Tp7EXs=">AAAB7HicbVA9SwNBEJ2LXzF+RS1tFoMgFuFOBC0sAjaWEbwkkJxhbzNJluztHbt7YjjyG2wsFLH1B9n5b9wkV2jig4HHezPMzAsTwbVx3W+nsLK6tr5R3Cxtbe/s7pX3Dxo6ThVDn8UiVq2QahRcom+4EdhKFNIoFNgMRzdTv/mISvNY3ptxgkFEB5L3OaPGSv7TQ3Y26ZYrbtWdgSwTLycVyFHvlr86vZilEUrDBNW67bmJCTKqDGcCJ6VOqjGhbEQH2LZU0gh1kM2OnZATq/RIP1a2pCEz9fdERiOtx1FoOyNqhnrRm4r/ee3U9K+CjMskNSjZfFE/FcTEZPo56XGFzIixJZQpbm8lbEgVZcbmU7IheIsvL5PGedVzq97dRaV2ncdRhCM4hlPw4BJqcAt18IEBh2d4hTdHOi/Ou/Mxby04+cwh/IHz+QPCTo6e</latexit><latexit sha1_base64="F1KMxJXvEGNv3MpOrIRy7Tp7EXs=">AAAB7HicbVA9SwNBEJ2LXzF+RS1tFoMgFuFOBC0sAjaWEbwkkJxhbzNJluztHbt7YjjyG2wsFLH1B9n5b9wkV2jig4HHezPMzAsTwbVx3W+nsLK6tr5R3Cxtbe/s7pX3Dxo6ThVDn8UiVq2QahRcom+4EdhKFNIoFNgMRzdTv/mISvNY3ptxgkFEB5L3OaPGSv7TQ3Y26ZYrbtWdgSwTLycVyFHvlr86vZilEUrDBNW67bmJCTKqDGcCJ6VOqjGhbEQH2LZU0gh1kM2OnZATq/RIP1a2pCEz9fdERiOtx1FoOyNqhnrRm4r/ee3U9K+CjMskNSjZfFE/FcTEZPo56XGFzIixJZQpbm8lbEgVZcbmU7IheIsvL5PGedVzq97dRaV2ncdRhCM4hlPw4BJqcAt18IEBh2d4hTdHOi/Ou/Mxby04+cwh/IHz+QPCTo6e</latexit><latexit sha1_base64="F1KMxJXvEGNv3MpOrIRy7Tp7EXs=">AAAB7HicbVA9SwNBEJ2LXzF+RS1tFoMgFuFOBC0sAjaWEbwkkJxhbzNJluztHbt7YjjyG2wsFLH1B9n5b9wkV2jig4HHezPMzAsTwbVx3W+nsLK6tr5R3Cxtbe/s7pX3Dxo6ThVDn8UiVq2QahRcom+4EdhKFNIoFNgMRzdTv/mISvNY3ptxgkFEB5L3OaPGSv7TQ3Y26ZYrbtWdgSwTLycVyFHvlr86vZilEUrDBNW67bmJCTKqDGcCJ6VOqjGhbEQH2LZU0gh1kM2OnZATq/RIP1a2pCEz9fdERiOtx1FoOyNqhnrRm4r/ee3U9K+CjMskNSjZfFE/FcTEZPo56XGFzIixJZQpbm8lbEgVZcbmU7IheIsvL5PGedVzq97dRaV2ncdRhCM4hlPw4BJqcAt18IEBh2d4hTdHOi/Ou/Mxby04+cwh/IHz+QPCTo6e</latexit>
x0<latexit sha1_base64="5huwj3zFVPbko7iCRKaz92eG43g=">AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0IOHghePFUxbaEPZbCft0s0m7G7EEvobvHhQxKs/yJv/xm2bg7Y+GHi8N8PMvDAVXBvX/XZKa+sbm1vl7crO7t7+QfXwqKWTTDH0WSIS1QmpRsEl+oYbgZ1UIY1Dge1wfDvz24+oNE/kg5mkGMR0KHnEGTVW8p/6uTvtV2tu3Z2DrBKvIDUo0OxXv3qDhGUxSsME1brruakJcqoMZwKnlV6mMaVsTIfYtVTSGHWQz4+dkjOrDEiUKFvSkLn6eyKnsdaTOLSdMTUjvezNxP+8bmai6yDnMs0MSrZYFGWCmITMPicDrpAZMbGEMsXtrYSNqKLM2HwqNgRv+eVV0rqoe27du7+sNW6KOMpwAqdwDh5cQQPuoAk+MODwDK/w5kjnxXl3PhatJaeYOYY/cD5/AMzzjqU=</latexit><latexit sha1_base64="5huwj3zFVPbko7iCRKaz92eG43g=">AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0IOHghePFUxbaEPZbCft0s0m7G7EEvobvHhQxKs/yJv/xm2bg7Y+GHi8N8PMvDAVXBvX/XZKa+sbm1vl7crO7t7+QfXwqKWTTDH0WSIS1QmpRsEl+oYbgZ1UIY1Dge1wfDvz24+oNE/kg5mkGMR0KHnEGTVW8p/6uTvtV2tu3Z2DrBKvIDUo0OxXv3qDhGUxSsME1brruakJcqoMZwKnlV6mMaVsTIfYtVTSGHWQz4+dkjOrDEiUKFvSkLn6eyKnsdaTOLSdMTUjvezNxP+8bmai6yDnMs0MSrZYFGWCmITMPicDrpAZMbGEMsXtrYSNqKLM2HwqNgRv+eVV0rqoe27du7+sNW6KOMpwAqdwDh5cQQPuoAk+MODwDK/w5kjnxXl3PhatJaeYOYY/cD5/AMzzjqU=</latexit><latexit sha1_base64="5huwj3zFVPbko7iCRKaz92eG43g=">AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0IOHghePFUxbaEPZbCft0s0m7G7EEvobvHhQxKs/yJv/xm2bg7Y+GHi8N8PMvDAVXBvX/XZKa+sbm1vl7crO7t7+QfXwqKWTTDH0WSIS1QmpRsEl+oYbgZ1UIY1Dge1wfDvz24+oNE/kg5mkGMR0KHnEGTVW8p/6uTvtV2tu3Z2DrBKvIDUo0OxXv3qDhGUxSsME1brruakJcqoMZwKnlV6mMaVsTIfYtVTSGHWQz4+dkjOrDEiUKFvSkLn6eyKnsdaTOLSdMTUjvezNxP+8bmai6yDnMs0MSrZYFGWCmITMPicDrpAZMbGEMsXtrYSNqKLM2HwqNgRv+eVV0rqoe27du7+sNW6KOMpwAqdwDh5cQQPuoAk+MODwDK/w5kjnxXl3PhatJaeYOYY/cD5/AMzzjqU=</latexit><latexit sha1_base64="5huwj3zFVPbko7iCRKaz92eG43g=">AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0IOHghePFUxbaEPZbCft0s0m7G7EEvobvHhQxKs/yJv/xm2bg7Y+GHi8N8PMvDAVXBvX/XZKa+sbm1vl7crO7t7+QfXwqKWTTDH0WSIS1QmpRsEl+oYbgZ1UIY1Dge1wfDvz24+oNE/kg5mkGMR0KHnEGTVW8p/6uTvtV2tu3Z2DrBKvIDUo0OxXv3qDhGUxSsME1brruakJcqoMZwKnlV6mMaVsTIfYtVTSGHWQz4+dkjOrDEiUKFvSkLn6eyKnsdaTOLSdMTUjvezNxP+8bmai6yDnMs0MSrZYFGWCmITMPicDrpAZMbGEMsXtrYSNqKLM2HwqNgRv+eVV0rqoe27du7+sNW6KOMpwAqdwDh5cQQPuoAk+MODwDK/w5kjnxXl3PhatJaeYOYY/cD5/AMzzjqU=</latexit>
# samples of class
Nk0<latexit sha1_base64="8Ml4gG7BbFmCtUqx3C4PlF933m0=">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</latexit><latexit sha1_base64="8Ml4gG7BbFmCtUqx3C4PlF933m0=">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</latexit><latexit sha1_base64="8Ml4gG7BbFmCtUqx3C4PlF933m0=">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</latexit><latexit sha1_base64="8Ml4gG7BbFmCtUqx3C4PlF933m0=">AAACynicjVHLSsNAFD2Nr1pfVZdugkVwVRIR7LLgxoVIBfuAWkoyndaheTGZCKV05w+41Q8T/0D/wjtjCmoRnZDkzLnn3Jl7r58EIlWO81qwlpZXVteK66WNza3tnfLuXiuNM8l4k8VBLDu+l/JARLyphAp4J5HcC/2At/3xuY6377lMRRzdqEnCe6E3isRQME8R1b7qT8d9Z9YvV5yqY5a9CNwcVJCvRlx+wS0GiMGQIQRHBEU4gIeUni5cOEiI62FKnCQkTJxjhhJ5M1JxUnjEjuk7ol03ZyPa65ypcTM6JaBXktPGEXli0knC+jTbxDOTWbO/5Z6anPpuE/r7ea6QWIU7Yv/yzZX/9elaFIaomRoE1ZQYRlfH8iyZ6Yq+uf2lKkUZEuI0HlBcEmbGOe+zbTypqV331jPxN6PUrN6zXJvhXd+SBuz+HOciaJ1UXafqXp9W6rV81EUc4BDHNM8z1HGBBpqmykc84dm6tKQ1saafUquQe/bxbVkPH298kdY=</latexit>
major minor
Q(k0|k) / 1 (Nk0
Nk)+
<latexit sha1_base64="4P/5yGI7mPJRAOS+dUzL1rT5mFA=">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</latexit><latexit sha1_base64="4P/5yGI7mPJRAOS+dUzL1rT5mFA=">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</latexit><latexit sha1_base64="4P/5yGI7mPJRAOS+dUzL1rT5mFA=">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</latexit><latexit sha1_base64="4P/5yGI7mPJRAOS+dUzL1rT5mFA=">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</latexit>
Q(k0|k)<latexit sha1_base64="JOFzpi7JU8iyOJ7GwGax5fw+8L8=">AAAB73icbVA9SwNBEJ2LXzF+RS1tFoMQm7AnghYWARvLBMwHJEfY2+wly+3tnbt7QjjzJ2wsFLH179j5b9wkV2jig4HHezPMzPMTwbXB+NsprK1vbG4Vt0s7u3v7B+XDo7aOU0VZi8YiVl2faCa4ZC3DjWDdRDES+YJ1/PB25ncemdI8lvdmkjAvIiPJA06JsVK3WQ0H+Ck8H5QruIbnQKvEzUkFcjQG5a/+MKZpxKShgmjdc3FivIwow6lg01I/1SwhNCQj1rNUkohpL5vfO0VnVhmiIFa2pEFz9fdERiKtJ5FvOyNixnrZm4n/eb3UBNdexmWSGibpYlGQCmRiNHseDbli1IiJJYQqbm9FdEwUocZGVLIhuMsvr5L2Rc3FNbd5Wanf5HEU4QROoQouXEEd7qABLaAg4Ble4c15cF6cd+dj0Vpw8plj+APn8wcIdY9H</latexit><latexit sha1_base64="JOFzpi7JU8iyOJ7GwGax5fw+8L8=">AAAB73icbVA9SwNBEJ2LXzF+RS1tFoMQm7AnghYWARvLBMwHJEfY2+wly+3tnbt7QjjzJ2wsFLH179j5b9wkV2jig4HHezPMzPMTwbXB+NsprK1vbG4Vt0s7u3v7B+XDo7aOU0VZi8YiVl2faCa4ZC3DjWDdRDES+YJ1/PB25ncemdI8lvdmkjAvIiPJA06JsVK3WQ0H+Ck8H5QruIbnQKvEzUkFcjQG5a/+MKZpxKShgmjdc3FivIwow6lg01I/1SwhNCQj1rNUkohpL5vfO0VnVhmiIFa2pEFz9fdERiKtJ5FvOyNixnrZm4n/eb3UBNdexmWSGibpYlGQCmRiNHseDbli1IiJJYQqbm9FdEwUocZGVLIhuMsvr5L2Rc3FNbd5Wanf5HEU4QROoQouXEEd7qABLaAg4Ble4c15cF6cd+dj0Vpw8plj+APn8wcIdY9H</latexit><latexit sha1_base64="JOFzpi7JU8iyOJ7GwGax5fw+8L8=">AAAB73icbVA9SwNBEJ2LXzF+RS1tFoMQm7AnghYWARvLBMwHJEfY2+wly+3tnbt7QjjzJ2wsFLH179j5b9wkV2jig4HHezPMzPMTwbXB+NsprK1vbG4Vt0s7u3v7B+XDo7aOU0VZi8YiVl2faCa4ZC3DjWDdRDES+YJ1/PB25ncemdI8lvdmkjAvIiPJA06JsVK3WQ0H+Ck8H5QruIbnQKvEzUkFcjQG5a/+MKZpxKShgmjdc3FivIwow6lg01I/1SwhNCQj1rNUkohpL5vfO0VnVhmiIFa2pEFz9fdERiKtJ5FvOyNixnrZm4n/eb3UBNdexmWSGibpYlGQCmRiNHseDbli1IiJJYQqbm9FdEwUocZGVLIhuMsvr5L2Rc3FNbd5Wanf5HEU4QROoQouXEEd7qABLaAg4Ble4c15cF6cd+dj0Vpw8plj+APn8wcIdY9H</latexit><latexit sha1_base64="JOFzpi7JU8iyOJ7GwGax5fw+8L8=">AAAB73icbVA9SwNBEJ2LXzF+RS1tFoMQm7AnghYWARvLBMwHJEfY2+wly+3tnbt7QjjzJ2wsFLH179j5b9wkV2jig4HHezPMzPMTwbXB+NsprK1vbG4Vt0s7u3v7B+XDo7aOU0VZi8YiVl2faCa4ZC3DjWDdRDES+YJ1/PB25ncemdI8lvdmkjAvIiPJA06JsVK3WQ0H+Ck8H5QruIbnQKvEzUkFcjQG5a/+MKZpxKShgmjdc3FivIwow6lg01I/1SwhNCQj1rNUkohpL5vfO0VnVhmiIFa2pEFz9fdERiKtJ5FvOyNixnrZm4n/eb3UBNdexmWSGibpYlGQCmRiNHseDbli1IiJJYQqbm9FdEwUocZGVLIhuMsvr5L2Rc3FNbd5Wanf5HEU4QROoQouXEEd7qABLaAg4Ble4c15cF6cd+dj0Vpw8plj+APn8wcIdY9H</latexit>
7. • Two evaluation metrics for imbalanced classification models:
• Balanced accuracy (bACC): arithmetic mean over class-wise sensitivity
• Geometric mean scores (GM): geometric mean over class-wise sensitivity
• Various types of baseline methods
• Re-sampling (RS) and re-weighting (RW)
• Re-sampling variants: SMOTE, deferred re-sampling (DRS)
[Chawla et al. 2002; Cao et al. 2019]
• Re-weighting variants: class-balanced RW (CB-RW) [Cui et al. 2019]
• Minority regularization: focal loss (Focal), label-dist. aware margin (LDAM)
[Lin et al. 2017; Cao et al. 2019]
Experiments
[Chawla et al. 2002] SMOTE: Synthetic Minority Oversampling Technique. In JAIR 2002
[Lin et al. 2017] Focal Loss for Dense Object Detection., CVPR 2017
[Cao et al. 2019] Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss. In NeurIPS 2019
[Cui et al. 2019] Class-balanced Loss Based on Effective Number of Samples. In CVPR, 2019
8. • Synthetically imbalanced version of CIFAR-10/100
• We control the imbalance ratio
• are set to follow an exponential decay across
Experiments: Long-tailed CIFAR-10/100
⇢ = N1/Nk<latexit sha1_base64="NwvRw269Yl0ESgPSsPYYzPTnY3I=">AAAC13icjVHLSsNAFD3GV62vWJdugkVwVRMR7EYouHElFexD2lKSdNqGpkmYTMRSijtx6w+41T8S/0D/wjtjCmoRnZCZM+fec2buXCfyvViY5uucNr+wuLScWcmurq1vbOpbuWocJtxlFTf0Q1537Jj5XsAqwhM+q0ec2UPHZzVncCrjtWvGYy8MLsUoYq2h3Qu8rufagqi2nmvyfmicGOftsTU5oHkwaet5s2CqYcwCKwV5pKMc6i9oooMQLhIMwRBAEPZhI6avAQsmIuJaGBPHCXkqzjBBlrQJZTHKsIkd0NyjXSNlA9pLz1ipXTrFp5+T0sAeaULK44TlaYaKJ8pZsr95j5WnvNuIVif1GhIr0Cf2L9008786WYtAF0VVg0c1RYqR1bmpS6JeRd7c+FKVIIeIOIk7FOeEXaWcvrOhNLGqXb6treJvKlOycu+muQne5S2pwdbPds6C6mHBMgvWxVG+VExbncEOdrFP/TxGCWcoo0LeN3jEE561K+1Wu9PuP1O1uVSzjW9De/gAVqKWDQ==</latexit><latexit sha1_base64="NwvRw269Yl0ESgPSsPYYzPTnY3I=">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</latexit><latexit sha1_base64="NwvRw269Yl0ESgPSsPYYzPTnY3I=">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</latexit><latexit sha1_base64="NwvRw269Yl0ESgPSsPYYzPTnY3I=">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</latexit>
M2m surpasses all the
baselines tested
M2m further improves
existing regularization
Nk<latexit sha1_base64="wk8D3NDoQs7Lt/vJBl8Aiej2MHo=">AAACyHicjVHLSsNAFD2Nr/quunQTLIKrkohglwU34kIqmLZQS0mm0zo0TcJkopTSjT/gVr9M/AP9C++MKahFdEKSM+fec2buvUESilQ5zmvBWlhcWl4prq6tb2xubZd2dhtpnEnGPRaHsWwFfspDEXFPCRXyViK5PwpC3gyGZzrevOMyFXF0rcYJ74z8QST6gvmKKO+yOxlOu6WyU3HMsueBm4My8lWPSy+4QQ8xGDKMwBFBEQ7hI6WnDRcOEuI6mBAnCQkT55hijbQZZXHK8Ikd0ndAu3bORrTXnqlRMzolpFeS0sYhaWLKk4T1abaJZ8ZZs795T4ynvtuY/kHuNSJW4ZbYv3SzzP/qdC0KfVRNDYJqSgyjq2O5S2a6om9uf6lKkUNCnMY9ikvCzChnfbaNJjW16976Jv5mMjWr9yzPzfCub0kDdn+Ocx40jiuuU3GvTsq1aj7qIvZxgCOa5ylqOEcdHnkLPOIJz9aFlVj31vgz1Srkmj18W9bDB8wYkTM=</latexit><latexit sha1_base64="wk8D3NDoQs7Lt/vJBl8Aiej2MHo=">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</latexit><latexit sha1_base64="wk8D3NDoQs7Lt/vJBl8Aiej2MHo=">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</latexit><latexit sha1_base64="wk8D3NDoQs7Lt/vJBl8Aiej2MHo=">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</latexit>
k<latexit sha1_base64="VJ9SyfzSOjNOMVbxQ3SrUcGK1J0=">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</latexit><latexit sha1_base64="VJ9SyfzSOjNOMVbxQ3SrUcGK1J0=">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</latexit><latexit sha1_base64="VJ9SyfzSOjNOMVbxQ3SrUcGK1J0=">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</latexit><latexit sha1_base64="VJ9SyfzSOjNOMVbxQ3SrUcGK1J0=">AAACxHicjVHLSsNAFD2Nr1pfVZdugkVwVRIR7LIgiMsW7ANqkSSd1qHTJEwmQin6A27128Q/0L/wzjgFtYhOSHLm3HvOzL03TAXPlOe9Fpyl5ZXVteJ6aWNza3unvLvXzpJcRqwVJSKR3TDImOAxaymuBOumkgWTULBOOD7X8c4dkxlP4is1TVl/EoxiPuRRoIhqjm/KFa/qmeUuAt+CCuxqJOUXXGOABBFyTMAQQxEWCJDR04MPDylxfcyIk4S4iTPco0TanLIYZQTEjuk7ol3PsjHttWdm1BGdIuiVpHRxRJqE8iRhfZpr4rlx1uxv3jPjqe82pX9ovSbEKtwS+5dunvlfna5FYYiaqYFTTalhdHWRdclNV/TN3S9VKXJIidN4QHFJODLKeZ9do8lM7bq3gYm/mUzN6n1kc3O861vSgP2f41wE7ZOq71X95mmlXrOjLuIAhzimeZ6hjks00DLej3jCs3PhCCdz8s9Up2A1+/i2nIcPSTqPZg==</latexit>
9. • Datasets with natural imbalances
• Vision tasks: CelebA-5 and SUN397
• NLP tasks: Twitter and Reuters
• Tend to have a much significant imbalance
Experiments: Real-world Imbalanced Datasets
M2m works even better
under a harsh imbalance
10. • A simple yet powerful over-sampling for imbalanced classification
• Diversity of majority is effective to overcome the scarcity of minority
• Adversarial examples could be a good feature at least for the minority
Conclusion
Thank you for your attention 😄