Source-linked AI summary
Dynamic Curriculum Learning for Imbalanced Data Classification
Yiru Wang, Weihao Gan, Jie Yang, Wei Wu, Junjie Yan
TL;DR
Human attribute classification is challenged by severely imbalanced data, while fixed resampling and cost-sensitive approaches require problematic distribution or cost choices. DCL dynamically schedules sampling and the relative roles of classification and metric-learning losses, achieving state-of-the-art performance on CelebA and RAP.
Problem
Human attribute analysis involves largely imbalanced data, and common resampling or cost-sensitive methods require distribution or cost choices that are difficult to determine.
Method
DCL uses sampling and loss curriculum schedulers to move training from imbalanced to balanced and easy to hard while adjusting classification and metric-learning emphasis.
Results
DCL achieves new state-of-the-art performance on the CelebA face-attribute and RAP pedestrian-attribute datasets.
Takeaways & Limitations
The framework demonstrates generalization and discriminative power across two human-attribute analysis datasets.
Takeaways & Limitations
On CIFAR-100, class-balanced accuracy equals class-biased accuracy because every class has the same number of instances.
Abstract
from arXiv · showhide
Human attribute analysis is a challenging task in the field of computer vision, since the data is largely imbalance-distributed. Common techniques such as re-sampling and cost-sensitive learning require prior-knowledge to train the system. To address this problem, we propose a unified framework called Dynamic Curriculum Learning (DCL) to online adaptively adjust the sampling strategy and loss learning in single batch, which resulting in better generalization and discrimination. Inspired by the curriculum learning, DCL consists of two level curriculum schedulers: (1) sampling scheduler not only manages the data distribution from imbalanced to balanced but also from easy to hard; (2) loss scheduler controls the learning importance between classification and metric learning loss. Learning from these two schedulers, we demonstrate our DCL framework with the new state-of-the-art performance on the widely used face attribute dataset CelebA and pedestrian attribute dataset RAP.
1. Introduction
Human attribute analysis faces severe class imbalance that can bias classifiers toward majority classes. DCL addresses this with curriculum-based sampling and loss scheduling, achieving state-of-the-art results on CelebA and RAP.
- Motivation: Human attribute datasets are naturally imbalanced, so equal sample importance can bias models toward majority classes and reduce minority-class accuracy.For example, Bald images are typically labeled No Bald.
- Motivation: Resampling can overfit duplicated minority samples or discard majority information, while cost-sensitive learning requires selecting costs for different distributions.These limitations motivate adaptive rather than fixed imbalance handling.
- Motivation: Fixed balanced sampling may weaken majority-class representations early, so learning should progress from general representations toward classification.The intended outcome is a balance between class-biased and class-balanced accuracy.
- Method: DCL uses two curriculum schedulers: sampling moves data from imbalanced to balanced and easy to hard, while loss scheduling adjusts classification and metric-learning weights.Both schedulers reflect the model’s learning status.
- Contributions: DCL introduces curriculum learning to imbalanced classification, unifies several existing methods under scheduler settings, and reports state-of-the-art performance on CelebA and RAP.The paper presents these as its main contributions.
2. Related Work
Prior imbalanced-learning methods modify data, losses, decision thresholds, or combinations of these strategies. Curriculum learning has mainly been studied in other vision settings, leaving the paper’s curriculum-based imbalance framework as a distinct direction.
- Imbalanced data learning: Data-level methods rebalance distributions through minority oversampling or majority downsampling, but each can introduce overfitting or information loss.Oversampling repeats minority examples, whereas downsampling removes majority samples.
- Imbalanced data learning: Algorithm-level methods impose heavier costs on minority misclassification, but determining appropriate costs across settings remains an open question.Threshold adjustment changes the decision threshold during testing.
- Imbalanced data learning: Hybrid approaches combine data-level and algorithm-level techniques, including ensembling methods.The related-work discussion groups these methods separately from data-level and algorithm-level approaches.
- Deep imbalanced learning: Deep imbalanced learning integrates sampling and cost-learning into end-to-end frameworks, extending imbalance handling to deep models.Examples include combining under-sampling with SMOTE-based over-sampling.
- Curriculum learning: Curriculum learning has improved generalization by ordering learning from easy to hard, but prior applications mainly concern discovery, tracking, and semi- or weakly-supervised learning.The paper applies this learning principle to imbalanced data classification.
3. Method
DCL uses scheduler functions to coordinate dynamic sampling and loss weighting for imbalanced classification. Its curriculum progresses from imbalanced to balanced data, from easy to hard samples, and from metric embedding toward classification.
- DCL framework: DCL combines a sampling scheduler and a loss scheduler to adapt training for imbalanced data classification.The sampling scheduler controls sample selection and distribution, while the loss scheduler controls the relative importance of classification and metric learning losses.
- Loss Scheduler: The loss scheduler shifts emphasis from soft feature embedding early in training toward classification later in training.DCL combines the Dynamic Selective Learning loss with triplet loss and adjusts their learning importance over time.
- Scheduler Function Design: Scheduler functions decrease monotonically from 1 to 0 across training epochs and encode different curriculum learning speeds.The paper considers convex, linear, concave, and composite functions, representing slow-to-fast, constant, fast-to-slow, and slow-to-fast-to-slow schedules.
- Sampling Scheduler: The sampling scheduler changes each batch from the training-set distribution toward a balanced distribution as training proceeds.It dynamically selects majority-class samples and reweights minority-class samples according to the epoch-specific target distribution.
- Sampling Scheduler: At the beginning, the target distribution matches the real-world training distribution; by the final epoch, its class elements approach 1, indicating balance.The scheduler function satisfies g(0) = 1 initially and approaches 0 at the final epoch.
- Metric Learning with Easy Anchors: The metric-learning component uses easy minority samples as anchors instead of all minority samples to stabilize feature-space learning.Easy anchors are high-confidence correctly predicted minority samples; the method pulls hard positives closer and pushes hard negatives farther away.
4.1. Datasets
The experiments use CelebA, RAP, and CIFAR-100, spanning imbalanced facial and pedestrian attribute datasets plus a balanced natural-image classification dataset.
- CelebA contains 202,599 images from 10,177 people with 40 binary facial-attribute classifications.
- RAP contains 41,585 surveillance images from 26 indoor cameras with 72 pedestrian attributes.
- RAP has imbalance ratios up to 1:1800, while CelebA’s majority-class rate minus 50% reaches 48 for some attributes.
- CIFAR-100 is balanced, containing 50,000 training images and 10,000 test images across 100 equally represented classes.
4.2. Evaluation Metric
Evaluation on CelebA and RAP uses class-balanced accuracy for each binary task and averages task accuracies into an overall metric; balanced CIFAR-100 equates this with class-biased accuracy.
- CelebA and RAP are evaluated with class-balanced accuracy on every binary classification task, then averaged across tasks.
- For each task, mA_i combines true-positive and true-negative rates using predicted positives, ground-truth positives, and their opposite counts.
- CIFAR-100’s equal class sizes make class-balanced accuracy equal to class-biased accuracy.
- Table 2 reports class-balanced mean accuracy by CelebA attribute and imbalance level, highlighting the first- and second-best results.
4.3. Experiments on CelebA Face Dataset
CelebA experiments compare DCL with DeepID2 and several imbalanced-learning methods, examine performance across imbalance levels, and analyze scheduler and component choices. DCL improves overall CelebA performance, remains stable across attributes, and gains most on highly imbalanced data.
- Computational cost: DCL training takes 20.5 hours versus 20 hours for DeepID2 under the same 300 epochs.
- Overall performance: DCL improves approximately 8% over the DeepID2 cross-entropy baseline on CelebA.
- Overall performance: DCL outperforms Selective Learning by 3.12%, CRL-I by 2.45%, LMLE by 5.22%, and CLMLE by 0.27%.
- Overall performance: DCL treats attributes individually and defines triplet loss at attribute level, supporting multi-attribute learning with minimal extra computational time.
- Effect of data imbalance level: DCL’s performance gain is plotted against imbalance ratio for DCL, CRL, and CLMLE relative to DeepID2.
- Effect of data imbalance level: DCL consistently improves across attributes, while CRL and CLMLE degrade on some attributes; for Blurry, DCL gains 8% over CRL and 21% over CLMLE.
- Component analysis: The sampling scheduler shifts target distributions from imbalanced to balanced and from easy to hard, while triplet loss with easy anchors and the loss scheduler provide additional components.
- Scheduler function selection: The scheduler-function study finds linear decrease substantially worse than alternatives and favors a convex function for sampling.
4.4. Experiments on RAP Pedestrian Dataset
On RAP, DCL is evaluated across highly imbalanced pedestrian attributes and achieves strong performance against established methods and a ResNet-50 cross-entropy baseline.
- Overall Evaluation: DCL outperforms the previous best method, LG-Net, by 5% in average class-balanced mean accuracy on RAP.The comparison includes Deep-Mar, Inception-v2, HP-net, JRL, VeSPA, and LG-Net.
- Overall Evaluation: The framework is end-to-end and incurs small extra cost, unlike LG-Net and HP-net, which use classwise attention with greater training and inference resource requirements.The resource comparison is reported alongside the overall RAP evaluation.
- Effect of Data Imbalance Ratio: RAP contains 70 attributes with imbalance ratios reaching 1:1800, motivating evaluation across three imbalance-ratio groups.The imbalance ratio is defined as the minority-to-majority sample ratio.
- Effect of Data Imbalance Ratio: DCL improves over the ResNet-50 cross-entropy baseline by 3.8% for imbalance ratios from 1∼25.This is the least imbalanced of the three evaluated groups.
- Effect of Data Imbalance Ratio: DCL gains 15.0% for imbalance ratios from 25∼50 and 17.5% when the ratio exceeds 50.The larger gains occur in the more severely imbalanced groups.
4.5. Experiments on CIFAR-100 Dataset
On balanced CIFAR-100, DCL is tested for generalization beyond imbalanced attribute datasets and outperforms both the baseline and CRL.
- Generalization Evaluation: DCL improves over the cross-entropy baseline by 3.4% and over CRL by 2.2% on CIFAR-100.CIFAR-100 provides a balanced-data setting for the generalization check.
- Generalization Evaluation: The easy-anchor triplet loss stabilizes training, while the loss scheduler helps rectify the feature space for general classification.These mechanisms are presented as explanations for DCL’s stronger representation in the balanced evaluation.
- Generalization Evaluation: On balanced CIFAR-100, several compared methods are equivalent to the cross-entropy baseline, while LMLE and CLMLE lack reported generalization results.The comparison is made with the baseline and CRL.
5. Conclusion
The conclusion presents DCL as a unified imbalanced-learning framework that dynamically schedules sampling and loss learning, with strong results on CelebA and RAP.
- Conclusion: DCL introduces curriculum learning through two schedulers governing sampling and loss backward propagation.The schedulers manage learning from imbalance to balance and from easy to hard.
- Conclusion: The framework includes a metric-learning triplet loss with an easy anchor to improve feature embedding.This component complements the curriculum schedulers.
- Conclusion: DCL achieves new state-of-the-art performance on the CelebA and RAP attribute-analysis datasets.The conclusion associates these evaluations with the model’s generalization and discriminative power.
- Conclusion: DCL shows strong classification ability when data is largely imbalance-distributed.This is the conclusion’s specific emphasis regarding imbalanced data.