Since that subject is mentioned on the Hosts (Call for Papers) Discussion Topic, I tried to bring some notes/papers about it.
\nHow to Tame the Long Tail in Machine Learning - By Sasha Harrison on June 29th, 2021
\n\"Many AI systems rely on supervised learning methods in which neural networks train on labeled data. The challenge with supervised methods is getting models to perform well on examples not adequately represented in the training dataset.\"
\n\"Typically, as the frequency of a particular category decreases, so does average model performance on this category. It is often difficult and costly to achieve strong performance on the rare edge cases that make up the long tail of a data distribution. In that blog post, the authors will take a deeper look at how sophisticated data curation tools can help machine learning teams target their experiments toward taming the long tail.\"
\nhttps://scale.com/blog/taming-long-tail
\nAuthors: Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, Sanjiv Kumar 28 Sept 2020 (modified: 10 Feb 2022)
\n\"Real-world classification problems typically exhibit an imbalanced or long-tailed label distribution, wherein many labels have only a few associated samples. This poses a challenge for generalisation on such labels, and also makes naive learning biased towards dominant labels.\"
\n\"In that paper, the authors presented a statistical framework that unifies and generalises several recent proposals to cope with these challenges. Their framework revisited the classic idea of logit adjustment based on the label frequencies, which encourages a large relative margin between logits of rare positive versus dominant negative labels. This yields two techniques for long-tail learning, where such adjustment is either applied post-hoc to a trained model, or enforced in the loss during training.\"
\n\"Those techniques are statistically grounded, and practically effective on four real-world datasets with long-tailed label distributions.\"
\nhttps://openreview.net/forum?id=37nvvqkCo5
\nTST: Two-Stage Training
\nIS: Instance Sampling
\nCBS: Class-Balanced Sampling
\nCLW: Class-Level Weighting
\nNC: Normalized Classifier
\nENS: Ensemble
\nDA: Data Augmentation
\n\"Information augmentation comprises transfer learning (TL) and data augmentation (Aug). Module improvement includes representation learning (RL), classifier design (CD), decoupled training (DT) and ensemble learning (Ensemble). According to this taxonomy, the authors sorted out existing deep long-tailed learning methods and will review them in detail.\"
\nhttps://arxiv.org/pdf/2110.04596.pdf
\n\"In recent years, a variety of visual datasets have been released for long-tailed learning, differing in tasks, class numbers and sample numbers. The authors summarized nine visual datasets that are widely used in the deep long-tailed learning community.\"
\n\"In long-tailed image classification, there are four benchmark datasets: ImageNet-LT , CIFAR100-LT , Places-LT , and iNaturalist 2018 . The previous three are sampled from ImageNet , CIFAR100 and Places365 following Pareto distributions, respectively, while iNaturalist is a real-world long-tailed dataset.\"
\n\"The imbalance ratio of ImageNet-LT, Places-LT and iNaturalist are 256, 996 and 500, respectively; CIFAR100-LT has three variants with various imbalance ratios {10, 50, 100}. In long-tailed object detection and instance segmentation, LVIS , providing precise bounding box and mask annotations, is the widely-used benchmark. In multi-label image classification, the benchmarks are VOC-LT and COCO-LT , which are sampled from PASCAL VOC 2012 and COCO, respectively.\"
\n\"Recently, a large-scale “untrimmed” video dataset, namely VideoLT , was released for long-tailed video recognition.\"
\nhttps://arxiv.org/pdf/2110.04596.pdf
\n\"In long-tailed learning, the overall performance on all classes and the performance for head, middle and tail classes are usually reported. The used evaluation metrics differ in various tasks. For example, Top-1 Accuracy (or Error Rate) is the widely-used metric for long-tailed image classification, while mean Average Precision (mAP) is adopted for long-tailed object detection and instance segmentation. Moreover, mAP is also used in long-tailed multilabel image classification as a metric, while video recognition applies both Top-1 Accuracy and mAP for evaluation.\"
\nhttps://arxiv.org/pdf/2110.04596.pdf
\n\"Information augmentation based methods seek to introduce additional information into model training, so that the model performance can be improved in long-tailed learning. There are two kinds of methods in this method type: transfer learning and data augmentation.\"
\nTRASFER LEARNING
\n\"Transfer learning seeks to transfer the knowledge from a source domain (e.g., datasets, tasks or classes) to enhance model training on a target domain. In deep long-tailed learning, there are four main transfer learning schemes, i.e., head-to tail knowledge transfer, model pre-training, knowledge distillation, and self-training.\"
\nhttps://arxiv.org/pdf/2110.04596.pdf
\n\"The most common application of long-tailed learning is multiclass classification. There are many artificially sampled long-tailed datasets from widely-used image classification datasets, i.e., ImageNet, CIFAR, and Places. Based on these datasets, various long-tailed learning methods have been proposed.\"
\n\"Besides these artificial tasks, long-tailed learning is also applied to real-world image classification tasks, including species classification, face recognition, age classification, logo detection, rail surface defect detection and medical image diagnosis.\"
\n\"In addition to multi-class classification, long-tailed learning is also applied to multi-label classification based on both artificial tasks (i.e., VOC-LT and COCO-LT) and real-world tasks, including web image classification, face attribute classification and cloth attribute classification.\"
\nIMAGE DETECTION and SEGMENTATION
\n\"Object detection and instance segmentation has attracted increasing attention in the long-tailed learning community, where most existing studies are conducted based on LVIS and COCO. In addition to these widely used benchmarks, many other applications have also been explored, including urban scene understanding, unmanned aerial vehicle detection, point cloud segmentation\"
\nhttps://arxiv.org/pdf/2110.04596.pdf
\nLearning from Long-Tailed Data with Noisy Labels https://arxiv.org/pdf/2108.11096.pdf
\nSelf-supervised Learning is More Robust to Dataset Imbalance https://arxiv.org/pdf/2110.05025.pdf
\nAdaptive Logit Adjustment Loss for Long-Tailed Visual Recognition https://arxiv.org/pdf/2104.06094.pdf
\nBalanced Knowledge Distillation for Long-tailed Learning https://arxiv.org/pdf/2104.10510.pdf
\nAdversarial Robustness under Long-Tailed Distribution https://arxiv.org/pdf/2104.02703.pdf
\nLearning From Multiple Experts: Self-paced Knowledge Distillation for Long-tailed Classification https://arxiv.org/pdf/2001.01536.pdf
\nDecoupling Representation and Classifier for Long-Tailed Recognition https://openreview.net/pdf?id=r1gRTCVFvB
\nBBN: Bilateral-Branch Network with Cumulative Learning for Long-Tailed Visual Recognition https://openaccess.thecvf.com/content_CVPR_2020/papers/Zhou_BBN_Bilateral-Branch_Network_With_Cumulative_Learning_for_Long-Tailed_Visual_Recognition_CVPR_2020_paper.pdf
\niNaturalist 2018 Competition https://github.com/visipedia/inat_comp/tree/master/2018
\nDeep Representation Learning on Long-tailed Data: A Learnable Embedding Augmentation Perspective https://openaccess.thecvf.com/content_CVPR_2020/papers/Liu_Deep_Representation_Learning_on_Long-Tailed_Data_A_Learnable_Embedding_Augmentation_CVPR_2020_paper.pdf
\nThe Devil is in the Tails: Fine-grained Classification in the Wild https://arxiv.org/pdf/1709.01450.pdf
\nSource: https://github.com/Stomach-ache/awesome-long-tailed-learning
", "messages": [ { "id": 1700277, "postDate": "2022-02-21T19:37:06.130Z", "content": "Since that subject is mentioned on the Hosts (Call for Papers) Discussion Topic, I tried to bring some notes/papers about it.
\nHow to Tame the Long Tail in Machine Learning - By Sasha Harrison on June 29th, 2021
\n\"Many AI systems rely on supervised learning methods in which neural networks train on labeled data. The challenge with supervised methods is getting models to perform well on examples not adequately represented in the training dataset.\"
\n\"Typically, as the frequency of a particular category decreases, so does average model performance on this category. It is often difficult and costly to achieve strong performance on the rare edge cases that make up the long tail of a data distribution. In that blog post, the authors will take a deeper look at how sophisticated data curation tools can help machine learning teams target their experiments toward taming the long tail.\"
\nhttps://scale.com/blog/taming-long-tail
\nAuthors: Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, Sanjiv Kumar 28 Sept 2020 (modified: 10 Feb 2022)
\n\"Real-world classification problems typically exhibit an imbalanced or long-tailed label distribution, wherein many labels have only a few associated samples. This poses a challenge for generalisation on such labels, and also makes naive learning biased towards dominant labels.\"
\n\"In that paper, the authors presented a statistical framework that unifies and generalises several recent proposals to cope with these challenges. Their framework revisited the classic idea of logit adjustment based on the label frequencies, which encourages a large relative margin between logits of rare positive versus dominant negative labels. This yields two techniques for long-tail learning, where such adjustment is either applied post-hoc to a trained model, or enforced in the loss during training.\"
\n\"Those techniques are statistically grounded, and practically effective on four real-world datasets with long-tailed label distributions.\"
\nhttps://openreview.net/forum?id=37nvvqkCo5
\nTST: Two-Stage Training
\nIS: Instance Sampling
\nCBS: Class-Balanced Sampling
\nCLW: Class-Level Weighting
\nNC: Normalized Classifier
\nENS: Ensemble
\nDA: Data Augmentation
\n\"Information augmentation comprises transfer learning (TL) and data augmentation (Aug). Module improvement includes representation learning (RL), classifier design (CD), decoupled training (DT) and ensemble learning (Ensemble). According to this taxonomy, the authors sorted out existing deep long-tailed learning methods and will review them in detail.\"
\nhttps://arxiv.org/pdf/2110.04596.pdf
\n\"In recent years, a variety of visual datasets have been released for long-tailed learning, differing in tasks, class numbers and sample numbers. The authors summarized nine visual datasets that are widely used in the deep long-tailed learning community.\"
\n\"In long-tailed image classification, there are four benchmark datasets: ImageNet-LT , CIFAR100-LT , Places-LT , and iNaturalist 2018 . The previous three are sampled from ImageNet , CIFAR100 and Places365 following Pareto distributions, respectively, while iNaturalist is a real-world long-tailed dataset.\"
\n\"The imbalance ratio of ImageNet-LT, Places-LT and iNaturalist are 256, 996 and 500, respectively; CIFAR100-LT has three variants with various imbalance ratios {10, 50, 100}. In long-tailed object detection and instance segmentation, LVIS , providing precise bounding box and mask annotations, is the widely-used benchmark. In multi-label image classification, the benchmarks are VOC-LT and COCO-LT , which are sampled from PASCAL VOC 2012 and COCO, respectively.\"
\n\"Recently, a large-scale “untrimmed” video dataset, namely VideoLT , was released for long-tailed video recognition.\"
\nhttps://arxiv.org/pdf/2110.04596.pdf
\n\"In long-tailed learning, the overall performance on all classes and the performance for head, middle and tail classes are usually reported. The used evaluation metrics differ in various tasks. For example, Top-1 Accuracy (or Error Rate) is the widely-used metric for long-tailed image classification, while mean Average Precision (mAP) is adopted for long-tailed object detection and instance segmentation. Moreover, mAP is also used in long-tailed multilabel image classification as a metric, while video recognition applies both Top-1 Accuracy and mAP for evaluation.\"
\nhttps://arxiv.org/pdf/2110.04596.pdf
\n\"Information augmentation based methods seek to introduce additional information into model training, so that the model performance can be improved in long-tailed learning. There are two kinds of methods in this method type: transfer learning and data augmentation.\"
\nTRASFER LEARNING
\n\"Transfer learning seeks to transfer the knowledge from a source domain (e.g., datasets, tasks or classes) to enhance model training on a target domain. In deep long-tailed learning, there are four main transfer learning schemes, i.e., head-to tail knowledge transfer, model pre-training, knowledge distillation, and self-training.\"
\nhttps://arxiv.org/pdf/2110.04596.pdf
\n\"The most common application of long-tailed learning is multiclass classification. There are many artificially sampled long-tailed datasets from widely-used image classification datasets, i.e., ImageNet, CIFAR, and Places. Based on these datasets, various long-tailed learning methods have been proposed.\"
\n\"Besides these artificial tasks, long-tailed learning is also applied to real-world image classification tasks, including species classification, face recognition, age classification, logo detection, rail surface defect detection and medical image diagnosis.\"
\n\"In addition to multi-class classification, long-tailed learning is also applied to multi-label classification based on both artificial tasks (i.e., VOC-LT and COCO-LT) and real-world tasks, including web image classification, face attribute classification and cloth attribute classification.\"
\nIMAGE DETECTION and SEGMENTATION
\n\"Object detection and instance segmentation has attracted increasing attention in the long-tailed learning community, where most existing studies are conducted based on LVIS and COCO. In addition to these widely used benchmarks, many other applications have also been explored, including urban scene understanding, unmanned aerial vehicle detection, point cloud segmentation\"
\nhttps://arxiv.org/pdf/2110.04596.pdf
\nLearning from Long-Tailed Data with Noisy Labels https://arxiv.org/pdf/2108.11096.pdf
\nSelf-supervised Learning is More Robust to Dataset Imbalance https://arxiv.org/pdf/2110.05025.pdf
\nAdaptive Logit Adjustment Loss for Long-Tailed Visual Recognition https://arxiv.org/pdf/2104.06094.pdf
\nBalanced Knowledge Distillation for Long-tailed Learning https://arxiv.org/pdf/2104.10510.pdf
\nAdversarial Robustness under Long-Tailed Distribution https://arxiv.org/pdf/2104.02703.pdf
\nLearning From Multiple Experts: Self-paced Knowledge Distillation for Long-tailed Classification https://arxiv.org/pdf/2001.01536.pdf
\nDecoupling Representation and Classifier for Long-Tailed Recognition https://openreview.net/pdf?id=r1gRTCVFvB
\nBBN: Bilateral-Branch Network with Cumulative Learning for Long-Tailed Visual Recognition https://openaccess.thecvf.com/content_CVPR_2020/papers/Zhou_BBN_Bilateral-Branch_Network_With_Cumulative_Learning_for_Long-Tailed_Visual_Recognition_CVPR_2020_paper.pdf
\niNaturalist 2018 Competition https://github.com/visipedia/inat_comp/tree/master/2018
\nDeep Representation Learning on Long-tailed Data: A Learnable Embedding Augmentation Perspective https://openaccess.thecvf.com/content_CVPR_2020/papers/Liu_Deep_Representation_Learning_on_Long-Tailed_Data_A_Learnable_Embedding_Augmentation_CVPR_2020_paper.pdf
\nThe Devil is in the Tails: Fine-grained Classification in the Wild https://arxiv.org/pdf/1709.01450.pdf
\nSource: https://github.com/Stomach-ache/awesome-long-tailed-learning
", "rawMarkdown": "Since that subject is mentioned on the Hosts (Call for Papers) Discussion Topic, I tried to bring some notes/papers about it.\n\n#Taming The Long Tail in Machine Learning\n\nHow to Tame the Long Tail in Machine Learning - By Sasha Harrison on June 29th, 2021\n\n\"Many AI systems rely on supervised learning methods in which neural networks train on labeled data. The challenge with supervised methods is getting models to perform well on examples not adequately represented in the training dataset.\"\n\n\"Typically, as the frequency of a particular category decreases, so does average model performance on this category. It is often difficult and costly to achieve strong performance on the rare edge cases that make up the long tail of a data distribution. In that blog post, the authors will take a deeper look at how sophisticated data curation tools can help machine learning teams target their experiments toward taming the long tail.\"\n\nhttps://scale.com/blog/taming-long-tail\n\n#Long-tail learning via logit adjustment\n\nAuthors: Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, Sanjiv Kumar 28 Sept 2020 (modified: 10 Feb 2022)\n\n\"Real-world classification problems typically exhibit an imbalanced or long-tailed label distribution, wherein many labels have only a few associated samples. This poses a challenge for generalisation on such labels, and also makes naive learning biased towards dominant labels.\"\n\n\"In that paper, the authors presented a statistical framework that unifies and generalises several recent proposals to cope with these challenges. Their framework revisited the classic idea of logit adjustment based on the label frequencies, which encourages a large relative margin between logits of rare positive versus dominant negative labels. This yields two techniques for long-tail learning, where such adjustment is either applied post-hoc to a trained model, or enforced in the loss during training.\"\n\n\"Those techniques are statistically grounded, and practically effective on four real-world datasets with long-tailed label distributions.\"\n\nhttps://openreview.net/forum?id=37nvvqkCo5\n\n\n#Types of Long-Tailed Learning Methods\n\nTST: Two-Stage Training\n\nIS: Instance Sampling\n\nCBS: Class-Balanced Sampling\n\nCLW: Class-Level Weighting\n\nNC: Normalized Classifier\n\nENS: Ensemble\n\nDA: Data Augmentation\n\n\"Information augmentation comprises transfer learning (TL) and data augmentation (Aug). Module improvement includes representation learning (RL), classifier design (CD), decoupled training (DT) and ensemble learning (Ensemble). According to this taxonomy, the authors sorted out existing deep long-tailed learning methods and will review them in detail.\"\n\nhttps://arxiv.org/pdf/2110.04596.pdf\n\n#Datasets used in the Deep Long-Tailed Learning\n\n\"In recent years, a variety of visual datasets have been released for long-tailed learning, differing in tasks, class numbers and sample numbers. The authors summarized nine visual datasets that are widely used in the deep long-tailed learning community.\"\n\n\"In long-tailed image classification, there are four benchmark datasets: ImageNet-LT , CIFAR100-LT , Places-LT , and iNaturalist 2018 . The previous three are sampled from ImageNet , CIFAR100 and Places365 following Pareto distributions, respectively, while iNaturalist is a real-world long-tailed dataset.\"\n\n\"The imbalance ratio of ImageNet-LT, Places-LT and iNaturalist are 256, 996 and 500, respectively; CIFAR100-LT has three variants with various imbalance ratios {10, 50, 100}. In long-tailed object detection and instance segmentation, LVIS , providing precise bounding box and mask annotations, is the widely-used benchmark. In multi-label image classification, the benchmarks are VOC-LT and COCO-LT , which are sampled from PASCAL VOC 2012 and COCO, respectively.\"\n\n\"Recently, a large-scale “untrimmed” video dataset, namely VideoLT , was released for long-tailed video recognition.\"\n\nhttps://arxiv.org/pdf/2110.04596.pdf\n\n#Evaluation Metrics in the Deep Long-Tailed Learning\n\n\"In long-tailed learning, the overall performance on all classes and the performance for head, middle and tail classes are usually reported. The used evaluation metrics differ in various tasks. For example, Top-1 Accuracy (or Error Rate) is the widely-used metric for long-tailed image classification, while mean Average Precision (mAP) is adopted for long-tailed object detection and instance segmentation. Moreover, mAP is also used in long-tailed multilabel image classification as a metric, while video recognition applies both Top-1 Accuracy and mAP for evaluation.\"\n\nhttps://arxiv.org/pdf/2110.04596.pdf\n\n\n#Information Augmentation in The Deep Long-Tailed Learning\n\n\"Information augmentation based methods seek to introduce additional information into model training, so that the model performance can be improved in long-tailed learning. There are two kinds of methods in this method type: transfer learning and data augmentation.\"\n\nTRASFER LEARNING\n\n\"Transfer learning seeks to transfer the knowledge from a source domain (e.g., datasets, tasks or classes) to enhance model training on a target domain. In deep long-tailed learning, there are four main transfer learning schemes, i.e., head-to tail knowledge transfer, model pre-training, knowledge distillation, and self-training.\"\n\nhttps://arxiv.org/pdf/2110.04596.pdf\n\n#Image Classification\n\n\"The most common application of long-tailed learning is multiclass classification. There are many artificially sampled long-tailed datasets from widely-used image classification datasets, i.e., ImageNet, CIFAR, and Places. Based on these datasets, various long-tailed learning methods have been proposed.\"\n\n\"Besides these artificial tasks, long-tailed learning is also applied to real-world image classification tasks, including species classification, face recognition, age classification, logo detection, rail surface defect detection and medical image diagnosis.\"\n\n\"In addition to multi-class classification, long-tailed learning is also applied to multi-label classification based on both artificial tasks (i.e., VOC-LT and COCO-LT) and real-world tasks, including web image classification, face attribute classification and cloth attribute classification.\"\n\nIMAGE DETECTION and SEGMENTATION\n\n\"Object detection and instance segmentation has attracted increasing attention in the long-tailed learning community, where most existing studies are conducted based on LVIS and COCO. In addition to these widely used benchmarks, many other applications have also been explored, including urban scene understanding, unmanned aerial vehicle detection, point cloud segmentation\"\n\nhttps://arxiv.org/pdf/2110.04596.pdf\n\n#Long-tailed Learning Papers\n\nLearning from Long-Tailed Data with Noisy Labels https://arxiv.org/pdf/2108.11096.pdf\n\nSelf-supervised Learning is More Robust to Dataset Imbalance https://arxiv.org/pdf/2110.05025.pdf\n\nAdaptive Logit Adjustment Loss for Long-Tailed Visual Recognition https://arxiv.org/pdf/2104.06094.pdf\n\nBalanced Knowledge Distillation for Long-tailed Learning https://arxiv.org/pdf/2104.10510.pdf\n\nAdversarial Robustness under Long-Tailed Distribution https://arxiv.org/pdf/2104.02703.pdf\n\nLearning From Multiple Experts: Self-paced Knowledge Distillation for Long-tailed Classification https://arxiv.org/pdf/2001.01536.pdf\n\nDecoupling Representation and Classifier for Long-Tailed Recognition https://openreview.net/pdf?id=r1gRTCVFvB\n\nBBN: Bilateral-Branch Network with Cumulative Learning for Long-Tailed Visual Recognition https://openaccess.thecvf.com/content_CVPR_2020/papers/Zhou_BBN_Bilateral-Branch_Network_With_Cumulative_Learning_for_Long-Tailed_Visual_Recognition_CVPR_2020_paper.pdf\n\niNaturalist 2018 Competition https://github.com/visipedia/inat_comp/tree/master/2018\n\nDeep Representation Learning on Long-tailed Data: A Learnable Embedding Augmentation Perspective https://openaccess.thecvf.com/content_CVPR_2020/papers/Liu_Deep_Representation_Learning_on_Long-Tailed_Data_A_Learnable_Embedding_Augmentation_CVPR_2020_paper.pdf\n\nThe Devil is in the Tails: Fine-grained Classification in the Wild https://arxiv.org/pdf/1709.01450.pdf\n\nSource: https://github.com/Stomach-ache/awesome-long-tailed-learning", "votes": 6 }, { "id": 1723593, "postDate": "2022-03-15T14:45:23.620Z", "content": "Thank you for sharing these ideas.
\nLove the two stage training procedure.
I learned something that I've never heard before about it.
\nThank you Med Ali for your participation. It means a lot to me.
", "rawMarkdown": "I learned something that I've never heard before about it.\n\nThank you Med Ali for your participation. It means a lot to me.", "votes": 1 } ] } ], "comments": [ { "id": 1723593, "author_name": "Med Ali Bouchhioua", "author_url": "", "post_date": "2022-03-15T14:45:23.620000", "content": "Thank you for sharing these ideas.
\nLove the two stage training procedure.
I learned something that I've never heard before about it.
\nThank you Med Ali for your participation. It means a lot to me.
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "1700277": "Since that subject is mentioned on the Hosts (Call for Papers) Discussion Topic, I tried to bring some notes/papers about it.\n\n#Taming The Long Tail in Machine Learning\n\nHow to Tame the Long Tail in Machine Learning - By Sasha Harrison on June 29th, 2021\n\n\"Many AI systems rely on supervised learning methods in which neural networks train on labeled data. The challenge with supervised methods is getting models to perform well on examples not adequately represented in the training dataset.\"\n\n\"Typically, as the frequency of a particular category decreases, so does average model performance on this category. It is often difficult and costly to achieve strong performance on the rare edge cases that make up the long tail of a data distribution. In that blog post, the authors will take a deeper look at how sophisticated data curation tools can help machine learning teams target their experiments toward taming the long tail.\"\n\nhttps://scale.com/blog/taming-long-tail\n\n#Long-tail learning via logit adjustment\n\nAuthors: Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, Sanjiv Kumar 28 Sept 2020 (modified: 10 Feb 2022)\n\n\"Real-world classification problems typically exhibit an imbalanced or long-tailed label distribution, wherein many labels have only a few associated samples. This poses a challenge for generalisation on such labels, and also makes naive learning biased towards dominant labels.\"\n\n\"In that paper, the authors presented a statistical framework that unifies and generalises several recent proposals to cope with these challenges. Their framework revisited the classic idea of logit adjustment based on the label frequencies, which encourages a large relative margin between logits of rare positive versus dominant negative labels. This yields two techniques for long-tail learning, where such adjustment is either applied post-hoc to a trained model, or enforced in the loss during training.\"\n\n\"Those techniques are statistically grounded, and practically effective on four real-world datasets with long-tailed label distributions.\"\n\nhttps://openreview.net/forum?id=37nvvqkCo5\n\n\n#Types of Long-Tailed Learning Methods\n\nTST: Two-Stage Training\n\nIS: Instance Sampling\n\nCBS: Class-Balanced Sampling\n\nCLW: Class-Level Weighting\n\nNC: Normalized Classifier\n\nENS: Ensemble\n\nDA: Data Augmentation\n\n\"Information augmentation comprises transfer learning (TL) and data augmentation (Aug). Module improvement includes representation learning (RL), classifier design (CD), decoupled training (DT) and ensemble learning (Ensemble). According to this taxonomy, the authors sorted out existing deep long-tailed learning methods and will review them in detail.\"\n\nhttps://arxiv.org/pdf/2110.04596.pdf\n\n#Datasets used in the Deep Long-Tailed Learning\n\n\"In recent years, a variety of visual datasets have been released for long-tailed learning, differing in tasks, class numbers and sample numbers. The authors summarized nine visual datasets that are widely used in the deep long-tailed learning community.\"\n\n\"In long-tailed image classification, there are four benchmark datasets: ImageNet-LT , CIFAR100-LT , Places-LT , and iNaturalist 2018 . The previous three are sampled from ImageNet , CIFAR100 and Places365 following Pareto distributions, respectively, while iNaturalist is a real-world long-tailed dataset.\"\n\n\"The imbalance ratio of ImageNet-LT, Places-LT and iNaturalist are 256, 996 and 500, respectively; CIFAR100-LT has three variants with various imbalance ratios {10, 50, 100}. In long-tailed object detection and instance segmentation, LVIS , providing precise bounding box and mask annotations, is the widely-used benchmark. In multi-label image classification, the benchmarks are VOC-LT and COCO-LT , which are sampled from PASCAL VOC 2012 and COCO, respectively.\"\n\n\"Recently, a large-scale “untrimmed” video dataset, namely VideoLT , was released for long-tailed video recognition.\"\n\nhttps://arxiv.org/pdf/2110.04596.pdf\n\n#Evaluation Metrics in the Deep Long-Tailed Learning\n\n\"In long-tailed learning, the overall performance on all classes and the performance for head, middle and tail classes are usually reported. The used evaluation metrics differ in various tasks. For example, Top-1 Accuracy (or Error Rate) is the widely-used metric for long-tailed image classification, while mean Average Precision (mAP) is adopted for long-tailed object detection and instance segmentation. Moreover, mAP is also used in long-tailed multilabel image classification as a metric, while video recognition applies both Top-1 Accuracy and mAP for evaluation.\"\n\nhttps://arxiv.org/pdf/2110.04596.pdf\n\n\n#Information Augmentation in The Deep Long-Tailed Learning\n\n\"Information augmentation based methods seek to introduce additional information into model training, so that the model performance can be improved in long-tailed learning. There are two kinds of methods in this method type: transfer learning and data augmentation.\"\n\nTRASFER LEARNING\n\n\"Transfer learning seeks to transfer the knowledge from a source domain (e.g., datasets, tasks or classes) to enhance model training on a target domain. In deep long-tailed learning, there are four main transfer learning schemes, i.e., head-to tail knowledge transfer, model pre-training, knowledge distillation, and self-training.\"\n\nhttps://arxiv.org/pdf/2110.04596.pdf\n\n#Image Classification\n\n\"The most common application of long-tailed learning is multiclass classification. There are many artificially sampled long-tailed datasets from widely-used image classification datasets, i.e., ImageNet, CIFAR, and Places. Based on these datasets, various long-tailed learning methods have been proposed.\"\n\n\"Besides these artificial tasks, long-tailed learning is also applied to real-world image classification tasks, including species classification, face recognition, age classification, logo detection, rail surface defect detection and medical image diagnosis.\"\n\n\"In addition to multi-class classification, long-tailed learning is also applied to multi-label classification based on both artificial tasks (i.e., VOC-LT and COCO-LT) and real-world tasks, including web image classification, face attribute classification and cloth attribute classification.\"\n\nIMAGE DETECTION and SEGMENTATION\n\n\"Object detection and instance segmentation has attracted increasing attention in the long-tailed learning community, where most existing studies are conducted based on LVIS and COCO. In addition to these widely used benchmarks, many other applications have also been explored, including urban scene understanding, unmanned aerial vehicle detection, point cloud segmentation\"\n\nhttps://arxiv.org/pdf/2110.04596.pdf\n\n#Long-tailed Learning Papers\n\nLearning from Long-Tailed Data with Noisy Labels https://arxiv.org/pdf/2108.11096.pdf\n\nSelf-supervised Learning is More Robust to Dataset Imbalance https://arxiv.org/pdf/2110.05025.pdf\n\nAdaptive Logit Adjustment Loss for Long-Tailed Visual Recognition https://arxiv.org/pdf/2104.06094.pdf\n\nBalanced Knowledge Distillation for Long-tailed Learning https://arxiv.org/pdf/2104.10510.pdf\n\nAdversarial Robustness under Long-Tailed Distribution https://arxiv.org/pdf/2104.02703.pdf\n\nLearning From Multiple Experts: Self-paced Knowledge Distillation for Long-tailed Classification https://arxiv.org/pdf/2001.01536.pdf\n\nDecoupling Representation and Classifier for Long-Tailed Recognition https://openreview.net/pdf?id=r1gRTCVFvB\n\nBBN: Bilateral-Branch Network with Cumulative Learning for Long-Tailed Visual Recognition https://openaccess.thecvf.com/content_CVPR_2020/papers/Zhou_BBN_Bilateral-Branch_Network_With_Cumulative_Learning_for_Long-Tailed_Visual_Recognition_CVPR_2020_paper.pdf\n\niNaturalist 2018 Competition https://github.com/visipedia/inat_comp/tree/master/2018\n\nDeep Representation Learning on Long-tailed Data: A Learnable Embedding Augmentation Perspective https://openaccess.thecvf.com/content_CVPR_2020/papers/Liu_Deep_Representation_Learning_on_Long-Tailed_Data_A_Learnable_Embedding_Augmentation_CVPR_2020_paper.pdf\n\nThe Devil is in the Tails: Fine-grained Classification in the Wild https://arxiv.org/pdf/1709.01450.pdf\n\nSource: https://github.com/Stomach-ache/awesome-long-tailed-learning", "1723593": "Thank you for sharing these ideas.\nLove the two stage training procedure." } }