Belhal Karimi

ML Research
Baidu Research, Cognitive Computing Lab
Send an email Twitter Google Scholar Github LinkedIn Soundcloud Instagram

[Curriculum Vitae] [Google Scholar] [Research Statement]
Previous: INRIA; Ecole Polytechnique; Samsung Research; MIT

Writing on pharmacokinetics related subjects: [blog]


I am currently an ML Researcher in the Cognitive Computing Lab at Baidu Research working with Dr. Ping Li.
I obtained my Ph.D. in Machine Learning at Ecole Polytechnique (CMAP) and INRIA, under the supervision of Marc Lavielle and Eric Moulines. From October 2016 to October 2019, I was a member of the XPOP team, a joint team between INRIA and CMAP focussing on statistical modelling for life sciences.

I had the opportunity to intern in the Samsung AI - HSE Lab (Moscow, RU, 2019), where I was working with Dr. Dmitry Vetrov on optimization methods for Bayesian Neural Networks and at MIT (Boston, USA, 2016), under the supervision of Dr. Vikash Mansinghka, where I've spent most of my time at the Probabilistic Computing Project working on MCMC methods.

I am developing, with other researchers, an open-source R package, called Saemix, devoted to the training of nonlinear mixed models in biostatistics, epidemiology etc. Check out our R bookdown [bookdown] and Twitter feed [twitter] for more information.



An Optimistic Acceleration of AMSGrad for Nonconvex Optimization
Jun-Kun Wang, Xiaoyun Li, Belhal Karimi and Ping Li.
Sumitted, 2020.
[abs] [pdf] [code]
Two-Timescale Stochastic EM Algorithms
Belhal Karimi and Ping Li.
Sumitted, 2020.
[abs] [pdf] [code]
Towards Better Generalization of Adaptive Gradient Methods
Yingxue Zhou, Belhal Karimi, Jinxing Yu, Zhiqiang Xu and Ping Li.
Advances in Neural Information Processing Systems (NeurIPS), 2020.
[abs] [pdf] [code]
FedSKETCH: Communication-Efficient Federated Learning via Sketching
Farzin Haddadpour, Belhal Karimi, Ping Li and Xiaoyun Li.
Submitted, 2020.
[abs] [pdf] [code]
Scaling Saemix, a dedicated R package for nonlinear mixed effects modeling
Belhal Karimi and Emmanuelle Comets.
Chan Zuckerberg Intitiative Proposal (CZI), 2020.
[abs] [pdf] [code]
f-SAEM: A fast Stochastic Approximation of the EM algorithm
Belhal Karimi, Marc Lavielle and Eric Moulines.
Computational Statistics and Data Analysis (CSDA), vol. 141, p. 123-138, 2020.
Accepted for a poster presentation at the Paris-Berlin Young Researchers Workshop : Stochastic Analysis with applications in Biology and Finance.
[abs] [pdf] [code] [poster]
Nonconvex Optimization for Latent Data Models: Algorithms, Analysis and Applications
Belhal Karimi.
Ph.D. thesis, Xpop at INRIA and Ecole Polytechnique, 2019.
[abs] [pdf] [slides]
On the Global Convergence of (Fast) Incremental Expectation Maximization Methods
Belhal Karimi, Hoi-To Wai, Eric Moulines and Marc Lavielle.
Advances in Neural Information Processing Systems (NeurIPS), 2019.
[abs] [pdf] [code] [slides] [poster]
Non-asymptotic Analysis of Biased Stochastic Approximation Scheme
Belhal Karimi, Blazej Miasojedow, Eric Moulines and Hoi-To Wai.
Proceedings of the 32nd Conference On Learning Theory (COLT), 2019.
[abs] [pdf] [code] [slides] [poster]
MISSO: Minimization by Incremental Stochastic Surrogate for large-scale nonconvex Optimization
Belhal Karimi and Eric Moulines.
1st Symposium on Advances in Approximate Bayesian Inference (AABI), 2018.
[abs] [pdf] [code] [slides] [poster]
Efficient Metropolis-Hastings sampling for nonlinear mixed effects models
Belhal Karimi and Marc Lavielle.
Proceedings of Bayesian Statistics and New Generations (BAYSM), 2018.
[abs] [pdf] [code] [slides] [poster]
On the Convergence Properties of the Mini-Batch EM and MCEM Algorithms
Belhal Karimi, Marc Lavielle and Eric Moulines.
Accepted for a poster presentation at the Data Science Summer School (DS3), 2017.
[abs] [pdf] [code] [slides] [poster]
Non linear Mixed Effects Models: Bridging the gap between Independent Metropolis Hastings and Variational Inference
Belhal Karimi, Marc Lavielle and Eric Moulines.
Accepted at the Implicit Models workshop (ICML), 2017.
[abs] [pdf] [code] [slides] [poster]
Probabilistic and Inferential Programming
Belhal Karimi.
MS thesis, ProbComp Project at MIT, 2016.
[abs] [pdf] [code] [slides] [poster]


NeurIPS 2019 On the Global Convergence of (Fast) Incremental EM Methods
Vancouver, Canada, Dec. 2019. [poster]
Ph.D. Defense Nonconvex Optimization for Latent Data Models: Algorithms, Analysis and Applications
Palaiseau, France, Sept. 2019. [slides]
Samsung AI - HSE Lab An Incremental and An Online Point of View of Nonconvex Optimization
Moscow, Russia, Aug. 2019. [slides]
Baidu Research Nonconvex Optimization for Latent Data Models: Algorithms, Analysis and Applications
Beijing, China, Aug. 2019. [slides]
COLT 2019 Non-asymptotic Analysis of Biased Stochastic Approximation Scheme
Phoenix, USA, Jun. 2019. [poster]
PGMODays 2018 MISSO Scheme
Palaiseau, France, May. 2018. [slides]
Compstat 2018 Acceleration of MLE algorithms
Iasi, Romania, Aug. 2018. [slides]
Facebook HQ Mixed effects models: Maximum Likelihood and Inference
Paris, France, Feb. 2017. [slides]
McGovern Institute Analysis of birth cohort studies in BayesDB
Boston, USA, May. 2016. [poster]
ML Tea at CSAIL Probabilisitc Computing Project
Boston, USA, May. 2016. [slides]


Marc Lavielle, Emmanuelle Comets, Audrey Lavenu and Belhal Karimi.
[git] [web] [R bookdown]


Research in Information retrieval in video streams. Video search engine and automatic trailer generation using Deep Learning.
PR: Axe IA accueille 5 nouvelles entreprises...
Monk AI
Machine Learning advisor to Monk AI, a French startup providing automated & objective condition reports, Paris, France. Damages detection using Mask R-CNN.
Samsung AI
Research intern at Samsung AI - HSE Lab, leading Russian lab in Bayesian Deep Learning, Moscow, Russia.
Project: Optimization for Bayesian Neural Networks with Dr. Dmitry Vetrov.
One-year freelance at OuiCar, French peer-to-peer car sharing platform, Paris, France.
Project: damages detection using Deep Learning (Mask R-CNN).
Freelance at Popsy, leading South American Classifieds App, NYC, USA.
Project: predict the Category and Price of any given listing based on their pictures.
Industry workshops at AMIES, Mathematics and Enterprises, Montreal, Canada.
Project: Inspection Route Optimization. [slides] [pdf]
Start-up project Agora, an innovative Scratches Detection computer vision engine, Paris, France.
See [pitch], [www] and [UI].


Visiting Student Researcher Grant
Obtained from the Jacques Hadamard Foundation [Junior Scientific Visibility] to pursue a research project on Bayesian Deep Learning at the HSE-Samsung AI Lab in Moscow with Dr. Dmitry Vetrov. ANR-11-LABX-0056-LMH.
Student Travel Award
Conference on Learning Theory (June 2019 Phoenix, USA).
Young Researcher Travel Award
International Conference on Bayesian Statistics in Action (July 2018 Warwick, UK).
Startup Pitch Award
Ranked 4th/130 at JSC 2017. Awarded to pitch at Axel Springer Plug and Play accelerator in Berlin, Aug 2017.

Reviewing Activities

NeurIPS ICBINB Workshop 2020
AABI 2019
ICML 2019
Statistics and Computing - Springer
Neural Networks - Elsevier

Teaching Activities

Executive Education Program at Ecole Polytechnique: Machine Learning (Orange, SFR,...)
MAP534 Machine Learning: Msc Ecole Polytechnique-HEC
MAP535 Regression: Msc Ecole Polytechnique-HEC
Bayesian Statistics: Msc Data Science Ecole Polytechnique
Innovation & Technology: 3rd-year students at Ecole Polytechnique


Ecole Polytechnique
Oct. 2016 -- Sep. 2019
Ph.D. candidate in Machine Learning
Advisors: Prof. Marc Lavielle, Prof. Eric Moulines
Paris Sciences et Lettres, PSL ITI
Sep. 2015 -- Jun. 2016
Masters in Computer Science
Advisor: Dr. Vikash Mansinghka while visiting MIT (Jan. -- Jun. 2016).
Professors: Prof. Gabriel Peyre (Computer Vision and Graphics).
Sep. 2011 -- Jun. 2015
Bachelors and Masters in Computer Science and Engineering
Professors: Prof. Olivier Pietquin, Prof. Matthieu Geist (Signals & Systems).


Belhal Karimi