BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Beyond Interpolation: Extrapolative Reasoning with Reinforcement L
 earning and Graph Neural Networks - Niccolò Grillo
DTSTART:20250224T184500Z
DTEND:20250224T191500Z
UID:TALK228859@talks.cam.ac.uk
CONTACT:Pietro Lio
DESCRIPTION:Despite incredible progress\, many neural architectures fail t
 o properly generalize beyond their training distribution. As such\, learni
 ng to reason in a correct and generalizable way is one of the current fund
 amental challenges in machine learning. In this respect\, logic puzzles pr
 ovide a great testbed\, as we can fully understand and control the learnin
 g environment. Thus\, they allow to evaluate performance on previously uns
 een\, larger and more difficult puzzles that follow the same underlying ru
 les. Since traditional approaches often struggle to represent such scalabl
 e logical structures\, we propose to model these puzzles using a graph-bas
 ed approach. Then\, we investigate the key factors enabling the proposed m
 odels to learn generalizable solutions in a reinforcement learning setting
 . Our study focuses on the impact of the inductive bias of the architectur
 e\, different reward systems and the role of recurrent modeling in enablin
 g sequential reasoning. Through extensive experiments\, we demonstrate how
  these elements contribute to successful extrapolation on increasingly com
 plex insights and frameworks offer a systematic way to design learning-bas
 ed systems capable of generalizable reasoning beyond interpolation.\n\n\ng
 oogle meet link:\nhttps://meet.google.com/vmn-iwhu-tas
LOCATION:Lecture Theatre 2\, Computer Laboratory\, William Gates Building
END:VEVENT
END:VCALENDAR