BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Learning with Generative Priors
DTSTART:20220825T090000
DTEND:20220825T110000
DTSTAMP:20260406T230502Z
UID:9aa81f8ac77b3d1ec7b89cd66725e8a232da8df506e0d59e2f3ca558
CATEGORIES:Conferences - Seminars
DESCRIPTION:Freya Behrens\nEDIC candidacy exam\nExam president: Prof. Volk
 an Cevher\nThesis advisor: Prof. Lenka Zdeborová\nCo-examiner: Prof. Mich
 aël Unser\n\nAbstract\nRecovering data points from their possibly nonline
 ar\nmeasurements is a ubiquotous problem in signal processing.\nWhen the m
 easurement process is ill-posed\, prior knowledge\nis required to identify
  the original input uniquely. In order to\nspecify this knowledge\, deep g
 enerative models and denoisers\nhave been succesfully employed in place of
  classical priors like\nsparsity. Based on the recent survey on the topic\
 , [Ong+20]\,\nthis report reviews some of these methods and applications a
 nd\nhighlights their wide empirical success. Then\, we describe a line\nof
  work that aims to develop a theoretical understanding of these\nmodels vi
 a message passing algorithms [Man+17\; Pan+20]. Such\nan analysis admits a
  computationally efficient inference algorithm and gives asymptotically ex
 act\npredictions of its performanceaccording to order parameters such as t
 he sample size.\nHowever\, as it concerns only multi-layered models with r
 andom parameters\,\nthis theory still falls short to describe real world s
 ettings\nwhere the generative models are trained on a data distribution\na
 nd thus have learned parameters. We conclude the report with\na discussion
  of open questions and possible research directions\nrelated to this gap.\
 n\nBackground papers\n- Multi-Layer Generalized Linear Estimation (http:/
 /arxiv.org/abs/1701.06981)\n- Deep Learning Techniques for Inverse Problem
 s in Imaging (10.1109/JSAIT.2020.2991563\; https://arxiv.org/abs/2005.060
 01)\n- Inference With Deep Generative Priors in High Dimensions (10.1109/J
 SAIT.2020.2986321\; https://ieeexplore.ieee.org/document/9061052/)
LOCATION:
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR