News
Please have a look at open positions in our group. See here for more details and join our team.
New preprint on improving multitask finetuning using Bayesian model merging is available on arXiv.
Our team have two accepted workshops at ICLR 2025: “Quantify Uncertainty and Hallucination” and “XAI4Science”.
Emtiyaz Khan gives a Distinguished Lecture at CISPA Helmholtz Center for Information Security.
The 2nd Bayes-duality workshop videos are now publicly available.
- Playlist: https://bayesduality.github.io/workshop_2024.html#videos
- Each talk separately: https://bayesduality.github.io/talks_2024.html
Our team has two open positions for research scientist or post-doc. The candidate will work on problems at the intersection of deep learning, Bayesian inference, optimization and reinforcement learning. Please check here for more details.
Thomas Möllenhoff will give invited talks:
- November 7, 2024 at the IBIS 2024 Machine Learning Theory workshop, Tokyo, Japan.
- December 4-5, 2024 at OCAMI Workshop, Osaka, Japan.
Thomas Möllenhoff will give invited talks:
- September 12, 2024 at the Seminar on Advances in Probabilistic Machine Learning (Aalto University) and the Mathematical Optimization Research Seminar (Saarland University).
- September 25-27, 2024 at KAIST in Daejong, Korea.
Emtiyaz Khan gave a keynote at the 3rd Conference on Lifelong Learning Agents (CoLLAs) 2024 [Slides].
Our paper Variational Learning is Effective for Large Neural Networks is highlighted at ICML 2024 as a spotlight. If you are at ICML, please make sure to check it out! Before heading to ICML in Vienna, Thomas Möllenhoff is giving four talks on the work:
- July 5, ISBA 2024, Ca’ Foscari University in Venice,
- July 8, The computer vision group at TU Munich,
- July 10, The data science group at University of Munich,
- July 18, The efficient learning and probabilistic inference for science group at Helmholtz AI.
Also check out our blog post, PyTorch optimizer and code for reproducing the experiments from the paper.
We are holding the 2nd Bayes-duality workshop 2024 (June 12-21), focusing on the design of AI that learns adaptively, robustly, and continually, like humans. Part 1 of the workshop is open to the public (through live streaming) and will feature 26 talks, 7hrs of tutorials, and panel discussions by world-leading researchers working on related topics. Register here
Our paper Variational Learning is Effective for Large Neural Networks is now published at ICML 2024. Also check out our blog post for additional remarks and explanations and find the code here.
Hugo Monzon and Thomas Möllenhoff are giving poster presentations at the RIKEN-LAMDA workshop in Nanjing, China.
Thomas Möllenhoff gave a talk at the joint symposium of RIKEN AIP and the Italian Institute of Technology.
Emtiyaz Khan gave an invited talk at the Workshop on Optimal Transport in Berlin. Thomas Möllenhoff gave an invited talk at Weierstraß-Instutit in Berlin and visited Jia-Jie Zhu’s group.
“Geometric Aspects of Data-Processing of Markov Chains” (G. Wolfer, S. Watanabe) is accepted for publication in Transactions of Mathematics and Its Applications.
We have two papers accepted to ICLR 2024:
- N. Daheim, T. Möllenhoff, E. M. Ponti, I. Gurevych, M. E. Khan, [Model Merging by Uncertainty-Based Gradient Matching]
- E. Guha, S. Natarajan, T. Möllenhoff, M. E. Khan, E. Ndiaye, [Conformal Prediction via Regression-as-Classification].
One paper accepted at TMLR: Improving Continual Learning by Accurate Gradient Reconstructions of the Past.
Emtiyaz Khan is co-organizing a Dagstuhl Seminar On the Role of Bayesianism in the Age of Modern AI. It will be held from Nov. 10 - Nov. 15, 2024.
One paper (The Memory Perturbation Equation: Understanding Model’s Sensitivity to Data) accepted at NeurIPS 2023.
“Empirical and Instance-Dependent Estimation of Markov Chain and Mixing Time” (G. Wolfer) is accepted for publication in Scandinavian Journal of Statistics.
Our paper on Bridging the Gap Between Target Networks and Functional Regularization is accepted at TMLR.
So Takao, Molei Tao and Rajesh Ranganath are visiting our group. See our reading-group page for more information about their talks.
“Systematic Approaches to Generate Reversiblizations of Markov Chains” (M. C.H. Choi and G. Wolfer) is accepted for publication in IEEE Transactions on Information Theory.
Thomas Moellenhoff and Emtiyaz Khan are organizing the Duality Principles for Modern ML Workshop at ICML 2023 in Hawaii.
From August 2 to August 9, Emtiyaz Khan, Gian Maria Marconi and Lu Xu will be giving invited talks at the Machine Learning Research School (MLRS), Bangkok, Thailand.
Ang Ming Liang’s (NUS) undergraduate thesis (done at RIKEN as a remote collaborator) was awarded “Lijen Industrial Development medal” for best academic projects in the discipline. His thesis was on the Knowledge Adaptation Prior paper by Khan and Swaroop (NeurIPS 2021).
G. Wolfer was awarded a KAKENHI Grant-in-Aid for Early-Career Scientists (Apr. 2023- Mar. 2026, USD ~32,000).
“Improved Estimation of Relaxation Time in Non-reversible Markov Chains” (G. Wolfer and A. Kontorovich is accepted for publication in Annals of Applied Probability.
“Geometric Reduction for Identity Testing of Reversible Markov Chains” (G. Wolfer and S. Watanabe, is accepted for an oral presentation at GSI’23.
Two papers accepted at ICML 2023 on “Memory-based Dual GPs” and Momentum-based Riemannian Optimization respectively.
Our Dagstuhl Seminar on AI for Social good (part III) is accepted (to be held Feb 19-23, 2024); see the first meeting details here
Emtiyaz Khan and Thomas Möllenhoff are co-organizing (with Mathieu Blondel and Zelda Mariet) the ICML 2023 Workshop on Duality Principles in Modern Machine Learning! See the workshop webpage for more details and the call for papers.
Our work “SAM as an Optimal Relaxation of Bayes” (T. Moellenhoff and M. E. Khan) is accepted at ICLR 2023 (notable top 5%). We have another paper “The Lie-Group Bayesian Learning Rule” (E. M. Kiral, T. Moellenhoff, M.E. Khan) accepted at AISTATS 2023.
Happy Buzaaba is co-organizing the ICLR2023 workshop on AfricaNLP to be held on 5th May in Kigali, Rwanda.
We organized the “Continual Lifelong Learning Workshop” at ACML2022 in Hyderabad, India.
Invited Prof. Michael Choi (National University of Singapore and Yale-NUS College) for a seminar talk. Title “From reversiblizations of non-reversible Markov chains to landscape modification”.
Emtiyaz Khan gave an invited talk at NeurIPS 2022 Workshop on Broadening Research Collaborations in Machine Learning [Slides].
Emtiyaz Khan gave an invited talk at University of Bath in Prof. Olga Isupova’s group [Slides].
Today, we hosted Prof. Yarin Gal for a seminar on “Uncertainty in Deep Learning: Lessons Learned from Medical Imaging”.
Forthcoming talks by Geoffrey Wolfer:
- [16 Nov, 2022] [ CCSP Seminar ] Talk: "Geometric Aspects of Markov Models"
- [22 Nov, 2022] [ IBIS'22 ] Poster: "Variance-Aware Estimation of Kernel Mean Embedding"
- [1 Dec, 2022] [ SITA'22 ] Talk: "Markovian Embeddings of Markov Chains"
Pierre will visit Toulouse School of Economics as an invited researcher in November. He will give a series of 5 lectures at on PAC-Bayes and Mutual Information bounds, on Nov. 8, 15, 21, 22 and 28. The lectures are intended for PhD students but will be also opened for master students. They will be essentially based on User-friendly introduction to PAC-Bayes bounds. He will also give the following talks:
- [10 Nov, 2022] [ MAD-Stat seminar, Toulouse School of Econonomics ] Talk: "The nice properties of MMD for statistical estimation"
- [14 Nov, 2022] [ Séminaire Parisien de Statistique à l'IHP ] Talk: "Concentration and robustness of discrepancy-based ABC"
- [17 Nov, 2022] [ Probability and Statistics Seminar, University of Luxembourg ] Talk: "Minimum MMD estimation"
Forthcoming talks by Geoffrey Wolfer:
- [19 Sep, 2022] [ IG4DS ] Talk: "Information Geometry of Reversible Markov Chains" [ Slides ] [ Video ]
- [20 Sep, 2022] [ IG4DS ] Talk: "Geometric Aspects of Data-Processing of Markov Chains" [ Slides ] [ Video ]
- [26 Sep, 2022] [ BIID'10 ] Talk: "Geometric Aspects of Data-Processing of Markov Chains"
Forthcoming talk by Pierre:
- [08 Sep, 2022] [ O'Bayes 2022, Santa Cruz, California ] Talk: "Concentration and robustness of discrepancy-based ABC"
[ ENTROPY ] special issue on Approximate Bayesian Inference (edited by Pierre) is now available as a book and as a pdf file that can be downloaded for free on the editor website: [ Book page ]
Recent talks by Pierre:
- [20 Jan, 2022] StatML CDT seminar (Imperial & Oxford): "Matrix factorization for time series analysis" [ Slides ]
- [02 Feb, 2022] Invited talk at [ AABI2022 ] "What to expect from PAC-Bayes bounds" [ Slides ] [ Video ]
- [03 Mar, 2022] RIKEN AIP High-dimensional modelling team seminar: "Deviation inequalities for Markov chains" [ Slides ] [ Video ]
- [19 Apr, 2022] [ DELTA ] Seminar: "PAC-Bayes bounds and contraction of the posterior"
- [21 Apr, 2022] [ International Bayes Club ] Seminar: "PAC-Bayes bounds and contraction of the posterior"
- [28 Apr, 2022] [ One World ABC ] Seminar: "Concentration and robustness of discrepancy-based ABC" [ Slides ] [ Video ]
Forthcoming talks:
- [31 May, 2022] [ Statistical estimation and deep learning in UQ for PDEs ] Talk: "Robust estimation via minimum distance estimation"
- [14 Jun, 2022] [ New Trends in Statistical Learning II ] Talk: "A theoretical analysis of catastrophic forgetting"
- [23 Jun, 2022] [ EcoDep 2022 Conference on Networks Reconstruction ] Talk: "A theoretical analysis of catastrophic forgetting"
Emtiyaz Khan gave the following talks in March and Feb, 2022
- [24 Mar, 2022] Talk at MLT_init_ [ Slides ]
- [16 Mar, 2022] Cambridge CBL Reading Group [ Slides ]
- [11 Mar, 2022] DeepMind/ELLIS CSML Seminar Series at UCL [ Slides ] [ Video ]
- [ 1 Mar, 2022] ATR-AIP Joint Seminar on Neuroscience-Inspired AI [ Slides ]
- [22 Feb, 2022] AI4Sec Seminar Series at Huwaei Research Munich [ Slides ] [ Summary ]
- [17 Feb, 2022] StatML CDT seminar at Imperial College London and University of Oxford [ Slides ] [ Summary ]
Emtiyaz Khan, Dharmesh Tailor and Siddharth Swaroop will be giving an invited talk on Adaptive and Robust Learning with Bayes at the NeurIPS 2021 Bayesian deep learning workshop on Dec. 14th 11:10 - 11:30 GMT.
We won both tracks of the NeurIPS competition on Approximate Inference in Bayesian Deep Learning! Please join our talks at the NeurIPS 2021 competition track (Dec. 9th, 6pm GMT) and at the BDL workshop (Dec. 14th, 4:40pm GMT) to learn about our solution.
Our Bayes-duality project is launched with a funding of $2.76 million by JST-ANR’s CREST proposal
Two papers accepted at NeurIPS 2021:
- Knowledge-Adaptation Priors, by M.E. Khan and S. Swaroop
- Dual parameterization of SVGP, by P. Chang, V. ADAM, M.E. Khan, A. Solin
Paper by Dimitri Meunier & Pierre Alquier accepted+published in Entropy: [Meta-strategy for Learning Tuning Parameters with Guarantees]
Invited talk by Emtiyaz Khan at Durham CDT Data Science Workshop.
Invited talk by Pierre Alquier at Ecology and Dependance 2021 (EcoDep) on [ Risk Bound for High Dimensional Time Series Completion ]
Invited talk by Emtiyaz Khan at Theory and Foundation of Continual Learning [ Slides ] [ SlidesLive Video ]
One paper accepted at UAI2021: Subset-of-Data Variational Inference for Deep Gaussian-Process Regression by A. Jain, P.K. Srijith, M.E. Khan.
Three papers were accepted at ICML2021:
- Scalable marginal likelihood estimation for model selection in deep learning by A. Immer, M. Bauer, V. Fortuin, G. Rätsch, M. E. Khan.
- Tractable structured natural gradient descent using local parameterizations by W. Lin, F. Nielsen, M. E. Khan, M. Schmidt.
- Non-exponentially weighted aggregation: regret bounds for unbounded loss functions by P. Alquier.
16th AIP Open Seminar: talks by Approximate Bayesian Inference Team. Talks by
- Emtiyaz Khan: Bayesian principles for Learning-Machines,
- Dharmesh Tailor: Memorable Experiences of Learning-Machines,
- Pierre Alquier: Meta-Strategy for Hyperparameter Tuning with Guarantees.
Two papers were accepted at AISTATS2021:
- A Theoretical Analysis of Catastrophic Forgetting through the NTK Overlap Matrix by T. Doan, M. Abbana Bennani, B. Mazoure, G. Rabusseau, P. Alquier.
- Improving predictions of Bayesian neural networks via local linearization by A. Immer, M. Korzepa, M. Bauer.
“Approximate Bayesian Inference”, the editorial of a forthcoming special issue in Entropy, written by P. Alquier, is now published: paper. You can submit a paper until the end of Feb. 2021. Special Issue on Approximate Bayesian Inference
The videos of two recent talks by Pierre Alquier are now online:
- “Regret bound for online variational inference” (Oct. 29) - Workshop on online decision making, Berkeley
- “Estimation with the MMD distance” (Nov. 4) - DataSig seminar series
A series of seminars on “Bayesian principles for learning machines” held by Emtiyaz Khan will take place at the following dates and locations.
- 3 Nov. 2020 - Waterloo AI institute
- 5 Nov. 2020 - TU Darmstadt
- 17 Nov. 2020 - Uppsala University
Our paper on continual learning by functional regularization on the memorable past is accepted as an oral presentation at NeurIPS 2020.
We have two tutorials at SMILES 2020! See Emtiyaz Khan’s tutorial on DL with Bayes (slides, video) and Pierre Alquier on Sequential Prediction Problems (video).
Emtiyaz Khan received a Kakenhi Grant (Series B) on “life-long learning” (approx. USD 158K).