MARS Seminar Series

Abstract

We propose an unsupervised deep-learning framework for the analysis of optical microscope data in the context of photovoltaic materials characterization. The goal is to recover spatially distributed carrier lifetime maps from a series of time-resolved fluorescence imaging (TR-FLIM) measurements. Accurate recovery of these parameters is critical for predicting solar cell efficiency and identifying early signatures of material degradation. However, observations can be photon-count-limited and corrupted by a challenging combination of noise sources whose joint statistical modelling is computationally intractable.

Our approach builds on the Noise2Noise (N2N) statistical learning framework, which exploits the theoretical result that a model trained to map one noisy realization of a signal to another independent noisy realization of the same signal converges, in expectation, to the underlying clean signal without ever requiring access to ground-truth data. We incorporate a physics-driven log-linear decay model into the N2N loss function, reflecting the known exponential decay structure of the emitted photoluminescence and enabling the simultaneous estimation of spatially varying amplitude and lifetime maps. The reconstruction is performed in a zero-shot regime, meaning no pre-training on external datasets is required: weight optimization is carried out directly on the set of experimental noisy observations available for a single scene.

This setting is well suited to low signal-to-noise ratio regimes arising from short acquisition protocols, which minimize potential damage caused by prolonged exposure to the laser excitation source. The method is demonstrated to be effective for the monitoring of lifetime evolution under controlled degradation conditions, including thermal stress and humidity exposure.

Joint work with Luca Calatroni, Stefania Cacovich and Tommaso Raimondi.

Abstract

Many real-world problems involve learning tasks where prediction accuracy is not equally important across the input space, but instead matters most in regions determined by the unknown function itself. When function evaluations are expensive, how can we learn efficiently without knowing these regions in advance? In this talk, we formalize this problem from an active learning perspective, focusing on Gaussian process regression under self-induced Boltzmann distributions. We introduce a novel acquisition strategy that approximates the unknown target distribution without estimating its partition function, while balancing exploration and weighted variance reduction. We present theoretical convergence guarantees and demonstrate the effectiveness of the proposed method on synthetic benchmarks and real-world applications in computational chemistry and drug discovery.

Abstract

Complex systems in nature and in applications (such as molecular systems, crowd dynamics, swarming, opinion formation, just to name a few) are often described by systems of stochastic differential equations (SDEs) and partial differential equations (PDEs). It is often analytically impossible or computationally prohibitively expensive to deal with the full models due to their high dimensionality (degrees of freedom, number of involved parameters, etc.). It is thus of great importance to approximate such large and complex systems by simpler and lower dimensional ones, while still preserving the essential information from the original model. This procedure is referred to as model reduction or coarse-graining in the literature. In this talk, I will present methods for qualitative and quantitative coarse-graining of several SDEs and PDEs, in the presence or absence of a scale-separation.

Abstract

Reconstructing images in low-photon scenarios, such as astronomy, biology and medical imaging, poses severe challenges due to high uncertainty and non-Gaussian noise. This talk explores Bayesian approaches that integrate data-driven priors with advanced sampling methods to achieve robust reconstruction and principled uncertainty quantification. I will present recent developments in accelerated proximal Markov chain Monte Carlo algorithms and plug-and-play Langevin sampling tailored to Poisson imaging, highlighting how optimisation theory and numerical analysis inform practical algorithm design in data-driven Bayesian inference. In my current work, I focus on fluorescence microscopy as a real-world application, where principled uncertainty quantification is essential for drawing reliable scientific conclusions. I will discuss some of the open problems in this field and outline future directions.

Abstract

TBC

Abstract

Sampling high dimensional probability measures is often made difficult by the multimodality of the target probability distribution. Markov chain Monte Carlo methods need to pass through low probability regions to switch from one mode to another, which is a rare event. An approach to making these transitions less rare is to identify a few selected (nonlinear) degrees of freedom of the system, which are at the origin of the slow mixing behavior, compute the associated free energies, and perform some importance sampling based on the latter function. Various tools have now recently been developed in molecular simulation to automatically find the most relevant nonlinear degrees of freedom hindering sampling, based on machine learning tools such as autoencoders. I will present a methodology to leverage these models for better sampling, and will also provide a mathematical analysis of the approach, relating it to principal manifolds and providing an interpretation based on conditional expectations. I will also discuss recent work in combining these techniques with supervised dimensionality reduction approaches such as Fisher's LDA. The results will be illustrated on biologically relevant systems such as HSP90.