Issam Falih

My research is in statistical machine learning, with a focus on multimodal representation learning under data scarcity: incomplete or unpaired modalities, few labels, small samples and limited compute. I work on three connected problems:

• Multimodal representation learning from incomplete and unpaired data: multi-view and collaborative clustering, non-negative matrix factorization, multilayer graphs, and shared latent representations across sensors.
• Cross-modal transfer and generation with optimal transport: hierarchical optimal transport for transfer under distribution shift, with generalization guarantees, and transport-based generative models (flow matching) for cross-sensor generation from few paired samples.
• Reliability under scarce supervision: credal uncertainty based on belief functions, and partial identification to certify transfers and synthetic data, i.e. how many paired or labeled observations are needed to reach a decision.

Applications include health (patient phenotyping in chronic pain, multimodal clinical cohorts), human activity recognition from video and inertial sensors, and multi-sensor perception (RGB, infrared) with industrial partners (Safran, Emo’Lis). I contributed to the European project NeuroDeRisk (H2020 IMI2).

Previously Associate Professor at Université Clermont Auvergne (LIMOS, 2019–2026). PhD in Computer Science, LIPN, Université Sorbonne Paris Nord (2018). MSc in Computer Science (MIDO, Intelligent Systems — Machine Learning track), Université Paris Dauphine – PSL (2014).

Research interests optimal transport · domain adaptation · generative models (flow matching) · multi-view learning · partial identification · uncertainty quantification · frugal learning