BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Semi-Unsupervised Learning with Deep Generative Models / Disentang
 ling Improves VAEs' Robustness to Adversarial Attacks - Matthew Willetts a
 nd Alexander Camuto\, University of Oxford / Alan Turing Institute
DTSTART:20190717T130000Z
DTEND:20190717T140000Z
UID:TALK127612@talks.cam.ac.uk
CONTACT:Eric T Nalisnick
DESCRIPTION:This talk will consist of two parts (approx 20 min each)\n\n#1
   "Semi-Unsupervised Learning with Deep Generative Models":https://arxiv.o
 rg/abs/1901.08560\n\nSemi-Unsupervised learning is a new regime that may b
 e quite common in the real world. It is a form of ultra-sparse semi-superv
 ised learning\, where for some classes in our data there are no labelled e
 xamples in the training set at all\, only unlabelled examples. It could be
  due to selection biases in how we obtained our annotated data compared to
  our larger amount of unlabelled data. Or it could just be that we were un
 lucky and missed some classes in the small labelled dataset.\n\nThis type 
 of data motivates models that can jointly perform clustering and semi-supe
 rvised learning. After training we want to have a classifier that accurate
 ly predicts both semi-supervised classes and classes for which it has neve
 r seen a single labelled example.\n\nIt is well known that some DGMs devel
 oped to learn a classifier in the semi-supervised regime can fail when tra
 ined on unlabelled data only ---they simply don't learn to discern between
  classes of data. We cover why this happens and present some models we hav
 e developed that enable us to learn in this set-up\, and show some results
 .\n\n\n\n\n#2    "Disentangling Improves VAEs' Robustness to Adversarial A
 ttacks":https://arxiv.org/abs/1906.00230\n\nWe highlight that conventional
  VAEs are brittle under attacks that aim to fool them into reconstructing 
 a subtly-distorted input to a chosen target image.\n\nHowever\, methods re
 cently introduced for disentanglement such as β-TCVAE (Chen et al.\, 2018
 ) improve robustness to proposed adversarial attacks that aim to match the
  distorted input to the target in a VAE's latent space. This motivated us 
 to develop Seatbelt-VAE\, a new hierarchical disentangled VAE that is desi
 gned to be significantly more robust to adversarial attacks than existing 
 approaches\, while retaining high quality reconstructions.
LOCATION:Engineering Department\, CBL Room BE-438.
END:VEVENT
END:VCALENDAR