BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Towards Unsupervised Transfer Learning
DTSTART:20230830T141500
DTEND:20230830T171500
DTSTAMP:20260407T055523Z
UID:608659dec31355adf7e251fc58052b57d56671a7df65c72d2b150cf3
CATEGORIES:Conferences - Seminars
DESCRIPTION:Artem Gadetskii\nEDIC candidacy exam\nExam president: Prof. Ma
 rtin Jaggi\nThesis advisor: Prof. Maria Brbic\nCo-examiner: Prof. Françoi
 s Fleuret\n\nAbstract\nTransfer learning is a fundamental paradigm in\nmac
 hine learning that leverages the pretraining of deep neural\nnetworks on l
 arge-scale datasets to transfer knowledge to a downstream\ntask with limit
 ed resources. A well-established approach\ninvolves fine-tuning the weight
 s of a linear classifier on top of\nfrozen representations in a supervised
  manner. This approach\nhas demonstrated success in various domains\, such
  as few-shot\nlearning\, domain adaptation\, and domain generalization. Re
 cently\,\nfoundation models trained on extensive data have achieved\nsigni
 ficant advancements across many modalities\, exhibiting\nexceptional repre
 sentation learning capabilities and enabling\nzero-shot predictions. Howev
 er\, existing approaches presume the\navailability of at least minimal sup
 ervision for downstream tasks\,\npreventing their application to more chal
 lenging fully-unsupervised\nsettings. To overcome this limitation\, deep c
 lustering methods can\nbe applied to solve the task of interest without re
 quiring human\nannotations. Yet\, the performance of such unsupervised met
 hods\nis lagging far behind. To bridge the gap between supervised\nand uns
 upervised learning we propose to consider a general\nunsupervised transfer
  learning setting pushing the limits of the\nrecent breakthroughs in repre
 sentation learning achieved by\nfoundational models.\n\nBackground papers\
 n\n	Learning transferable visual models from natural language supervision.
  Alec Radford et al. International Conference on Machine Learning 2021. (l
 ink)\n	SCAN: Learning to classify images without labels. Wouter Van Gansbe
 ke et al. European Conference on Computer Vision 2020. (link)\n	Learning t
 o discover novel visual categories via deep transfer clustering. Kai Han e
 t al. International Conference on Computer Vision 2019 (link)\n
LOCATION:BC 129 https://plan.epfl.ch/?room==BC%20129
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR