William De Vazelhes, Ph.D. Email & Phone Number

Dubai, Ae

Abu Dhabi, Abu Dhabi Emirate, United Arab Emirates

R&D on Large Language Models.

Abu Dhabi, United Arab Emirates

Worked on optimization (first order and zeroth-order) and sparsity (hard-thresholding). Published in NeurIPS, ICLR, JMLR, and AAAI. Reviewer at ICML (2024), ICLR (2023) and NeurIPS (2023).

Région De Boulogne-Billancourt, France

Huawei and Université Paris-Est, on Information Theory and Machine Learning.Topics researched: Information Theory (in particular, Rate Distortion Theory with logarithmic loss), Minimax Density Estimation (in particular using KDE), Variational Autoencoders, Optimal Transport (in particular Wasserstein Barycenter).Interrupted to start a PhD anew at MBZUAI

Région De Lille, France

- Developed a scikit-learn compatible package for metric learning. (https://github.com/scikit-learn-contrib/metric-learn) (1.3K+ stars on Github, 60+ cit.). - Improved the speed, memory cost, and robustness of pre-existing algorithms by vectorizing computations in novel ways, changing the iterative methods for optimizing the algorithms, solving problems due to numerical uncertainties, and check-proofing the algorithms by testing them on mathematical toy problems.- Contributed to scikit-learn. Main contribution: NeighborhoodComponentsAnalysis (NCA).

Région De Boulogne-Billancourt, France

- Improved the number of records successfully fuzzy matched in the company's heterogeneous database by approximately 30%, using Elasticsearch + query randomization in Python + ML-based results filtering in scikit-learn. Developed both the prototype and part of the production code (pickling, and efficient batch prediction API).- Prototyped mail classification algorithms to help send the appropriate invoice reminder or warning to consumers.- Developed shared Python libraries with the team. Prepared the team's GPU server to integrate it with Dataiku's built-in machine learning interface.

Orange Gardens, Châtillon, France

Deep Reinforcement Learning for Human-Machine negociation dialogue. - Adapted and tuned several deep reinforcement algorithms (DQN, Actor-Critic) to maximise score for a human-machine sequential two-players task (agenda-setting task). - Implemented them to work in the Java simulation environment, with the Deeplearning4j framework. - Compared the performance of the algorithms to pre-existing algorithms for this problem.- Found significative variables to monitor along training and visualized them with Python Matplotlib.

Région De Montréal, Canada

"Metal artefacts correction in CT reconstruction with polar coordinates."I had to :- Analytically compute and implement a likelihood criterion (and its gradient) for reconstructing a 3D image (prior) having some observed X-ray projections of it, taking into account an innovative projection method for matrices with polar coordinates developed by the lab (computationally more efficient thanks to invariances in the case of tomography), and the non-linearities in the projection due to the X-ray spectrum polychromaticity.- Optimize this criterion with an existing method from Matlab (here l-bfgs-b (a gradient conjugate method)).- Find under what conditions the argument of the criterion converges towards the desired 3D image to reconstruct (tomography parameters, characteristics of the object to reconstruct, criterion and optimization parameters)Topics:- Inverse problems- Non-linear optimization (Conjugate Gradient methods (l-bfgs-b))- Matlab object-oriented programming

Antony

Benchmarking: - Functional benchmark of competitors' ventilators in order to support the specification of a new product- Sharing of outputs with project teamKnowledge management: - Creation of a wiki for the competition data shared between the R&D and marketing departments as a basis for innovationOrganized information research and restitution- Market research: data crunching for market study

Doctor Of Philosophy - Phd, Machine Learning, Gpa 3.96/4

Worked on optimization (first order and zeroth-order) and sparsity (hard-thresholding). Published in NeurIPS, ICLR, JMLR, and AAAI..

Master Of Science - Ms

Last Year Key Courses: Machine Learning (inc. Deep Learning), Reinforcement Learning, Data Analytics, Numerical Optimization, Image.

Education record

- Telecommunication Networks, Communication Theory - Advanced Topics in Medical Image (labs: registration, segmentation, binary images.

Pcsi-Psi

French preparatory school (two-years highly selective classes to prepare for the competitive exam to the Top Engineering Schools)

Doctor Of Philosophy - Phd, Machine Learning, 3.96

Education record