Speakers: Ettore Saetta, University of Naples

Hosted at: SISSA, International School of Advanced Studies, Trieste, Italy, room 004.

Abstract:

AEs and their variants have received considerable attention in recent literature for their ability  to make accurate predictions in fields such as Fluid Mechanics . However, the question of  which information an AE extracts during data compression remains open. Some studies show  that AEs can recover database design parameters by ordering the latent space, but it is still  unclear whether the latent representation also encodes additional, physically useful  information about the underlying phenomena. Some articles in literature propose an  interpretability of the latent spaces achieved by injecting a priori knowledge of the governing  physics; without such prior knowledge it is difficult to assign meaning to latent variables. The analysis must also account for the geometry of the latent space, which is a manifold; Kelshaw  and Magri have discussed this aspect extensively and proposed the Proper Latent  Decomposition as a generalization of Proper Orthogonal Decomposition on manifolds. Focus of the talk will be the introduction of a Machine Learning procedure based on sequential training and unsupervised latent conditioning that aims to separate distinct physical processes. By conditioning the latent representation sequentially, we show that a systematic analysis of latent components can provide insight into different physical contributions. Sample applications to Fluid Dynamic datasets will be presented, comparing the latent  spaces produced by conventional AEs with those produced by a sequential AE. A brief  discussion of the manifold geometry will conclude the talk.

Zoom link:
https://sissa-it.zoom.us/j/84131589707?pwd=Wg46kAalwmZyPrrTBbadWZgm0ZSL4O.1