Learning the Latent Noisy Data Generative Process for Label-Noise Learning

Yexiong Lin, Yu Yao, Tongliang Liu

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Learning with noisy labels, noise transition, image classification

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

In learning with noisy labels, the noise transition reveals how an instance relates from its clean label to its noisy one. Accurately inferring an instance's noise transition is crucial for inferring its clean label. However, when only a noisy dataset is available, noise transitions can typically be inferred only for a ``special'' group of instances. To use these learned transitions to assist in inferring others, it is essential to understand the connections among different transitions across different instances.
Existing work usually addresses this by introducing assumptions that explicitly define the similarity of noise transitions across various instances. However, these similarity-based assumptions often lack empirical validation and may not be aligned with real-world data. The misalignment can lead to misinterpretations of both noise transitions and clean labels.
In this work, instead of directly defining similarity, we propose modeling the generative process of noisy data. Intuitively, to understand the connections among noise transitions across different instances, we represent the causal generative process of noisy data using a learnable graphical model. Relying solely on noisy data, our method can effectively discern the underlying causal generative process, subsequently inferring the noise transitions of instances and their clean labels. Experiments on various datasets with different types of label noise further demonstrate our method's effectiveness.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7806