Hi there 👋, I'm Timothée Lesort, a Senior Data Scientist at Aignostics GmbH in Berlin. My work focuses on training large-scale self-supervised vision models for histopathology (mostly tweaking Dinov2 training), aiming to improve cancer and rare disease diagnostics ( or at least the numbers in the benchmarks 🙃 ).
My expertise lies in deep learning for vision and language, with a strong interest in continual learning and representation learning for robust generalization and efficient scaling.
Previously, I conducted postdoctoral research at UdeM, Mila – Quebec Artificial Intelligence Institute under the supervision of Irina Rish, where I worked on large-scale continual pretraining of LLMs (large language models).
I earned my PhD in Computer Science from IP Paris - Institut Polytechnique de Paris (France) in the U2IS lab under the supervision of David Filliat. My doctoral research, titled "Continual Learning: Tackling Catastrophic Forgetting in Deep Neural Networks with Replay Processes," investigated the use of replay mechanisms, particularly generative models, for continual learning. I also explored replay for continual reinforcement learning and the theoretical limitations of regularizing dynamic architectures in continual learning. I hold a Master's degree in Electronics and Robotics from CPE Lyon.
I love train 🚞 and bike 🚲 travelling, big trees 🌳 and playing chess ♟
Featured Research Projects:
- Pretraining of Vision Transformers for Histopathology
- Continual pre-training of large language models.
- Characterization of data distribution drifts.
- A better understanding of continual learning models.
- The impact of large pre-trained models for continual learning.
- Continuum: A PyTorch-based library designed to facilitate continual learning experimentation through diverse benchmarks, aiming to accelerate progress in the field.
- continual_learning_papers: A curated catalogue of continual learning papers
Google Scholar LinkedIn Research Gate Semantic Scholar Twitter