BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Forward Pass as Heat Flow - Kartik Tandon
DTSTART:20260403T140000Z
DTEND:20260403T150000Z
UID:TALK245311@talks.cam.ac.uk
CONTACT:Pietro Lio
DESCRIPTION:Strong machine learning models have demonstrated a remarkable 
 ability to leverage the underlying geometric and topological structure of 
 datasets. This has been observed not just in explicitly geometric domains 
 (such as graph or mesh-based data)\, but even when this underlying structu
 re is implicit (eg satisfies the manifold hypothesis). In this talk\, we s
 hall explore the unifying perspective that both regimes may be understood 
 as performing heat diffusion intrinsic to the underlying geometry in the m
 odel’s forward pass. As examples of this philosophy\, we will discuss a 
 far-reaching generalization of the convergence results of Belkin-Niyogi th
 at unites several geometric deep learning architectures as well as a manif
 old-theoretic framework underlying the ‘emergent’ ability of in-contex
 t learning in large models.
LOCATION:Computer Laboratory\, William Gates Building\, Room LT1
END:VEVENT
END:VCALENDAR