ANC Workshop - Lorenzo Loconte

Tuesday, 18th June 2024

Charting the Expressiveness of Tractable Generative Models - Lorenzo Loconte

Abstract: Generative models are popular methods powering AI systems, whose performance depends on their expressiveness, i.e., their ability of capturing complex data distributions. Generative models supporting tractable probabilistic inference are particularly suitable for building trustworthy AI systems, but they often exhibit a limited expressiveness. In this talk, I will show how investigating the limitations of tractable generative models in approximating complex distributions gives us a better understanding of how to boost their expressiveness. Concretely, I will present a class of deep mixture models that subtract probability density, which makes them exponentially more expressive than traditional mixtures. Finally, by unifying various tractable generative models arising from different fields, I will sketch how jagged their relative expressiveness landscape can be.

Event type: Workshop

Date: Tuesday, 18th June

Time: 11:00

Location: G.03

Speaker(s): Lorenzo Loconte