Friday, 29th of November

11h (room R2014, 660 building) (see location)

Luigi Gresele

(Max Planck Institute, Tübingen)

The Incomplete Rosetta Stone Problem: Multi-View Nonlinear ICA, with applications to neuroimaging

Independent Component Analysis (ICA) provides a principled framework to
study unsupervised feature extraction and blind source separation. The
theory of ICA studies under which conditions the latent sources which
generated the data can be reconstructed from observations. If this is
possible, the model is said to be identifiable.
While identifiability has been thoroughly characterized for the linear
setting, thus triggering the development of multiple techniques for
independent component estimation, the general nonlinear problem is
unidentifiable.
In recent years, the theory of nonlinear ICA witnessed some crucial
advancements, with new constructive proofs of identifiability under
additional assumptions on the data generative model.
I will discuss a novel theoretical result, the "Rosetta Stone
Identifiability", characterizing the case in which multiple noisy views
of the sources are available. I will also discuss the applicability of
this result to the analysis of group studies in neuroimaging, in
particular for shared response modeling.

Joint work with Paul K. Rubenstein, Arash Mehrjou, Francesco Locatello
and Bernhard Schölkopf

Reference

http://auai.org/uai2019/proceedings/papers/53.pdf



All TAU seminars: here