Coupled Confusion Correction: Learning from Crowds with Sparse Annotations

TL;DR

This paper tackles learning from crowds with sparse, noisy annotations by introducing Coupled Confusion Correction (CCC), a bi-level, meta-learning framework that trains two models to jointly correct annotator confusion matrices. CCC groups annotators by expertise and distills supervision from a meta set via a two-model distillation scheme to mitigate sparsity and confirmation bias. The method achieves superior performance over state-of-the-art baselines on both synthetic (independent and correlated confusions) and real-world datasets (LabelMe, CIFAR-10N, MUSIC), even approaching clean-label performance in some cases. These results demonstrate CCC’s practical potential for robust learning from crowds in large-scale, real-world applications.

Abstract

As the size of the datasets getting larger, accurately annotating such datasets is becoming more impractical due to the expensiveness on both time and economy. Therefore, crowd-sourcing has been widely adopted to alleviate the cost of collecting labels, which also inevitably introduces label noise and eventually degrades the performance of the model. To learn from crowd-sourcing annotations, modeling the expertise of each annotator is a common but challenging paradigm, because the annotations collected by crowd-sourcing are usually highly-sparse. To alleviate this problem, we propose Coupled Confusion Correction (CCC), where two models are simultaneously trained to correct the confusion matrices learned by each other. Via bi-level optimization, the confusion matrices learned by one model can be corrected by the distilled data from the other. Moreover, we cluster the ``annotator groups'' who share similar expertise so that their confusion matrices could be corrected together. In this way, the expertise of the annotators, especially of those who provide seldom labels, could be better captured. Remarkably, we point out that the annotation sparsity not only means the average number of labels is low, but also there are always some annotators who provide very few labels, which is neglected by previous works when constructing synthetic crowd-sourcing annotations. Based on that, we propose to use Beta distribution to control the generation of the crowd-sourcing labels so that the synthetic annotations could be more consistent with the real-world ones. Extensive experiments are conducted on two types of synthetic datasets and three real-world datasets, the results of which demonstrate that CCC significantly outperforms state-of-the-art approaches. Source codes are available at: https://github.com/Hansong-Zhang/CCC.

Figures (4)

• Figure 1: The histogram of the number of labels provided by annotators. The green patch on the left denotes the annotators who only provide seldom labels.
• Figure 2: The illustration of the our method. "N/A" means the label is missing.
• Figure 3: The distances of true confusion matrices of every two annotators in LabelMe and CIFAR-10N datasets. The $\#$ denotes the index of annotators. The more blue (red) the pixel is, the smaller (larger) value it is.
• Figure 4: The ablation studies on sparsity level (top) and number of annotator groups (bottom). The horizontal axis represents average number of annotators (top) and number of annotator groups (bottom), while the vertical axis denotes test accuracy.