Learning to Reweight Examples for Robust Deep Learning

Abstract

Deep neural networks have been shown to be very powerful modeling tools for many supervised learning tasks involving complex input patterns. However, they can also easily overfit to training set biases and label noises. In addition to various regularizers, example reweighting algorithms are popular solutions to these problems, but they require careful tuning of additional hyperparameters, such as example mining schedules and regularization hyperparameters. In contrast to past reweighting methods, which typically consist of functions of the cost value of each example, in this work we propose a novel meta-learning algorithm that learns to assign weights to training examples based on their gradient directions. To determine the example weights, our method performs a meta gradient descent step on the current mini-batch example weights (which are initialized from zero) to minimize the loss on a clean unbiased validation set. Our proposed method can be easily implemented on any type of deep network, does not require any additional hyperparameter tuning, and achieves impressive performance on class imbalance and corrupted label problems where only a small amount of clean validation data is available.

13 pages; Published at ICML 2018; Code released at: https://github.com/uber-research/learning-to-reweight-examples

References in corpus (5)

• SMOTE: Synthetic Minority Over-sampling Technique
• Understanding Black-box Predictions via Influence Functions
• The Cityscapes Dataset for Semantic Urban Scene Understanding
• Training Deep Neural Networks on Noisy Labels with Bootstrapping
• Stochastic Hyperparameter Optimization through Hypernetworks

Cited by in corpus (194)

• Generalized Cross Entropy Loss for Training Deep Neural Networks with Noisy Labels
• Deep Multi-modal Object Detection and Semantic Segmentation for Autonomous Driving: Datasets, Methods, and Challenges
• Regularized Deep Networks in Intelligent Transportation Systems: A Taxonomy and a Case Study
• Co-teaching: Robust Training of Deep Neural Networks with Extremely Noisy Labels
• Meta-Weight-Net: Learning an Explicit Mapping For Sample Weighting
• Image Classification with Deep Learning in the Presence of Noisy Labels: A Survey
• Balanced Meta-Softmax for Long-Tailed Visual Recognition
• Early-Learning Regularization Prevents Memorization of Noisy Labels
• Are Anchor Points Really Indispensable in Label-Noise Learning?
• Masking: A New Perspective of Noisy Supervision
• Gradient Descent with Early Stopping is Provably Robust to Label Noise for Overparameterized Neural Networks
• Robust Federated Learning with Noisy Labels
• Using Trusted Data to Train Deep Networks on Labels Corrupted by Severe Noise
• LongReMix: Robust Learning with High Confidence Samples in a Noisy Label Environment
• Dual T: Reducing Estimation Error for Transition Matrix in Label-noise Learning
• Toward Open-World Electroencephalogram Decoding Via Deep Learning: A Comprehensive Survey
• Long-tailed Visual Recognition via Gaussian Clouded Logit Adjustment
• Towards Intersectionality in Machine Learning: Including More Identities, Handling Underrepresentation, and Performing Evaluation
• Distribution Aligning Refinery of Pseudo-label for Imbalanced Semi-supervised Learning
• Self-Adaptive Training: beyond Empirical Risk Minimization
• Coresets via Bilevel Optimization for Continual Learning and Streaming
• Learning to Generalize from Sparse and Underspecified Rewards
• Meta Balanced Network for Fair Face Recognition
• Combating noisy labels by agreement: A joint training method with co-regularization
• Self-supervised Learning is More Robust to Dataset Imbalance
• Co-Correcting: Noise-tolerant Medical Image Classification via mutual Label Correction
• Identifying and Compensating for Feature Deviation in Imbalanced Deep Learning
• SELF: Learning to Filter Noisy Labels with Self-Ensembling
• Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect
• On the Minimal Supervision for Training Any Binary Classifier from Only Unlabeled Data
• Feature Fusion from Head to Tail for Long-Tailed Visual Recognition
• Denoising Multi-Source Weak Supervision for Neural Text Classification
• Identifying Mislabeled Data using the Area Under the Margin Ranking
• IMAE for Noise-Robust Learning: Mean Absolute Error Does Not Treat Examples Equally and Gradient Magnitude's Variance Matters
• An Entropic Optimal Transport Loss for Learning Deep Neural Networks under Label Noise in Remote Sensing Images
• Label Noise Types and Their Effects on Deep Learning
• Balancing Unobserved Confounding with a Few Unbiased Ratings in Debiased Recommendations
• Learning from Pixel-Level Label Noise: A New Perspective for Semi-Supervised Semantic Segmentation
• FINE Samples for Learning with Noisy Labels
• Distribution-Balanced Loss for Multi-Label Classification in Long-Tailed Datasets
• Adaptive Self-training for Few-shot Neural Sequence Labeling
• CReST: A Class-Rebalancing Self-Training Framework for Imbalanced Semi-Supervised Learning
• L_DMI: An Information-theoretic Noise-robust Loss Function
• Boosting Facial Expression Recognition by A Semi-Supervised Progressive Teacher
• Deep Imbalanced Learning for Face Recognition and Attribute Prediction
• Open-set Label Noise Can Improve Robustness Against Inherent Label Noise
• The DNNLikelihood: enhancing likelihood distribution with Deep Learning
• Rethinking Class-Balanced Methods for Long-Tailed Visual Recognition from a Domain Adaptation Perspective
• Learning with Bounded Instance- and Label-dependent Label Noise
• MESA: Boost Ensemble Imbalanced Learning with MEta-SAmpler
• Provably End-to-end Label-Noise Learning without Anchor Points
• TGDM: Target Guided Dynamic Mixup for Cross-Domain Few-Shot Learning
• Uncertainty Based Detection and Relabeling of Noisy Image Labels
• Penalty Method for Inversion-Free Deep Bilevel Optimization
• Not All Unlabeled Data are Equal: Learning to Weight Data in Semi-supervised Learning
• Hyperbolic Geometric Graph Representation Learning for Hierarchy-imbalance Node Classification
• Heteroskedastic and Imbalanced Deep Learning with Adaptive Regularization
• Distribution Density, Tails, and Outliers in Machine Learning: Metrics and Applications
• MetaLabelNet: Learning to Generate Soft-Labels from Noisy-Labels
• Learning Soft Labels via Meta Learning
• M2m: Imbalanced Classification via Major-to-minor Translation
• How does Early Stopping Help Generalization against Label Noise?
• Feature-Balanced Loss for Long-Tailed Visual Recognition
• MetaSAug: Meta Semantic Augmentation for Long-Tailed Visual Recognition
• How to Train Your MAML to Excel in Few-Shot Classification
• Robust Learning Under Label Noise With Iterative Noise-Filtering
• Simple and Effective Regularization Methods for Training on Noisily Labeled Data with Generalization Guarantee
• SimAug: Learning Robust Representations from Simulation for Trajectory Prediction
• Deep Self-Learning From Noisy Labels
• SWIPENET: Object detection in noisy underwater images
• Exploring Memorization in Adversarial Training
• ProSelfLC: Progressive Self Label Correction for Training Robust Deep Neural Networks
• Co-Learning Meets Stitch-Up for Noisy Multi-label Visual Recognition
• FR-Train: A Mutual Information-Based Approach to Fair and Robust Training
• Learning Bounds for Risk-sensitive Learning
• Balancing Training for Multilingual Neural Machine Translation
• Narrowing the Gap: Improved Detector Training with Noisy Location Annotations
• Rethinking Importance Weighting for Deep Learning under Distribution Shift
• Underwater object detection using Invert Multi-Class Adaboost with deep learning
• Detecting, Localising and Classifying Polyps from Colonoscopy Videos using Deep Learning
• Improving Medical Image Classification with Label Noise Using Dual-uncertainty Estimation
• On Tilted Losses in Machine Learning: Theory and Applications
• Affect Expression Behaviour Analysis in the Wild using Consensual Collaborative Training
• Imbalanced Continual Learning with Partitioning Reservoir Sampling
• Consensual Collaborative Training And Knowledge Distillation Based Facial Expression Recognition Under Noisy Annotations
• A Survey on Curriculum Learning
• Searching to Exploit Memorization Effect in Learning from Corrupted Labels
• Adversarial Robustness under Long-Tailed Distribution
• Universal Stagewise Learning for Non-Convex Problems with Convergence on Averaged Solutions
• Derivative Manipulation for General Example Weighting
• Learning to Impute: A General Framework for Semi-supervised Learning
• DoubleEnsemble: A New Ensemble Method Based on Sample Reweighting and Feature Selection for Financial Data Analysis
• Exploiting All Samples in Low-Resource Sentence Classification: Early Stopping and Initialization Parameters
• Teacher Supervises Students How to Learn From Partially Labeled Images for Facial Landmark Detection
• Combining Self-Supervised and Supervised Learning with Noisy Labels
• Label-Noise Robust Generative Adversarial Networks
• Task Agnostic Robust Learning on Corrupt Outputs by Correlation-Guided Mixture Density Networks
• Faster Meta Update Strategy for Noise-Robust Deep Learning
• Adaptive Sample Selection for Robust Learning under Label Noise
• MetaMixUp: Learning Adaptive Interpolation Policy of MixUp with Meta-Learning
• Countering Noisy Labels By Learning From Auxiliary Clean Labels
• Self-Adaptive Training: Bridging Supervised and Self-Supervised Learning
• Imbalanced Image Classification with Complement Cross Entropy
• Improving Generalization by Controlling Label-Noise Information in Neural Network Weights
• Weakly Supervised Learning with Side Information for Noisy Labeled Images
• Few-Shot Anomaly Detection for Polyp Frames from Colonoscopy
• Robust and On-the-fly Dataset Denoising for Image Classification
• Improving Robustness of Learning-based Autonomous Steering Using Adversarial Images
• The Resistance to Label Noise in K-NN and DNN Depends on its Concentration
• Tilted Empirical Risk Minimization
• Select-ProtoNet: Learning to Select for Few-Shot Disease Subtype Prediction
• Teaching with Commentaries
• A Novel Self-Supervised Re-labeling Approach for Training with Noisy Labels
• Optimizing Data Usage via Differentiable Rewards
• Discriminative-Generative Dual Memory Video Anomaly Detection
• Deep Learning Classification With Noisy Labels
• No Regret Sample Selection with Noisy Labels
• Reducing and Exploiting Data Augmentation Noise through Meta Reweighting Contrastive Learning for Text Classification
• Learning with Noisy Labels by Efficient Transition Matrix Estimation to Combat Label Miscorrection
• Online Hyperparameter Meta-Learning with Hypergradient Distillation
• Learning to Combat Noisy Labels via Classification Margins
• How Out-of-Distribution Data Hurts Semi-Supervised Learning
• Joint Negative and Positive Learning for Noisy Labels
• Data Manipulation: Towards Effective Instance Learning for Neural Dialogue Generation via Learning to Augment and Reweight
• Deep k-NN for Noisy Labels
• Less Is Better: Unweighted Data Subsampling via Influence Function
• Using Latent Codes for Class Imbalance Problem in Unsupervised Domain Adaptation
• NLNL: Negative Learning for Noisy Labels
• Safeguarded Dynamic Label Regression for Generalized Noisy Supervision
• Learning Image Labels On-the-fly for Training Robust Classification Models
• A Simple yet Effective Baseline for Robust Deep Learning with Noisy Labels
• Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning
• Robust Deep Graph Based Learning for Binary Classification
• From Recognition to Prediction: Analysis of Human Action and Trajectory Prediction in Video
• Minimizing Close-k Aggregate Loss Improves Classification
• MetaPerturb: Transferable Regularizer for Heterogeneous Tasks and Architectures
• Optimizing Black-box Metrics with Iterative Example Weighting
• Adversarial-Based Knowledge Distillation for Multi-Model Ensemble and Noisy Data Refinement
• Unified Gradient Reweighting for Model Biasing with Applications to Source Separation
• Class-Wise Difficulty-Balanced Loss for Solving Class-Imbalance
• Distilling effective supervision for robust medical image segmentation with noisy labels
• Improving Medical Annotation Quality to Decrease Labeling Burden Using Stratified Noisy Cross-Validation
• Negative-ResNet: Noisy Ambulatory Electrocardiogram Signal Classification Scheme
• Direct Differentiable Augmentation Search
• Mobility-aware Content Preference Learning in Decentralized Caching Networks
• Medical Image Segmentation with Limited Supervision: A Review of Deep Network Models
• Improving Training on Noisy Stuctured Labels
• ARM: A Confidence-Based Adversarial Reweighting Module for Coarse Semantic Segmentation
• Multiplicative Reweighting for Robust Neural Network Optimization
• Self-Paced Deep Regression Forests with Consideration on Underrepresented Examples
• Few-Shot Text Ranking with Meta Adapted Synthetic Weak Supervision
• Self-Supervised Person Detection in 2D Range Data using a Calibrated Camera
• Constrained Instance and Class Reweighting for Robust Learning under Label Noise
• Meta Label Correction for Noisy Label Learning
• Learning to Auto Weight: Entirely Data-driven and Highly Efficient Weighting Framework
• Leveraging Local Variation in Data: Sampling and Weighting Schemes for Supervised Deep Learning
• Energy Aligning for Biased Models
• Graph convolutional networks for learning with few clean and many noisy labels
• Meta Corrupted Pixels Mining for Medical Image Segmentation
• Sample Prior Guided Robust Model Learning to Suppress Noisy Labels
• Exploiting Class Similarity for Machine Learning with Confidence Labels and Projective Loss Functions
• Learning to Reweight with Deep Interactions
• Self-Training with Weak Supervision
• Semi-Supervised Learning with Meta-Gradient
• Gradient-based Hyperparameter Optimization Over Long Horizons
• Boosting Mapping Functionality of Neural Networks via Latent Feature Generation based on Reversible Learning
• Improving Generalization of Deep Fault Detection Models in the Presence of Mislabeled Data
• Ada-Segment: Automated Multi-loss Adaptation for Panoptic Segmentation
• Two-Stream Compare and Contrast Network for Vertebral Compression Fracture Diagnosis
• MSD: Saliency-aware Knowledge Distillation for Multimodal Understanding
• Robust Temporal Ensembling for Learning with Noisy Labels
• Spending Money Wisely: Online Electronic Coupon Allocation based on Real-Time User Intent Detection
• Simple and Effective Input Reformulations for Translation
• Calibrating Class Activation Maps for Long-Tailed Visual Recognition
• Meta Adaptation using Importance Weighted Demonstrations
• IPOF: An Extremely and Excitingly Simple Outlier Detection Booster via Infinite Propagation
• Noise Robust Generative Adversarial Networks
• Procrustean Training for Imbalanced Deep Learning
• MULTIMODAL ANALYSIS: Informed content estimation and audio source separation
• Biquality Learning: a Framework to Design Algorithms Dealing with Closed-Set Distribution Shifts
• Meta Auxiliary Learning for Facial Action Unit Detection
• Learning with Noisy Labels for Sentence-level Sentiment Classification
• Temporal-aware Language Representation Learning From Crowdsourced Labels
• Sum of Ranked Range Loss for Supervised Learning
• Compensation Learning
• Learning to Transfer Learn: Reinforcement Learning-Based Selection for Adaptive Transfer Learning
• Learning from Web Data with Self-Organizing Memory Module
• Which Samples Should be Learned First: Easy or Hard?
• DIVA: Dataset Derivative of a Learning Task
• One-shot Weakly-Supervised Segmentation in Medical Images
• Mitigating Class Boundary Label Uncertainty to Reduce Both Model Bias and Variance
• Task-Driven Data Verification via Gradient Descent
• Learning from Mistakes based on Class Weighting with Application to Neural Architecture Search
• Learning advisor networks for noisy image classification