Manuel Haussmann

I'm a postdoctoral researcher at the University of Southern Denmark in Odense, where I work in the group of Melih Kandemir. I did my PhD at Heidelberg University with Fred Hamprecht and Melih, and spent two great years at Aalto University in Finland working with Samuel Kaski and Harri Lähdesmäki. My research focus is on Bayesian probabilistic machine learning with applications ranging from physics over health data to reinforcement learning. I am an action editor for TMLR, a reviewer for most major machine learning conferences (AAAI, AISTATS, ICLR, ICML, NeurIPS), and a member of the ICBINB initiative.

email  /  dblp  /  scholar  /  github

Research

Fixing plasticity issues in PPO via evidential learning.

Improving Bayesian optimization by incorporating prior knowledge into the latent space.

An evaluation of uncertainty calibration with BNNs and ensembles on LHC physics data.

We propose a novel actor-critic algorithm for continuous control in sparse reward settings, using a critic ensemble and PAC-Bayesian bounds on Bellman operator error to enable deep exploration via posterior sampling.

We introduce a moment-matching based uncertainty penalty for the Bellman operator in an offline RL setting.

A structured variational auto-encoder constrained with a mixed model for longitudinal data.

We introduce the first PAC Bayesian bound trainable in a modern deep reinforcement learning pipeline to solve continuous control tasks.

A model for longitudinal data that can handle irregular observations and model missing not at random data.

We introduce a probabilistic model to combine different patient populations for trial effect estimation.

An evaluation on the differences between a curated medical trial data set and uncurated electronic health records data.

We introduce a method to incorporate high-dimensional equivariances into (conditional) neural processes.

An evaluation of Bayesian normalizing flow for LHC events.

Evidential deep learning combined with concepts from neural Turing machines and neural processes.

We introduce a probabilistic method to learn dynamical systems that allows for partial input incorporation.

We combine evidential deep learning and PAC-based regularization with moment-matching approaches from Bayesian deep learning.

We discuss Bayesian deep-learning methods for jet classification in particle physics.

A sampling-free approach to BNN learning via ReLU decomposition.

We introduce a way to adaptively learn the acquisition function for active learning with reinforcement feedback.

We present LeMoNADe, an end-to-end learned motif detection method directly operating on calcium imaging videos.

Instance level inference with closed-form updates in a Gaussian processed-based model for weakly supervised multiple instance learning and object detection.

We introduce a model for multiple instance learning with Gaussian processes for bag-level predictions with closed-form updates.

Design and source code based on Jon Barron's website.