BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Generative models for black-box optimization of complex objectives
DTSTART:20260709T100000
DTEND:20260709T120000
DTSTAMP:20260523T201418Z
UID:63b9aafa362613fe6b72e05b05b9275dfe6757d5ea65b550b1925531
CATEGORIES:Conferences - Seminars
DESCRIPTION:Edouard Dufour\nEDIC candidacy exam\nExam president: Prof. Nic
 olas Flammarion\nThesis advisor: Prof. Pascal Fua\nCo-examiner: Prof. Alex
 ander Mathis\n\nAbstract\nMany high-impact problems in science and enginee
 ring reduce to optimizing a complex objective\, where each evaluation is c
 ostly and often unreliable: physical experiments fail\, simulators diverge
 \, and stochastic systems return noisy feedback. Classical black-box optim
 izers struggle in this regime\, where evaluation budgets are tight and the
  feasible region is hard to characterize a priori. Our work investigates h
 ow the statistical expressivity of modern generative samplers can be lever
 aged in black-box optimization. We propose that these samplers enable effi
 cient navigation of feasible sets and optimization of geometrically comple
 x objectives under limited\, unreliable observations. \n\nTwo complementa
 ry lines of work have already been studied. The main one\, SPARROW\, demon
 strated that a sampler-driven optimizer can substantially reduce the numbe
 r of function calls needed to maximize complex and unreliable objectives. 
 A parallel line established that samplers can be steered to satisfy hard c
 onstraints without sacrificing the statistical diversity of the generated 
 samples. Because SPARROW imposes minimal assumptions on the underlying sam
 pler\, these constrained samplers can drive it directly\, opening a path t
 o constrained black-box optimization.\n\nBuilding on these results\, two d
 irections are planned. The first is theoretical: a deeper mathematical stu
 dy of sampler-driven optimization\, drawing on the broader literature on s
 ample-efficient black-box optimization\, in order to derive stronger guara
 ntees and more efficient algorithms. The second is applied: applying the m
 ethods to concrete domains and exploiting the structure those domains expo
 se to sharpen performance beyond what a pure black-box treatment can achie
 ve.\n\nThe applications of low-budget black-box optimization are diverse\,
  including engineering design (e.g. aerodynamic shape optimization\, mecha
 nical components)\, scientific discovery (e.g. protein and molecule design
  under synthesizability and stability constraints)\, and the safety and se
 curity of opaque ML systems (e.g. query-efficient adversarial probing\, re
 d-teaming of black-box models).\n\nSelected papers\n 
LOCATION:BC 333 https://plan.epfl.ch/?room==BC%20333
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR