SAE: Estimation for Transition Matrix in Annotation Algorithms
Abstract: The transition matrix plays a critical role in label-noise learning tasks, which refers to the transition from clean labels to noisy labels. The majority of recent methods for inferring the transition matrix concentrate on the manually hand-crafted label noise but with bearing the high cost of time and labor. In light of this, several straightforward and effective algorithms are introduced for automatically annotating the label noise. However, the automatic annotation algorithms easily generate wrong pseudo labels for similar semantic categories. Moreover, a special instance-dependent transition matrix is launched due to the mapping from a specific category to other similar categories during the annotation process. To address this issue, we propose a semantic adaption estimator (SAE) to indirectly infer the instance-dependent transition matrix. Specifically, we decouple the original instance-dependent transition matrix to several easy-to-estimate semantic-dependent transition matrices by introducing a semantic adaption loss function. In this way, the original datasets can be decoupled into some simple semantic regions. Then the instance-dependent transition matrix can be built from multiple learned semantic-dependent matrices. Empirical evaluations on two real-world datasets (i.e., S3DIS and ScanNet) demonstrate the superior performance of our method, in comparison with the state-of-the-art.
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.
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
Submission Guidelines: Yes
Please Choose The Closest Area That Your Submission Falls Into: Unsupervised and Self-supervised learning
4 Replies
Loading