BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Obtaining Robust Classifiers via Regularization and Averaging Sche
 mes.
DTSTART:20180830T110000
DTEND:20180830T130000
DTSTAMP:20260407T043421Z
UID:050df3fb58912604241043512bd79002b7454bc186ac85dee4a57c97
CATEGORIES:Conferences - Seminars
DESCRIPTION:Fabian Latorre Gomez\nEDIC candidacy exam\nExam president: Pro
 f. Emre Telatar\nThesis advisor: Prof. Volkan Cevher\nCo-examiner: Prof. D
 aniel Kuhn\n\nAbstract\nNeural networks have shown great promise for task 
 automation. However\, the\ndiscovery of adversarial examples has upset thi
 s promise: small perturbations\nof the input data make the misclassificati
 on rate greatly increase. Recent\napproaches have been shown empirically t
 o provide some level of robustness to\nperturbed data\, but fail to provid
 e theoretical guarantees for perturbations of\ndifferent sizes. We address
  this shortcoming by introducing\nWasserstein-Lipschitz Regularization (WL
 R)\, an optimization objective whose\nsolution provides theoretical bounds
  on the robustness against perturbations of\nlarge size. We also study the
  effect of model averaging on the robustness and\nestablish sufficient con
 ditions for it to improve performance. Our results\nsuggest similar regula
 rization schemes can be derived for unsupervised learning\nmethods such as
  GANs\, as well as regression problems. Further research will\nfocus on de
 veloping the necessary theory and algorithms to control the lipschitz\ncon
 stant of neural networks\, as well as understanding the trade-offs between
 \nthe robustness to adversarial examples and the choice of network archite
 cture.\n\nBackground papers\nIntriguing Properties of Neural Networks htt
 ps://arxiv.org/pdf/1312.6199.pdf\nTowards Deep Learning Models Resistant t
 o Adversarial Attacks https://arxiv.org/abs/1706.06083\nWasserstein Distr
 ibutional Robustness and regularization in Machine Learning https://arxiv
 .org/abs/1712.06050\n 
LOCATION:ELD 120 https://plan.epfl.ch/?room=ELD120
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR