BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:Séminaire 09/10/2026 - Gabriel Peyré and Vianney Perchet
DTSTART:20261009T121500Z
DTEND:20261009T143000Z
DTSTAMP:20261009T022100Z
UID:indico-event-16833@indico.math.cnrs.fr
DESCRIPTION:14h15-15h15 Gabriel Peyré (ENS)\nThe Expressive Power of Larg
 e Language Models\nAbstract: Large language models process vast sequences 
 of input tokens by alternating between classical multi-layer perceptron l
 ayers and self-attention mechanisms. While the approximation capabilities
  of perceptrons are relatively well understood\, those of attention mechan
 isms remain less explored. In this talk\, I will compare the proof techni
 ques and approximation results associated with these two types of layers\
 , emphasizing key open questions that connect large language models with 
 approximation theory in infinite-dimensional spaces representing input tok
 en distributions.\n \n15h30-16h30 Vianney Perchet (Criteo et ENSAE)\nAI-p
 owered AI. Robustness\, Consistency and Smoothness of Algorithms.\nAbstrac
 t: AI systems no longer operate as isolated algorithms\, machines\, or age
 nts\; instead\, theyinteract with each other. An algorithm designed for a 
 specific task might use the output orprediction of another as its input. T
 his has serious consequences: blindly relying on anotheralgorithm can be e
 xtremely helpful if it is accurate\, but devastating if it is not\, as a s
 ingle error -even the tiniest one - could break down the whole system. New
  algorithms will need to incorporatethese predictions to improve their con
 sistency\, while remaining robust against errors. We will usesimple toy ex
 amples to show how the generic framework of learning-augmented algorithms 
 (LAA) isperfectly suited for this challenge\, highlighting the complexity 
 of designing practical and potentiallyvastly more efficient algorithms.\n
  \n \n\nhttps://indico.math.cnrs.fr/event/16833/
LOCATION:Amphithéâtre Yvonne Choquet-Bruhat (IHP - Bâtiment Perrin)
URL:https://indico.math.cnrs.fr/event/16833/
END:VEVENT
END:VCALENDAR