cv
General Information
| Full Name | Paul Novello |
| Date of Birth | 24/07/1994 |
| Phone | +33 6 28 34 35 58 |
| paul.novello@outlook.fr | |
| Languages | English, French |
Education
-
2018 - 2022 PhD, Applied Mathematics
Ecole Polytechnique, French Alternative Energies and Atomic Energy Commission (CEA) CESTA, INRIA Saclay, France - Combining supervised deep learning and scientific computing: some contributions and application to computational fluid dynamics. Significant works:
-
2017 - 2018 MSc, Statistics
Imperial College, London, UK - Specialization in Machine Learning and Computational Statistics
- Master Thesis on Unsupervised Deep Learning with application to:
- Unsupervised representations of the concept of counting from images using Autoencoders and Generative Adversarial Networks (GAN)
- Evaluation of randomness in artificial dynamics such as cellular automata using Recurrent Neural Networks (RNN).
-
2014 - 2018 Diplôme d'Ingénieur
Télécom ParisTech, Paris, France - Major in Statistical Modeling and Scientific Computing, Signal Processing.
Experience
-
2022 - present ML Research Engineer
IRT Saint Exupery, France - Core member of the DEEL Team, a research initiative focused on trustworthy AI => Deel homepage.
- Lead of the Out-Of-Distribution (OOD) detection research project
- Collaboration with academics (MILA) and industrials (Airbus and Renault)
- Lead developer of Oodeel, an open-source library for post-hoc OOD detection on pre-trained Pytorch and Tensorflow deep neural networks.
- One Class Classification using Lipschitz Neural Networks. ICML 2023.
- Combining OOD and Conformal Prediction (CP) (preprint)
- Anomaly Detection using Diffusion models
- Ongoing projects:
OOD for Object Detection and Segmentation
Anomaly Detection with Self-Supervised Learning (SSL)
Decomposition of neural network's feature space for OOD
OOD score ensembling...
- Research on XAI
- Attribution method for XAI in Image Classification and Object Detection using Hilbert-Schmidt Independence Criterion (HSIC). NeurIPS 2022. Implemented the method in Xplique, a Neural Networks Explainability Toolbox.
- Unlocking Feature Visualization for Deep Network with MAgnitude Constrained Optimization. NeurIPS 2023.
- 3D prediction of pollutant diffusion in urban areas
- Used Fourier Neural Operators (FNO) as Surrogate models for the prediction of pollutant diffusion in 3D meshes of urban areas, extending the 2D model of Mendil et al. that was based Multi-Layer Perceptrons (MLP). Joint work with The French Alternative Energies and Atomic Energy Commission (CEA).
- Finding instance-dependent approximation guarantees for scientific machine learning using Lipschitz neural networks. Supervised a group of 3 PhD candidates during the CEMRACS hackathon. Project still ongoing.
-
2018 - 2022 PhD Candidate
Ecole Polytechnique, French Alternative Energies and Atomic Energy Commission (CEA) CESTA, INRIA Saclay, France - Combining supervised deep learning and scientific computing: some contributions and application to computational fluid dynamics. Significant works:
- Leveraging local variation in data: sampling and weighting schemes for supervised deep learning.
📚 Published in the Journal of Machine Learning for Modeling and Computing. Short version at NeurIPS's Second Workshop on Machine Learning and the Physical Sciences.
🔧 Software used: Tensorflow - Goal-oriented sensitivity analysis of hyperparameters in deep learning.
📚 Published in the Journal of Scientific Computing
🔧 Software used: Tensorflow, Slurm (conducted hyperparameter optimization on 400 GPUs on CEA's private cluster) - An analogy between solving Partial Differential Equations with Monte-Carlo schemes and the Optimisation process in Machine Learning (and a few illustrations of its benefits).
📚 Chapter of PhD manuscript
🔧 Software used: Jax - Accelerating hypersonic reentry simulations using deep learning-based hybridization (with guarantees)
📚 Published in the Journal of Computational Physics
🔧 Software used: Tensorflow, Tensorflow C api, C++, Fortran: Embedded of a Multi-Layer Perceptron (MLP) trained in tensorflow in a real-world Fortran/C++ simulation code using Tensorflow C++ API. Modification of the original Fortran C++ code to leverage neural network’s vectorization
- Leveraging local variation in data: sampling and weighting schemes for supervised deep learning.
- Combining supervised deep learning and scientific computing: some contributions and application to computational fluid dynamics. Significant works:
-
2017 - 2018 Data scientist
Intento> - First as intern, then remote while I was at Imperial College London (has been though...).
- Implementation of data analysis algorithms: website visitor's fidelity scoring, visitor's profile clustering and classification based on their navigation path, using TraMineR a Genomics sequence clustering package in R.
- Contribution to a server built with Flask, automatically processing MariaDB visitor databases from client websites.
- First as intern, then remote while I was at Imperial College London (has been though...).
Open Source Projects
Academic Interests
-
Trustworthy AI
- OOD Detection
- XAI
- Robustness
-
Scientific Machine Learning
- Hybridization between AI and classical Numerical solvers
- Approximation guarantees
Other Interests
- Sports: climbing, hiking, tennis
- Indie rock, contemporary jazz and techno music. I'm working towards confidently participating in jam sessions with my bass.
- Reading and watching about non-mathematical sciences, economy, geopolitics ...
- Got married in Sep. 2020, despite Covid19 !
- Permis B