DM3L 2nd Annual Symposium 2026
sOrganized by: R. Furrer, Trolese
Home | Registration | Programme | Speakers | PicturesSpeakers
Prof. Dr. Julien Mairal: Machine Learning for Scientific Imaging
"In this presentation, we will present a few scientific imaging problems where hybrid approaches that combine physical models of image formation and deep learning are highly successful. We will then address a fundamental challenge in image restoration: the choice of estimator, as perceptual quality often does not align with traditional objective criteria such as minimizing the mean squared error. Finally, we will show how algorithms related to diffusion—highly successful in generative image modeling—can provide an effective solution to this problem."
Prof. Dr. Huyen Pham: Learning Stochastic Generative Dynamics from Distributional Observations
"Generative models are typically trained to reproduce a target distribution, but in many applications one also observes intermediate distributions describing how a population evolves over time. We propose a framework for learning stochastic dynamics from such distributional snapshots. The problem is formulated as a McKean–Vlasov control problem in which discrepancies from the observed distributions are imposed as soft penalties. This provides flexibility when observations are noisy or incomplete. The optimality conditions lead to a forward-backward stochastic differential equation, which also forms the basis of a neural simulation algorithm. Experiments on synthetic examples and human-motion data show that intermediate distributional supervision improves the temporal coherence of generated trajectories compared with training only on initial and terminal distributions."
Prof. Dr. Ulrike Schneider: Pattern Recovery in Penalized Estimation: A Unified Framework
"We study penalized estimators with polyhedral gauge penalties which reveal structural patterns in parameters, including LASSO, generalized LASSO, SLOPE, OSCAR, and PACS. We define the notion of patterns via subdifferentials, introduce a measure for pattern complexity, and derive a minimal criterion for a pattern to be detected with positive probability, the accessibility condition. We further propose the stronger noiseless recovery condition, which generalizes and unifies the LASSO irrepresentability condition. For appropriately thresholded penalized estimators, we show that accessibility alone ensures exact pattern recovery under sufficiently small noise. Our results also admit a clear geometric interpretation."
Prof. Dr. Christa Zoufal: Learning from Quantum Data
"Recent advances in quantum machine learning theory have established that learning models trained on data generated by quantum devices can provably access information about many-body ground states that is unattainable from purely classical data. However, translating these guarantees into practice has remained challenging, with experimental demonstrations largely restricted to small systems or highly structured states due to the difficulty of preparing complex many-body ground states on quantum hardware.
In recent work, we demonstrate a scalable realization of learning from quantum data in an interacting two-dimensional many-body system. Leveraging approximate ground-state preparation techniques—including sampled-based Krylov diagonalization and approximate quantum compilation—combined with high-performance computing and modern machine learning, we generate quantum datasets for the 2D Heisenberg XXZ model with system sizes of up to 115 qubits. The dataset comprises local observables, two-point correlations, and higher-order loop correlations across the antiferromagnetic phase.
We train neural networks on this experimentally derived quantum data and show accurate prediction of spatially resolved observables for previously unseen Hamiltonian parameters, both within and beyond the training distribution, approaching the phase boundary. Our results provide a concrete step toward practical quantum-enhanced learning and outline a path to regimes where quantum devices can supply training data beyond the reach of classical approximation methods."
References:
https://arxiv.org/abs/2112.00778
https://arxiv.org/pdf/2606.15983