BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Electronic Structure Reading Group: Quantifying and propagating mo
 del-form uncertainty via misspecification-aware Bayesian regression
DTSTART:20250401T160000
DTEND:20250401T173000
DTSTAMP:20260411T101746Z
UID:8aa3e5196f9178177f14ad363ec700ac92f7203f02c86d5b1660e852
CATEGORIES:Conferences - Seminars
DESCRIPTION:Thomas D Swinburne\nEssentially all models are misspecified\, 
 i.e. no one parameter choice is able to exactly reproduce observations. As
  a result\, model parameters are fundamentally uncertain\, i.e. have model
 -form error\, as there is no unique “best choice”. In addition\, finit
 e capacity models such as polynomials or partially frozen neural networks 
 are often underparametrized\, i.e. the number of training data is much gre
 ater than the number of parameters\, meaning epistemic uncertainties are m
 inimal.\n\nI will discuss recent work[1] which treats the true generalisat
 ion error\, a misspecification-aware error measure for which the misspecif
 ication-blind log likelihood of Bayesian inference is only an (Hoeffding/J
 ensen) upper bound\, closely analogous to the Gibbs-Bogoliubov bound (F<U)
 . Whilst direct minimisation of the generalisation error is not tractable\
 , we derive a novel condition any valid posterior must obey\, then design 
 an ansatz valid for any model\, which we efficiently evaluate for linear (
 or linearised) models. Our method\, POPS[1\,2] provides a new form of post
 erior distribution on model parameters which enables robust *bounding* of 
 test errors. Fitting and inference has minimal (< x2) overhead compared to
  standard Bayesian regression schemes. Importantly\, by assigning error to
  parameters rather than simply the model output we have an excellent start
 ing point for propagation of model-form uncertainty through multi scale si
 mulations[3\,4\,5].\n\n[1]T. Swinburne\, D. Perez\, Mach. Learn.: Sci. Tec
 hnol.\, 6\, 015008 (2025)\n[2] https://github.com/tomswinburne/POPS-Regres
 sion.git <https://github.com/tomswinburne/POPSRegression.git>\n[3] I. Mali
 yov\, P. Grigorev\, T. Swinburne\, npj. Comput. Mater.\, 11\, 22 (2025)\n[
 4] D. Perez\, A. P. A. Subramanyam\, I. Maliyov\, T. D. Swinburne\, arXiv:
 2502.07104 (2025)\n[5] T. D. Swinburne\, C. Lapointe\, M-C. Marinica\, arX
 iv:2502.18191 (2025)\n\nThis event is jointly organized with the COSMO sem
 inar.\n---\nThe electronic structure reading group brings together researc
 hers and students interested in mathematical aspects of electronic structu
 re problems and adjacent topics\, including:\n\n	Density Functional Theory
 \n	Many-body Schrödinger equation for electrons\n	Born-Oppenheimer Molecu
 lar Dynamics\n	Numerical analysis and error control\n\nFor updates\, join 
 the matrix chat room at #electronic-structure:epfl.ch (requires a GASPAR
  account).\n\nWebsite: https://matmat.org/readinggroup/
LOCATION:MA B1 524 https://plan.epfl.ch/?room==MA%20B1%20524
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR