Training Deep Neural Networks on Noisy Labels with Bootstrapping

Abstract

Current state-of-the-art deep learning systems for visual object recognition and detection use purely supervised training with regularization such as dropout to avoid overfitting. The performance depends critically on the amount of labeled examples, and in current practice the labels are assumed to be unambiguous and accurate. However, this assumption often does not hold; e.g. in recognition, class labels may be missing; in detection, objects in the image may not be localized; and in general, the labeling may be subjective. In this work we propose a generic way to handle noisy and incomplete labeling by augmenting the prediction objective with a notion of consistency. We consider a prediction consistent if the same prediction is made given similar percepts, where the notion of similarity is between deep network features computed from the input data. In experiments we demonstrate that our approach yields substantial robustness to label noise on several datasets. On MNIST handwritten digits, we show that our model is robust to label corruption. On the Toronto Face Database, we show that our model handles well the case of subjective labels in emotion recognition, achieving state-of-the- art results, and can also benefit from unlabeled face images with no modification to our method. On the ILSVRC2014 detection challenge data, we show that our approach extends to very deep networks, high resolution images and structured outputs, and results in improved scalable detection.

References in corpus (5)

• Improving neural networks by preventing co-adaptation of feature detectors
• Going Deeper with Convolutions
• Training Convolutional Networks with Noisy Labels
• DeepID-Net: multi-stage and deformable deep convolutional neural networks for object detection
• Analysis of Semi-Supervised Learning with the Yarowsky Algorithm

Cited by in corpus (58)

• Temporal Ensembling for Semi-Supervised Learning
• Collaborative Unsupervised Domain Adaptation for Medical Image Diagnosis
• Gradient Descent with Early Stopping is Provably Robust to Label Noise for Overparameterized Neural Networks
• Hard Sample Aware Noise Robust Learning for Histopathology Image Classification
• Combating noisy labels by agreement: A joint training method with co-regularization
• SELF: Learning to Filter Noisy Labels with Self-Ensembling
• Weakly Supervised Vessel Segmentation in X-ray Angiograms by Self-Paced Learning from Noisy Labels with Suggestive Annotation
• Multi-Granularity Denoising and Bidirectional Alignment for Weakly Supervised Semantic Segmentation
• Data Distillation: Towards Omni-Supervised Learning
• FaceNet2ExpNet: Regularizing a Deep Face Recognition Net for Expression Recognition
• Unsupervised Domain Adaptation of Black-Box Source Models
• Boosting Facial Expression Recognition by A Semi-Supervised Progressive Teacher
• Learning From Noisy Labels By Regularized Estimation Of Annotator Confusion
• Towards Understanding Label Smoothing
• Uncertainty Based Detection and Relabeling of Noisy Image Labels
• Gunrock 2.0: A User Adaptive Social Conversational System
• Learning Deep Networks from Noisy Labels with Dropout Regularization
• NRGNN: Learning a Label Noise-Resistant Graph Neural Network on Sparsely and Noisily Labeled Graphs
• Does label smoothing mitigate label noise?
• Robust Learning Under Label Noise With Iterative Noise-Filtering
• Segmentation-Aware Image Denoising without Knowing True Segmentation
• CIA-Net: Robust Nuclei Instance Segmentation with Contour-aware Information Aggregation
• Breast Mass Detection with Faster R-CNN: On the Feasibility of Learning from Noisy Annotations
• Robustness of Accuracy Metric and its Inspirations in Learning with Noisy Labels
• In Defense of the Triplet Loss Again: Learning Robust Person Re-Identification with Fast Approximated Triplet Loss and Label Distillation
• Unified Multisensory Perception: Weakly-Supervised Audio-Visual Video Parsing
• Coresets for Robust Training of Neural Networks against Noisy Labels
• Weakly Supervised Learning with Side Information for Noisy Labeled Images
• Attend in groups: a weakly-supervised deep learning framework for learning from web data
• Less is More: Learning Highlight Detection from Video Duration
• A Theory-Driven Self-Labeling Refinement Method for Contrastive Representation Learning
• Exploiting Web Images for Weakly Supervised Object Detection
• Robust and On-the-fly Dataset Denoising for Image Classification
• Conservative Predictions on Noisy Financial Data
• A Novel Self-Supervised Re-labeling Approach for Training with Noisy Labels
• MetaInfoNet: Learning Task-Guided Information for Sample Reweighting
• Robust Deep Graph Based Learning for Binary Classification
• Joint Negative and Positive Learning for Noisy Labels
• A Simple yet Effective Baseline for Robust Deep Learning with Noisy Labels
• Co-matching: Combating Noisy Labels by Augmentation Anchoring
• Robust Deep Learning with Active Noise Cancellation for Spatial Computing
• KNN-enhanced Deep Learning Against Noisy Labels
• Learning Deep Neural Networks under Agnostic Corrupted Supervision
• Data-driven Meta-set Based Fine-Grained Visual Classification
• Temporal Self-Ensembling Teacher for Semi-Supervised Object Detection
• Learning from Noisy Labels with Noise Modeling Network
• A Practical Framework for Relation Extraction with Noisy Labels Based on Doubly Transitional Loss
• Improving Medical Annotation Quality to Decrease Labeling Burden Using Stratified Noisy Cross-Validation
• Consistency Regularization Can Improve Robustness to Label Noise
• Multi-source weak supervision for saliency detection
• Decoupled Gradient Harmonized Detector for Partial Annotation: Application to Signet Ring Cell Detection
• Deep Learning for Earth Image Segmentation based on Imperfect Polyline Labels with Annotation Errors
• Ranking-Based Autoencoder for Extreme Multi-label Classification
• Simpler Certified Radius Maximization by Propagating Covariances
• Suppressing Mislabeled Data via Grouping and Self-Attention
• Robust Temporal Ensembling for Learning with Noisy Labels
• Mitigating Memorization in Sample Selection for Learning with Noisy Labels
• Joint Deep Learning of Facial Expression Synthesis and Recognition