Agenda

14 May 2025 11:15

Predictive risk for Poisson inverse problems

Sala Riunioni B, edificio ZETA - Campus Scientifico via Torino

Speaker: Federico Benvenuto, Università di Genova

Abstract:
In this work, we present an estimator for predictive risk in Poisson data using the Kullback-Leibler divergence as the loss function, aimed at guiding the choice of regularization parameters in the Expectation Maximization (EM) algorithm.

We establish a Poisson version of Stein’s Lemma for Gaussian variables, which leads to our proposed estimator, drawing parallels with Stein’s Unbiased Risk Estimator for quadratic loss.

Our estimator is asymptotically unbiased as the number of measured counts increases, under mild conditions that are satisfied by the EM algorithm.

We apply this estimator to optimize the reconstruction process and conduct numerical tests in image deconvolution, comparing its performance against existing methods in both inverse crime and non-inverse crime scenarios.

Language

The event will be held in Italian

Organized by

Gabriele Santin

Search in the agenda