Master Assignment
[B] Can Normalizing Disentangled Flows Detect OOD data?
Type: Master EE/CS/ITC
Period: TBD
Student: (Unassigned)
If you are interested please contact :
Background:
Invertible flow networks can lead to disentangled latent spaces for interpretable results. Self-supervised normalizing flows have been used for the challenging problem of image anomaly detection and localization in [1], and detection of OOD data [2-4]. However, they can fail for OOD data as well [4]. In this project you should examine in depth the advantages and limitations of normalizing flows for detecting OOD data by investigating benchmarking datasets with existing and artificially added noise. Explainability methods can be used to localize OOD data and understand how it affects the accuracy of anomaly detection.
- Self-Supervised Normalizing Flows for Image Anomaly Detection and Localization – CVPRW23 https://openaccess.thecvf.com/content/CVPR2023W/VAND/html/Chiu_Self-Supervised_Normalizing_Flows_for_Image_Anomaly_Detection_and_Localization_CVPRW_2023_paper.html
- AE-FLOW: Autoencoders with Normalizing Flows for Medical Images Anomaly Detection https://openreview.net/forum?id=9OmCr1q54Z
- Normalizing Flows for Out-of-Distribution Detection: Application to Coronary Artery Segmentation https://www.mdpi.com/2076-3417/12/8/3839
- Out-of-Distribution Detection of Melanoma using Normalizing Flows https://pure.tue.nl/ws/portalfiles/portal/198572416/2103.12672v1.pdf
- Why Normalizing Flows Fail to Detect Out-of-Distribution Data https://github.com/howardyclo/papernotes/issues/74
- Gyoza Python package for invertible flow networks https://github.com/TimHenry1995/gyoza/tree/main