BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Insights on the generalization ability of deep neural networks usi
 ng sensitivity analysis
DTSTART:20180628T100000
DTEND:20180628T120000
DTSTAMP:20260916T060036Z
UID:6472c50e1fdf225396fb0a524121c2889afc30070417b46608aeb6f9
CATEGORIES:Conferences - Seminars
DESCRIPTION:Mahsa Forouzesh\nEDIC candidacy exam\nExam president: Prof. Ru
 ediger Urbanke\nThesis advisor: Prof. Patrick Thiran\nCo-examiner: Prof. M
 artin Jaggi\n\nAbstract\nThere is a growing line of research on understand
 ing what drives generalization in deep learning settings. Sharpness analys
 is of the loss surface gives intuition on the generalization process. Robu
 stness analysis provides generalization error bounds which complexity meas
 ures of the model appear in. However\, none is sufficient to explain the g
 eneralization ability of an over-parameterized deep neural network to unse
 en data. In this proposal\, we would like to find insights on tackling thi
 s phenomenon using mathematical tools.  In particular\, we would like to 
 apply sensitivity analysis to both forward pass and backward pass of the m
 odel\, and by considering a probabilistic framework\, we would like to pro
 vide a better understanding of the performance of various algorithms. Theo
 retical explanation on why and how deep neural networks work is the starti
 ng point for designing new regularization techniques that are not only jus
 tified by empirical results but also have mathematical fundamentals.\n\nBa
 ckground papers\nMathematics of Deep Learning\, by Vidal\, R.\, et al.\nEn
 tropy-SGD: Biasing gradient descent into wide valleys\, by Chaudhari\, P.\
 , et al.\nLayer Normalization\, by Lei Ba\, J.\, et al.\n\n\n​\n 
LOCATION:BC 229 https://plan.epfl.ch/?room==BC%20229
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR