Master Assignment
[B] Self-Supervised Out-of-Distribution Detection
Type: Master EE/CS/ITC
Period: TBD
Student: (Unassigned)
If you are interested please contact :
Background:
Self-Supervised Out-of-Distribution Detection
Shifts in the data distribution may compromise the performance of deep learning methods. As these shifts are often unknown, methods from self-supervised learning can help identify feature clusters on the original and augmented data for more robust out of distribution detection. The work of [1] claims to successfully detect OOD data in unlabeled medical imaging datasets compared to other self-supervised, contrastive learning approaches. These methods will be compared in a principled manner on data with simulated semantic and distribution shifts, real-world medical imaging data with domain shift, to determine their respective advantages and limitations.
- SOoD: Self-Supervised Out-of-Distribution Detection Under Domain Shift for Multi-Class Colorectal Cancer Tissue Types https://cibm.ch/wp-content/uploads/Bozorgtabar_SOoD_Self-Supervised_Out-of-Distribution_Detection_Under_Domain_Shift_for_Multi-Class_Colorectal_ICCVW_2021_paper.pdf
- GOAD https://github.com/lironber/GOAD