BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Over-parameterized deep neural networks: optimization\, robustness
 \, and generalization
DTSTART:20230629T090000
DTEND:20230629T110000
DTSTAMP:20260407T051414Z
UID:82b45ea398ab91bdf633328750605175a27a02d6ac9a7803d8082280
CATEGORIES:Conferences - Seminars
DESCRIPTION:Zhenyu Zhu\nEDIC candidacy exam\nExam president: Prof. Lenka Z
 deborova\nThesis advisor: Prof. Volkan Cevher\nCo-examiner: Prof. Nicolas 
 Flammarion\n\nAbstract\nDeep neural networks (DNNs) have demonstrated rema
 rkable achievements in various fields\, and the networks used in\npractica
 l applications are continuously becoming wider and deeper. While it is kno
 wn that overparameterized neural networks are easy to\nlearn\, researchers
  in machine learning still have concerns about certain weaknesses in deep 
 neural networks\, such as convergence\,\nrobustness\, and generalization. 
 This report aims to discuss recent important advancements in understanding
  deep neural networks\,\nspecifically focusing on the works of [1]\, [2]\,
  [3]\, in order to gain a better comprehension of their behavior. The anal
 ysis of deep neural\nnetworks will be a primary focus\, highlighting key m
 athematical models and their algorithmic implications. Furthermore\, we wi
 ll explore\nthe challenges associated with understanding deep neural netwo
 rks and discuss current research directions in this field.\n\nBackground p
 apers\n1. A convergence theory for deep learning via over-parameterization
  (Allen-Zhu\, https://arxiv.org/abs/1811.03962).\n2. A Universal Law of Ro
 bustness via Isoperimetry (Sébastien Bubeck\, https://arxiv.org/abs/2105.
 12806).\n3. Benign Overfitting without Linearity: Neural Network Classifie
 rs Trained by Gradient Descent for Noisy Linear Data" (https: https://arxi
 v.org/abs/2202.05928).\n\n 
LOCATION:ELD 120 https://plan.epfl.ch/?room==ELD%20120
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR