Detecting Distribution Shift with Deep Generative Models
by Dr. Eric Nalisnick (Assistant professor at the University of Amsterdam)
Deep generative models (DGMs) have demonstrated impressive generative capabilities. Moreover, for models that allow tractable density computations (e.g. normalizing flows), detecting anomalies and distribution shift is often cited as another benefit. I will present results that question this claim, showing that DGMs’ density function fails to distinguish non-training data in scenarios in which a human would never be fooled. I will go on to talk about some methodological improvements as well as fundamental limitations for DGMs for this task.
We have two more speakers at this event:
- Estefanía Talavera (University of Twente) will tell us about the interpretation of anomalies related to human behavior
- Samet Akcay (Intel Corporation) will explain Anomalib, an innovative library that serves as a bridge between state-of-the-art research and real-world product applications.
Share this Post