FI Computational Methods and Data Science Journal Club
Flatiron Institute, 162 5th Avenue
Speaker: Jeremy Cohen (CREATIS, CNRS, Villeurbanne, France)
Title: Semi-supervised Low-rank Approximation, from variational methods to deep learning
Abstract: Low-Rank Approximations can be understood as a class of inverse problems when the input data is a matrix or a tensor. They have been in particular used extensively in the context of blind source separation/pattern mining, when little or no side information on the patterns to extract is available.
However, in many practical applications of LRA, such additional information is available, and can be used to inform the algorithms and improve performances. I will call such models Semi-Supervised LRA. After introducing LRA, in this presentation I will explore existing literature on SSLRA, showing that it bridges the gap between traditional unsupervised variational methods, and recent deep learning supervised models. Through this presentation, it will be made clear that SSLRA actually covers a wide range of models, e.g. when labels, priors, generative modes or denoisers are known, and is a fast evolving topic with lots of opportunities for both theoretical and practical contributions.
FI employees are welcome to attend in person. Please email firstname.lastname@example.org for the Zoom link if you wish to attend remotely.
Visitors (w/out an FI badge) please email email@example.com to be registered to the building or to obtain Zoom information. All visitors must be vaccinated for entry into our buildings.