Paper link:

https://arxiv.org/abs/2201.08984

Code link:

https://github.com/hbzju/PiCO

The training of modern deep neural networks usually requires a large amount of labeled data, which brings great difficulties to the collection of training data.This is because labeling real-world data (data annotation) has the effect of label ambiguity and noise.
As shown in Figure 1, for a human annotator to label an image, selecting the correct label, Alaskan sled, from its candidate label set {husky, alaskan, samoyed} isIt's difficult because these types of dogs look very similar.The problem of label ambiguity is prevalent in many applications, but is often overlooked.This highlightsThe importance of partial label learning (PLL).
Although PLLs have promising applications, they suffer from label ambiguity.Therefore, a core challenge of PLL is label disambiguation, that is, identifying the correct ground-truth label from the candidate label set.Some existing label disambiguation models usually assume that data points with similar distances in the feature space are more likely to share the same correct label, which leads to the fact that they often require a better feature representation to achieve better results.Performance.
This reliance on representation creates a major non-trivial dilemma:Label uncertainty can adversely affect the process of representation learning, resulting in poor representations, which in turn impede effective label disambiguation (label disambiguation). So far, little has been done to address this issue.Based on this, this paper seeks to address this dilemma by reconciling the inherent contradiction between these two interdependent parts: representation learning and label disambiguation.