Internship offers – year 2024-2025
- Neural Architecture Growth for Frugal Learning
- Instead of training large models and reducing them afterwards, we develop a framework to start from tiny networks and let them grow during training, guided by the task to solve.
- Several possible internship topics: optimization (theory, algorithmics, implementation) , exploration strategies, statistical analysis (estimator convergence)
- Frugal Dictionary Learning for Imaging
- Geometric Deep Learning for Glassy Materials
- Internship Proposal: Transfer Learning for Multi-Output Regression with Few-Shot Data
- Fairness: Knowing better what is good for you?
- Learning HJB Viscosity Solutions with PINNs for Optimal Control and Continuous-Time Reinforcement Learning
- Links between explainability and formal proofs of neural networks (extraction of key concepts), in collaboration with PEPR IA SAIF. Contact Guillaume
(previously…)
Internship offers – year 2022-2023
- Toward efficient then explainable models: Choquetization of a Neural Net.
- The goal of the internship is to educate a trained neural net into being interpretable: i) forcing the decision to be a readable aggregate (Hierarchical Choquet Integral) of latent nodes; ii) formulating binary problems to give a name to each latent node, with the expert in the loop.
- contact: sebag@lri.fr, cohen@lri.fr
- Details here
- Stability-based causal discovery: an adversarial approach
- An important desired property for causal discovery is identifiability. The internship makes a step toward identifiability and focuses on stability, that is, the property of finding the same model from i) true data D; ii) data D’ generated after the model learned from D.
- contact: alessandro.leite@inria.fr, sebag@lri.fr
- Details here