BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20260408T100101EDT-1238fOEwL4@132.216.98.100
DTSTAMP:20260408T140101Z
DESCRIPTION:\n	ISS Informal Systems Seminar\n		Speaker: Petar Veličković – Sta
 ff Research Scientist\, DeepMind\, United Kingdom \n	\n\n\n \n	\n\n	Présentat
 ion sur YouTube\n\n	The last decade has witnessed an experimental revolutio
 n in data science and machine learning\, epitomised by deep learning metho
 ds. Indeed\, many high-dimensional learning tasks previously thought to be
  beyond reach –such as computer vision\, playing Go\, or protein folding –
  are in fact feasible with appropriate computational scale. Remarkably\, t
 he essence of deep learning is built from two simple algorithmic principle
 s: first\, the notion of representation or feature learning\, whereby adap
 ted\, often hierarchical\, features capture the appropriate notion of regu
 larity for each task\, and second\, learning by local gradient-descent typ
 e methods\, typically implemented as backpropagation.\n\n	While learning ge
 neric functions in high dimensions is a cursed estimation problem\, most t
 asks of interest are not generic\, and come with essential pre-defined reg
 ularities arising from the underlying low-dimensionality and structure of 
 the physical world. This talk is concerned with exposing these regularitie
 s through unified geometric principles that can be applied throughout a wi
 de spectrum of applications.\n\n	Such a ‘geometric unification’ endeavour i
 n the spirit of Felix Klein's Erlangen Program serves a dual purpose: on o
 ne hand\, it provides a common mathematical framework to study the most su
 ccessful neural network architectures\, such as CNNs\, RNNs\, GNNs\, and T
 ransformers. On the other hand\, it gives a constructive procedure to inco
 rporate prior physical knowledge into neural architectures and provide pri
 ncipled way to build future architectures yet to be invented.\n\n	Biography
 : Dr. Veličković is a Staff Research Scientist at DeepMind\, Affiliated Le
 cturer at the University of Cambridge\, and an Associate of Clare Hall\, C
 ambridge. He holds a PhD in Computer Science from the University of Cambri
 dge (Trinity College)\, obtained under the supervision of Pietro Liò. His 
 research concerns geometric deep learning—devising neural network architec
 tures that respect the invariances and symmetries in data. For his contrib
 utions\, he is recognised as an ELLIS Scholar in the Geometric Deep Learni
 ng Program. Particularly\, he focuses on graph representation learning and
  its applications in algorithmic reasoning (featured in VentureBeat). He i
 s the first author of Graph Attention Networks—a popular convolutional lay
 er for graphs—and Deep Graph Infomax—a popular self-supervised learning pi
 peline for graphs (featured in ZDNet). His research has been used in subst
 antially improving travel-time predictions in Google Maps (featured in the
  CNBC\, Endgadget\, VentureBeat\, CNET\, the Verge and ZDNet)\, and guidin
 g intuition of mathematicians towards new top-tier theorems and conjecture
 s (featured in Nature\, Science\, Quanta Magazine\, New Scientist\, The In
 dependent\, Sky News\, The Sunday Times\, la Repubblica and The Conversati
 on). See homepage.\n\n
DTSTART:20220916T180000Z
DTEND:20220916T190000Z
LOCATION:CA\, ZOOM
SUMMARY:Geometric Deep Learning
URL:https://www.mcgill.ca/channels/channels/event/geometric-deep-learning-3
 51686
END:VEVENT
END:VCALENDAR