BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Multi-Task Scene Representations
DTSTART:20210824T130000
DTEND:20210824T150000
DTSTAMP:20260407T123440Z
UID:0e253dba7a10f975181a29af0da3ac73e4df5ecb14932362b9177b91
CATEGORIES:Conferences - Seminars
DESCRIPTION:Roman Bachmann\nEDIC candidacy exam\nexam president: Prof. Nic
 olas Boumal\nthesis advisor: Prof. Amir Zamir\nco-examiner: Prof. Mackenzi
 e Mathis\n\nAbstract\nCurrent supervised and self-supervised representatio
 n learning literature focuses heavily on using large-scale classification 
 datasets to train a network to produce image-level features that can be us
 ed for transfer learning. The following questions arise: Does training on 
 classification tasks/datasets really produce the best representations for 
 diverse downstream task learning\, and why do we perform transfers from in
 dependent image-level features instead of scene-level representations that
  aggregate information over time and space? Indeed\, there is evidence tha
 t no pre-training task is the best single choice for all other visual down
 stream tasks. We propose to learn scene-level representations by merging i
 mage-level representations of multiple diverse tasks over the spatial and 
 temporal dimensions\, with the goal of creating powerful visual priors for
  downstream learning. Using such multi-task priors should improve the cove
 rage of the space of features that are useful for visual tasks. Furthermor
 e\, the use of scene representations can allow for global and out-of-sight
  reasoning.\n\nBackground papers\n1) Big Transfer (BiT): General Visual R
 epresentation Learning. Kolesnikov et al. 2019. https://arxiv.org/abs/19
 12.11370\n2) Neural scene representation and rendering. Eslami et al. 201
 8.: https://storage.googleapis.com/deepmind-media/papers/Neural_Scene_Rep
 resentation_and_Rendering_preprint.pdf)\n3) On the Theory of Transfer Lea
 rning: The Importance of Task Diversity. Tripuraneni et al. 2020. https:
 //arxiv.org/abs/2006.11650\n 
LOCATION:
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR