We would like to emphasize that all faculty, students and postdocs are invited to attend the talks.
Moreover, if desired, we would like to offer assistance on traveling to the CRM via metro鈥攊f there is interest, we may organize traveling as a group.
If you would like to sign up to the mailing list, please contact scarpa.carlo [at] uqam.ca
Title: Elements of compressed sensing with deep generative models
Abstract: Compressed sensing (CS) is an impressive modern theory in applied mathematics that has revolutionized signal processing. CS permits recovery of high-dimensional structured signals from random, underdetermined and corrupted linear measurements. Typically, the measurement matrix of the linear system is random (e.g., has independent identically distributed Gaussian entries), and recovery is guaranteed with high probability on its realization. On the other hand, deep generative models have garnered renown for their impressive ability to effectively model the manifold of natural images. Such models are frequently comprised of alternating compositions of affine transformations and piecewise linear nonlinearities. Recently, it was proved that this type of generative model can serve as the underlying structure for compressed sensing problems. In this talk we provide a foundational understanding of compressed sensing, an overview of deep learning relevant to the design of effective natural signal modelling, and highlight the key elements that connect these seemingly disparate topics.
Zoom link: on request
Free coffee 鈽� and cookies 馃崻 will be offered after the talk.
Venue: Salle 4336-4384, Pav. Andr茅 Aisenstadt, UdeM