Robert Bamler
I am a machine-learning scientist and engineer. From 2020 to 2025, I was a university professor, directing the group of Machine Learning and Data Science at the Cluster of Excellence "Machine Learning: New Perspectives for Science" at University of Tübingen, Germany. I then joined Nokia TECH in Munich as a Principal Researcher in the MPEG-AI team, where I invented new methods for efficient AI model compression until Nokia announced the closure of its Munich lab. I am currently looking for my next opportunity in machine learning research and development, so please feel free to reach out to me if you are hiring.
Full Resume: PDF
My expertise spans probabilistic machine learning and uncertainty quantification, AI for science (in particular, hybrid physics/ML models), and efficient AI (model quantization, compression, and inference acceleration). I derive machine-learning methods from a deep understanding of the mathematical models, and I implement them to run fast on real hardware. See my publications and open-source contributions below for details.
Before becoming a professor, I worked at Disney Research (part of Walt Disney Imagineering) in Pittsburgh and Los Angeles, and as a postdoc at UC Irvine (in the group of Prof. Stephan Mandt). I received my PhD in theoretical statistical and quantum physics from University of Cologne, Germany, in 2016, advised by Achim Rosch and with support from the Deutsche Telekom Foundation.
YouTube playlist Lecture notes, problem sets, and solutions
- Representative Publications
- Open-Source Contributions
- Publications (Machine Learning and AI)
- Physics Publications
- Selected Talks & Video Abstracts
Representative Publications
- Probabilistic ML & AI for science: hybrid ML models for predicting physico-chemical properties (Digital Discovery 2025), uncertainty quantification in Bayesian deep learning (NeurIPS 2024), and physics-inspired improvements to black box variational inference (NeurIPS 2017).
- Efficient AI: compression of deep-learning models (arXiv preprint, 2025), image compression with deep-learning models (ICML 2020), and faster training for very large output spaces (ICLR 2020).
See the full list of publications below.
Selected Open-Source Contributions
- wgpu, the WebGPU implementation used by Firefox (5 merged pull requests): packed int8 dot products on all GPU backends, a building block for quantized inference on consumer GPUs.
- constriction: my Rust/Python library of fast entropy coders for AI-based compression (100+ GitHub stars).
- GPU kernels for compressed neural networks: a proof of concept that decodes entropy-coded weights on the fly to cut memory traffic in LLM decoding (related paper).
- kladde: my library of durable data structures (think "SQLite without SQL") in Rust and C++.
Peer-Reviewed Publications
For a complete list of my publications and citation statistics, please see my Google Scholar profile.
-
Well-Defined Function-Space Variational Inference in Bayesian Neural Networks via Regularized KL-DivergenceUAI 2025PDF
-
What Actually Matters for Materials Discovery: Pitfalls and Recommendations in Bayesian OptimizationAI for Accelerated Materials Design workshop at ICLR 2025PDF
-
On the Challenges and Opportunities in Generative AITMLR 2025Journal Website
-
Verbalized Machine Learning: Revisiting Machine Learning with Language ModelsTMLR 2025PDF
-
Your Finetuned Large Language Model is Already a Powerful Out-of-distribution DetectorAISTATS 2025PDF
-
Balancing Molecular Information and Empirical Data in the Prediction of Physico-Chemical PropertiesDigital Discovery, 2025PDF Journal Website
-
A Note on Generalization in Variational Autoencoders: How Effective Is Synthetic Data & Overparameterization?TMLR, 2024PDF
-
FSP-Laplace: Function-Space Priors for the Laplace Approximation in Bayesian Deep LearningNeurIPS 2024PDF
-
Predictive, scalable and interpretable knowledge tracing on structured domainsICLR 2024PDF
-
Differentiable Annealed Importance Sampling Minimizes The Jensen-Shannon Divergence Between Initial and Target DistributionICML 2024PDF
-
A Compact Representation for Bayesian Neural Networks By Removing Permutation SymmetryNeurIPS 2023 Workshop on Unifying Representations in Neural ModelsPDF
-
The SVHN Dataset Is Deceptive for Probabilistic Generative Models Due to a Distribution MismatchNeurIPS 2023 Workshop on Distribution ShiftsPDF
-
Resampling Gradients Vanish in Differentiable Sequential Monte Carlo SamplersICLR 2023 (TinyPapers track)PDF
-
Trading Information between Latents in Hierarchical Variational AutoencodersICLR 2023PDF
-
Post-Training Neural Network Compression With Variational Bayesian QuantizationNeurIPS 2022 Workshop on Challenges in Deploying and Monitoring Machine Learning SystemsPDF
-
Making thermodynamic models of mixtures predictive by machine learning: matrix completion of pair interactionsChemical Science 13 (17), 4854-4862PDF
-
Improving Inference for Neural Image CompressionNeurIPS 2020PDF
-
User-Dependent Neural Sequence Models for Continuous-Time Event DataNeurIPS 2020PDF
-
Hybridizing physical and data-driven prediction methods for physicochemical propertiesChemical Communications, 2020PDF Journal Website
-
Variational Bayesian Quantization(former title: Variable-Bitrate Neural Compression via Bayesian Arithmetic Coding)ICML 2020PDF Video abstract by Yibo Yang (15 min)
-
Machine Learning in Thermodynamics: Prediction of Activity Coefficients by Matrix CompletionThe Journal of Physical Chemistry Letters, 2020(11) 981-985PDF Journal website
-
Extreme Classification via Adversarial Softmax ApproximationICLR 2020PDF Video abstract (5 min.) Slides
-
Tightening Bounds for Variational Inference by Revisiting Perturbation TheoryJ. Stat. Mech. (2019) 124004PDF Journal website
-
Augmenting and Tuning Knowledge Graph EmbeddingsUAI 2019PDF Video abstract (1 min.) Slides
-
A Quantum Field Theory of Representation LearningICML 2019 Workshop on Theoretical Physics for Deep LearningPDF Video abstract (4 min.) Slides
-
Improving Optimization for Models With Continuous Symmetry BreakingICML 2018PDF Related video abstract (4 min.) Slides
-
Perturbative Black Box Variational InferenceNeurIPS 2017PDF
-
Bayesian Paragraph VectorsNeurIPS 2017 workshop on Advances in Approximate Bayesian InferencePDF
-
Structured Black Box Variational Inference for Latent Time Series ModelsICML 2017 Time Series WorkshopPDF
-
Dynamic Word EmbeddingsICML 2017PDF Promotional video (3 min.) Conference talk (17 min.)
Physics Publications
-
Phase-Space Berry Phases in Chiral Magnets: Skyrmion Charge, Hall Effect, and Dynamics of Magnetic SkyrmionsPh.D. thesis, University of Cologne (2016)PDF
-
Equilibration and approximate conservation laws: Dipole oscillations and perfect drag of ultracold atoms in a harmonic trapPhys. Rev. A 91, 063604 (2015)PDF
-
Phase-space Berry phases in chiral magnets: Dzyaloshinskii-Moriya interaction and the charge of skyrmionsPhys. Rev. B 88, 214409 (2013)PDF
-
Giant generic topological Hall resistivity of MnSi under pressurePhys. Rev. B 87, 134424 (2013)PDF
-
Hydrodynamic Object Recognition: When Multipoles CountPhys. Rev. Lett. 102, 058104 (2009)PDF
Selected Talks & Video Abstracts
Unfortunately, video recordings or slides are publicly available only for a few of my past talks. If you have any questions about a talk that's not in the list below (or if the slides don't resolve your question), please feel free to contact me.
-
It’s Not Just ChatGPT: Underappreciated Machine-Learning Methods for the Natural SciencesCITENI, Universidade da Coruña, Spain, 9/2025Slides
-
Scalable Bayesian Inference: New Tools for New Challengesml4science Cluster Colloquium, University of Tübingen, Germany, 2/2021Video recording (50 min.) Slides
-
Extreme Classification via Adversarial Softmax ApproximationICLR 2020Video abstract (5 min.) Slides Paper (PDF)
-
Revisiting Variational Expectation MaximizationAI/ML Seminar @ UC Irvine, 2019Slides Related paper 1 (PDF) Related paper 2 (PDF)
-
Marrying Big Data And Probabilistic Models With Variational InferenceGerman Aerospace Center (DLR), 2019Slides
-
Augmenting and Tuning Knowledge Graph EmbeddingsUAI 2019Video abstract (1 min.) Slides Paper (PDF)
-
A Quantum Field Theory of Representation LearningTheoretical Physics for Deep Learning Workshop at ICML 2019Video abstract (4 min.) Slides Workshop paper (PDF) Related paper (PDF)
-
Dynamic Word EmbeddingsICML 2017Video recording (17 min.) Paper (PDF)