Probabilistic divergences such as Kullback-Leibler, Rényi and f-divergence play an increasingly important role in machine learning (ML) offering a notion of distance between probability distributions. In the recent past, divergence estimation has been developed on the premise of variational formulas and function parametrization via neural networks. Despite the successes, the statistical estimation of a divergence is still considered a very challenging problem mainly due to high variance of the neural-based estimators.

Particularly, hard cases include high dimensional data, large divergence values and Rényi divergence when its order is larger than one. His recent work focuses on reducing the variance by regularizing the function space of the variational formulas.

 

Prof. Pantazis will present novel families of divergences which enjoy enhanced statistical and stability properties. Those function-space regularized divergences have been tested against a series of ML application including generative adversarial networks, mutual information estimation and rare sub-population detection. If time permits, he will also present the construction of gradient flows based on those regularized divergences which result in transport particle algorithms and show that such particle algorithms are data efficient.

 

 

Yannis Pantazis, PhD, is a Principal Researcher at the Institute of Applied and Computational Mathematics of the Foundation of Research and Technology - Hellas (IACM-FORTH), Greece. He received his BSc, MSc and PhD in Computer Science from the University of Crete in 2004, 2006 and 2010, respectively. After his PhD, he became Postdoctoral Research Associate at the University of Massachusetts, Amherst working on Applied and Stochastic Mathematics and, upon return to Greece, at the University of Crete developing algorithms for sparse dynamical system inference from temporal measurements.

His research interests revolve around theoretical and algorithmic advancements in Applied Probability, Information Theory and Machine Learning applied on a variety of scientific and engineering fields including Stochastic Processes, Uncertainty Quantification, Speech Processing and Bioinformatics.

He is author of 55 referred scientific publications that has been presented in various journals, conferences and workshops. He has received the 1st honorable mention (bronze medal) in DARPA’s Forecasting Chikungunya Outbreak Challenge (2015) and won the 1st place in CTTSO's Challenge on "Algorithmic Identification of Material Mixtures from Raman Spectra" (2016).

---

 

 

 

---