Janis Postels

I am a PhD candidate at the Computer Vision Lab at ETH Zurich under the supervision of Luc Van Gool (ETH Zurich) and Federico Tombari (Google).

My research primarily focuses on uncertainty quantification in deep neural networks. I am specifically interested in efficiently estimating epistemic uncertainty of neural networks. Moreover, I have also worked on neural compression algorithms and generative models for point cloud data.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github  /  LinkedIn

• ManiFlow: Implicitly Representing Manifolds with Normalizing Flows was accepted to 3DV 2022
• Implicit Neural Representations for Image Compression was accepted to ECCV 2022
• Our workshop on Probabilistic Robotics was accepted to IROS 2022
• On the Practicality of Deterministic Epistemic Uncertainty was accepted to ICML 2022
• SHIFT: A Synthetic Driving Dataset for Continuous Multi-Task Domain Adaptation was accepted to CVPR 2022
• Go with the Flows: Mixtures of Normalizing Flows for Point Cloud Generation and Reconstruction was accepted to 3DV 2021
• The OOD Blind Spot of Unsupervised Anomaly Detection was accepted to MIDL 2021
• Variational Transformer Networks for Layout Generation was accepted to CVPR 2021
• I will join CVL at ETH as a PhD student starting from 02.2020
• Sampling-free epistemic uncertainty estimation using approximated variance propagation was accepted as an oral to ICCV 2019

We propose an approach for sampling from lower dimensional manifolds using normalizing flows. Moreover, we introduce an improved version of our algorithm tailored to 3D point clouds.

We propose a neural image compression algorithm by leveraging neural implicit representations and meta-learning.

We show that the uncertainty predicted by a recent family of methods for uncertainty estimation which treat the weights of a neural network deterministically is poorly calibrated.

We introduce a synthetic dataset for multi-task learning that allows fine-grained control over continuous distributional shifts.

We mitigate drawbacks of prior generative models based on normalizing flows for point clouds by introducing a mixture of normalizing flows.

We demonstrate the vulnerability of recent approaches based on VAEs for unsupervised anomaly detection in medical images.

We introduce a generative model for layouts based on VAEs and attention layers and demonstrate its strong inductive bias.

We demonstrate that the uncertainty of a neural network can be quantified using the distribution of its hidden representations.

We propose a method to estimate the uncertainty of Bayesian neural networks in a single forward pass by applying error propagation.

In this blogpost I first give an introduction to the commonly considered types of uncertainty. Then, I demonstrate on toy regression and classification examples how one can separate these types of uncertainties using Bayesian Neural Networks.