Unpacking the Black Box: From PCA to EOFlows
Details
How can we uncover the most meaningful factors hidden in complex data and understand what each of them actually represents?
We are happy to welcome Prof. Dr. Ullrich Köthe from Heidelberg University to heidelberg.ai to talk about EOFlows, a non-linear extension of Principal Component Analysis (PCA).
PCA identifies the main factors of variation in a dataset, but it is limited to linear relationships. Normalizing flows can capture more complex patterns, but they lack interpretability.
EOFlows aim to combine the flexibility of normalizing flows with the interpretability of PCA. Using orthogonality regularization, they learn latent dimensions that capture distinct factors of variation, which can be ranked by their explained entropy.
In this talk, Prof. Köthe will explain how EOFlows work and show what they discover in the CelebA portrait dataset. He will also discuss how stable the results are across training runs and how the method compares with existing approaches.
About Ullrich Köthe
Ullrich Köthe is an Associate Professor at Heidelberg University, where he has led a group in the Visual Learning Lab since 2018. His research spans image analysis, machine learning, and explainable AI, with a focus on developing methods that make the inner workings of neural networks easier to understand.
For more info go to: https://www.heidelberg.ai/2026/11/10/ullrich-koethe.html
After the event, we will also upload a video recording to our YouTube channel: https://www.youtube.com/channel/UCfHWBneOsb7SfOxJepnMQKA