BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//TUC//Events//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:Europe/Athens
TZNAME:EEST
DTSTART:19700329T030000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0300
TZNAME:EET
DTSTART:19701025T040000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CREATED:20220627T100348Z
LAST-MODIFIED:20220627T100348Z
DTSTAMP:20260811T201505Z
UID:1786468505@tuc.gr
SUMMARY:Ομιλία Επ. Καθ. Α. Κυριλλίδη (Rice U
 niversity)
LOCATION:Λ - Κτίριο Επιστημών/ΗΜΜΥ, 145Π-58
DESCRIPTION:https://www.ece.tuc.gr/el/katalogos-
 ekdiloseon?tx_tucevents2_tuceventsdi
 splay%5Baction%5D=show&tx_tucevents2
 _tuceventsdisplay%5Bcontroller%5D=Ev
 ent&tx_tucevents2_tuceventsdisplay%5
 Bevent%5D=5600&cHash=a360c61c6934f7c
 cb883adf45dff0c26\nTitle: \n A Tale 
 of Sparsity in Deep Learning: Lotter
 y Tickets, Subset Selection, and Eff
 iciency in Distributed Learning\n\nA
 bstract:\n Neural network pruning is
  useful for discovering efficient, h
 igh-performing subnetworks within pr
 e-trained, dense network architectur
 es. Yet, more often than not, it inv
 olves a computationally expensive pr
 ocedure, as the dense model must be 
 fully pre-trained to achieve top-not
 ch performance, at least for a numbe
 r of iterations/epochs. Most existin
 g works in this area remain empirica
 l or impractical, based on heuristic
  rules about when, how and how much 
 one needs to pre-train to recover me
 aningful sparse subnetworks. In this
  talk, we will scratch the surface o
 f open questions in the general area
  of ``pruning techniques for neural 
 network training'', with special foc
 us on what can be theoretically char
 acterized, in order to move from heu
 ristics to provable protocols.\n\n I
 f time permits, the talk is split in
 to the following three themes/questi
 ons:\n - Can we theoretically charac
 terize how much SGD-based pre-traini
 ng is sufficient to identify meaning
 ful sparse subnetworks?\n - Can we d
 rive connections between pruning met
 hods and classical sparse recovery, 
 in order to leverage decades of know
 ledge on theory for subset selection
 ?\n - From a practical standpoint, c
 an we combine the above ideas with e
 xisting efficient distributed protoc
 ols in order to achieve end-to-end s
 parse neural network training, even 
 avoiding heavy full-model pre-traini
 ng phases?\n\nBio:\n Anastasios Kyri
 llidis is a Noah Harding Assistant P
 rofessor at the Computer Science dep
 artment at Rice University. Prior to
  that, he was a Goldstine postdoctor
 al fellow at IBM T. J. Watson Resear
 ch Center (NY), and a Simons Foundat
 ion postdoc member at the University
  of Texas at Austin. He finished his
  PhD at the CS Department of EPFL (S
 witzerland) under the supervision of
  Volkan Cevher. Tasos got his M.Sc. 
 and Diploma from the Electronic and 
 Computer Engineering Dept. at the Te
 chnical University of Crete (Chania)
 . His research interests include (bu
 t not limited to): Optimization for 
 machine learning, convex and non-con
 vex algorithms and analysis, large-s
 cale optimization, any problem that 
 includes a math-driven criterion and
  requires an efficient method for it
 s solution.\n https://akyrillidis.gi
 thub.io\n  \n
STATUS:CONFIRMED
ORGANIZER;RSVP=FALSE;CN=TUC;CUTYPE=TUC:mailto:webmaster@tuc.gr
DTSTART:20220705T143000
DTEND:20220705T160000
TRANSP:OPAQUE
CLASS:DEFAULT
END:VEVENT
END:VCALENDAR