MLGdańsk #139 – DCID: Deep Canonical Information Decomposition

We would like to invite you to 139. MLGdańsk meeting on Monday, 17th April 2023 at 18.00 CET

The meeting will be held online on jitsi:

Our guests speaker is:
Alexander Rakowski
Hasso Plattner Institute, Poczdam

The topic of the meeting is:
DCID: Deep Canonical Information Decomposition

“We consider the problem of separating the signal shared between two scalar variables, from their individual signals. In the case of non-scalar, multivariate variables, the probabilistic interpretation of CCA allows to retrieve the shared signal. However, in the univariate case, CCA can only yield trivial solutions. Inspired by the field of Deep Multi-Task Learning, we propose to overcome this problem by approximating the variables of interest with Deep Neural Network classifiers and using their latent features as multivariate decompositions of the variables, allowing us to utilize the standard CCA machinery to discover the shared signal.

The meeting is open to all interested – feel free to join!