In modern settings, ML models are highly complex, trained on partly labeled and even synthetic data and then deployed in dynamic environments, making inference on unstructured data. Unfortunately, this challenges the classical statistical inference tools and procedures to a large extent. Indeed, classical inference methods were not designed for this new landscape, and most often are not valid when deployed in such settings. To ensure that data driven conclusions are truly reliable in real-world settings, statistical tools need to be updated, and this requires substantial practical and theoretical development.
My work is focused on the development of new learning algorithms that bridge the gap between the ability to generate predictions and the ability to ensure that these predictions are reliable and suitable for guiding real-world decisions.
My current research interests include:
Statistical Inference: advancing techniques in machine learning and statistical inference to make new, replicable data-driven discoveries.
Robustness: developing principled strategies to measure, monitor, and adapt to train–test mismatch, aiming to maintain reliability when deployment conditions differ from training.
Other topics I've worked on are:
Probabilistic Graphical Models: representation of high-dimensional distributions, and inference using Markov and Bayesian Networks.
Dependence-Structure modeling: part of my early work has been focused on the copula construction. Together with my PhD advisor, Prof. Gal Elidan, we introduced the Copula-Bayesian-Network construction that allows flexible modeling of high-dimensional distribution. I was also lucky to make some contribution to the theory of copula construction.