BEGIN:VCALENDAR

VERSION:2.0

PRODID:-//wp-events-plugin.com//7.4.2//EN

TZID:Asia/Jerusalem

X-WR-TIMEZONE:Asia/Jerusalem
BEGIN:VEVENT

UID:499@math.technion.ac.il

DTSTART;TZID=Asia/Jerusalem:20251027T140000

DTEND;TZID=Asia/Jerusalem:20251027T150000

DTSTAMP:20251015T055712Z

URL:https://math.technion.ac.il/events/short-range-oversquashing/

SUMMARY:Short-Range Oversquashing
DESCRIPTION:Lecturer:Yakov Mishayev\n Location:Amado 814\n Message Passing 
 Neural Networks (MPNNs) are widely used for learning on\n\ngraphs\, but th
 eir ability to process long-range information is limited by the\n\nphenome
 non of oversquashing. This limitation has led some researchers to\n\nadvoc
 ate Graph Transformers as a better alternative\, whereas others suggest th
 at\n\nit can be mitigated within the MPNN framework\, using virtual nodes 
 or other\n\nrewiring techniques.\n\nIn this work\, we demonstrate that ove
 rsquashing is not limited to long-range\n\ntasks\, but can also arise in s
 hort-range problems. This observation allows us\n\nto disentangle two dist
 inct mechanisms underlying oversquashing: (1) the bot-\n\ntleneck phenomen
 on\, which can arise even in low-range settings\, and (2) the\n\nvanishing
  gradient phenomenon\, which is closely associated with long-range\n\ntask
 s.\n\nWe further show that the short-range bottleneck effect is not captur
 ed by existing\n\nexplanations for oversquashing\, and that adding virtual
  nodes does not resolve it.\n\nIn contrast\, transformers do succeed in su
 ch tasks\, positioning them as the more\n\ncompelling solution to oversqua
 shing\, compared to specialized MPNNs.\n\n&nbsp\;\n\nAdvisor: Dr. Nadav Dy
 m\n\n&nbsp\;\n 
CATEGORIES:Graduation Seminars,Seminars
LOCATION:Amado 814

END:VEVENT

BEGIN:VTIMEZONE

TZID:Asia/Jerusalem

X-LIC-LOCATION:Asia/Jerusalem

BEGIN:STANDARD

DTSTART:20251026T010000

TZOFFSETFROM:+0300

TZOFFSETTO:+0200

TZNAME:IST

END:STANDARD

END:VTIMEZONE
END:VCALENDAR