Best student paper award to Kimia Nadjahi at ICASSP 2020

Kimia Nadjahi, PhD candidate at the Image, Data, Signal department (supervisors: Umut Şimşekli and Roland Badeau), received the best student paper award at ICASSP 2020.

Kimia NadjahiICASSP 2020 Conference, held remotely in May 2020, is about Signal Processing: from Sensors to Information, at the heart of Data Science. The best student paper award was granted by a committee of experts who evaluated the novelty, impact and quality of the paper, as well as its oral presentation.

Kimia is a 2nd-year PhD student at Télécom Paris, advised by Umut Şimşekli and Roland Badeau. Her main research interests are statistical inference and optimal transport. Specifically, her research focuses on understanding how computational optimal transport, approximate inference and implicit generative modeling algorithms can be used jointly for large-scale machine learning applications.

The awarded paper is entitled Approximate Bayesian Computation with the Sliced-Wasserstein Distance
by Kimia Nadjahi, Valentin De Bortoli, Alain Durmus, Roland Badeau, Umut Şimşekli.

Approximate Bayesian Computation (ABC) is a popular method for approximate inference in generative models with intractable but easy-to-sample likelihood. It constructs an approximate posterior distribution by finding parameters for which the simulated data are close to the observations in terms of summary statistics. These statistics are defined beforehand and might induce a loss of information, which has been shown to deteriorate the quality of the approximation. To overcome this problem, Wasserstein-ABC has been recently proposed, and compares the datasets via the Wasserstein distance between their empirical distributions, but does not scale well to the dimension or the number of samples. The paper proposes a new ABC technique, called Sliced-Wasserstein ABC and based on the Sliced-Wasserstein distance, which has better computational and statistical properties. In this paper, the autors derive two theoretical results showing the asymptotical consistency of their approach, and illustrate its advantages on synthetic data and an image denoising task.