CV
Basics
Name | Michaël Soumm |
Label | Deep Learning Researcher |
soumm@telecom-paris.fr | |
Url | https://msoumm.github.io |
Summary | Deep Learning researcher with expertise in model evaluation, fairness, and representation learning. Three years of experience in deep learning across computer vision, NLP, and recommendation systems. Solid applied (MVA MSc.) and theoretical (ENSAE Paris MSc.) knowledge, consistently ranked in the top 5% in DL, CV, and NLP courses. Three papers published at WACV, and one paper currently under review. |
Work
-
2025.02 - Present Postdoctoral Researcher in Deep Learning
IDS (Télécom Paris) & MLIA (Sorbonne University)
Disentangled and Controllable Latent Representations for Computer Vision and Medical Imaging. Under the supervision of Pietro Gori and Dr. Alasdair Newson.
- Developping structured latent representations for medical imaging, focusing on disentangling healthy and pathological brain patterns in neuroimaging.
- Adapting contrastive learning to enhance interpretability in psychiatric disorder diagnostics.
- Investigating diffusion models, optimizing their latent space structure for improved synthesis and analysis of neuroimaging data.
-
2021.12 - 2024.12 PhD in Deep Learning
CEA LIST
Refining Machine Learning Evaluation: Statistical Insights into Model Performance and Fairness. Under the direction of Betrand Delezoide and Adrian Popescu.
- Developed and integrated statistical methodologies to improve the robustness and interpretabilityy of deep learning model evaluations in 3 fields
- Studied the impact of pre-training and architectures for Class Incremental Learning (CIL), identifying key factors affecting performance, with results published in WACV 2024.
- Investigated biases in Face Recognition systems, quantifying demographic disparities, with findings under review. Showed how using conditioned generation can reduce biases, accepted in WACV 2025.
- Developed novel information metrics to characterize difficult users for Recommender systems, and studied their impact on performance. Work under review.
-
2021.05 - 2024.11 Research Intern in Deep Learning
CEA LIST / Valeo
Self-supervised depth prediction from monocular images. Goal: perform depth prediction for autonomous driving. Under the supervision of Florian Chabot.
- Developed and integrated statistical methodologies to improve the robustness and interpretabilityy of deep learning model evaluations in 3 fields
- Crafted a Vision Transformer U-Net pipeline for Depth prediction.
- Adapted the model to a self-supervised setup.
-
2020.05 - 2020.08 Research Intern in Statistics
CREST (ENSAE Paris)
Investigated variable selection in high-dimensional data, using Sequential Monte-Carlo methods. Under the supervision of Nicolas Chopin.
Education
-
2021.12 - 2024.12 Palaiseau, France
-
2020.09 - 2021.05 Palaiseau, France
MSc
ENS Paris-Saclay
MVA (Mathematics, Vision, and Learning)
- Deep Learning (V. Lepetit)
- Object Recognition and Computer Vision (I. Laptev)
- Deep Learning in Practice (G. Charpiat)
- Image denoising : the human machine competition (J.M. Morel)
- Graphs in Machine Learning(M. Valko)
- Geometrical Deep Learning (J. Feydy)
- Geometric Approaches in Statistical Learning: The Example of Longitudinal Data (S. Durrleman)
-
2018.09 - 2021.05 Palaiseau, France
MSc
ENSAE Paris
Statistics
- Statistics 1 & 2
- Deep Learning
- NLP
- Reinforcement Learning
- Optimization
- Time Series
- Bayesian Statistics
- Econometrics
Publications
-
2025 Fairer Analysis and Demographically Balanced Face Generation for Fairer Face Verification
Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision
Develops fairness-aware analysis and face generation techniques to mitigate demographic biases in face verification models.
-
2024 An Analysis of Initial Training Strategies for Exemplar-Free Class-Incremental Learning
Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision
Investigates the impact of initial training strategies on exemplar-free class-incremental learning, highlighting key factors affecting model adaptation over time.
-
2024 Quantifying User Coherence: A Unified Framework for Cross-Domain Recommendation Analysis
arXiv preprint
Proposes a framework to measure user coherence in cross-domain recommendation, providing insights into user behavior consistency across different domains.
-
2023 Vis2Rec: A Large-Scale Visual Dataset for Visit Recommendation
Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision
Introduces Vis2Rec, a large-scale dataset for visit recommendation based on visual data, addressing challenges in recommendation systems with multimodal information.
Skills
Deep Learning | |
Computer Vision | |
Model Evaluation | |
Representation Learning | |
Fairness | |
Recommendation Systems | |
NLP | |
Class Incremental Learning |
Mathematics | |
Probability | |
Statistics | |
Optimization | |
Linear Algebra | |
Analysis | |
Graph Theory |
Coding | |
Python | |
Pytorch | |
Tensorflow | |
Scikit-learn | |
Pandas | |
Numpy |
Languages
🇫🇷 French | |
Native speaker |
🇬🇧 English | |
Fluent (TOEIC 990/990) |
🇷🇺 Russian | |
Native speaker |
🇩🇪 German | |
Basics (B1) |
Projects
- 2020 - 2020
NLP and Political Programs – Applied Statistics Project (ENSAE)
Used Natural Language Processing (NLP) techniques to analyze the representation of French political programs in the media, assessing their visibility and narrative influence.
- Topic Modeling with LDA
- Text Embeddings using Word2Vec and TF-IDF
- Advanced Techniques: LSTM and Transformers (BERT)