[ABE-L] Convite: Seminário DEST/UFMG em 21/10/2022.

[Vinicius Mayrink]

Caros,

Na próxima sexta (*21 de outubro*, às *13:30h*) o ciclo de Seminários
do *Departamento
de Estatística da UFMG* terá a apresentação de *Rafael Izbicki*.

Rafael é professor do Departamento de Estatística da UFSCar. Ele obteve o
grau de Doutor em Estatística pela Carnegie Mellon University (EUA). Suas
principais áreas de interesse são: Machine learning, Bioestatística,
Astroestatística, Fundamentos da estatística, Inferência Bayesiana,
Inferência não paramétrica e Inferência em Dados com alta dimensionalidade.

O seminário será transmitido ao vivo pelo canal do Youtube "*Seminários
DEST - UFMG <https://www.youtube.com/channel/UCoZC2_pME9ca_-Hx4djd60w>*".

At.te,
Vinícius Mayrink

*********** Título e Resumo ***********

Rafael Izbicki (Departamento de Estatística, UFSCar)

*Diagnostics and recalibration of predictive distributions.*

Uncertainty quantification is crucial for assessing the predictive ability
of AI algorithms. A large body of work (including normalizing flows and
Bayesian neural networks) has been devoted to describing the entire
predictive distribution (PD) of a target variable Y given input features X.
However, off-the-shelf PDs are usually far from being conditionally
calibrated; i.e., the probability of occurrence of an event given input X
can be significantly different from the predicted probability. Most current
research on predictive inference (such as conformal prediction) concerns
constructing calibrated prediction sets only. It is often believed that the
problem of obtaining and assessing entire conditionally calibrated PDs is
too challenging. In this work, we show that recalibration, as well as
validation of full/entire PDs, are indeed attainable goals in practice. Our
proposed method relies on the idea of regressing probability integral
transform (PIT) scores against X. This regression gives full diagnostics of
conditional coverage across the entire feature space and can be used to
recalibrate misspecified PDs. We benchmark our corrected prediction bands
against oracle bands and state-of-the-art predictive inference algorithms
for synthetic data, including settings with a distributional shift.
Finally, we produce calibrated PDs for two applications: (i) probabilistic
forecasting based on sequences of satellite images, and (ii) estimation of
galaxy distances based on imaging data (photometric redshifts).

-- 
*Vinícius D. Mayrink*
*Professor Associado - Departamento de Estatística*

*ICEx, Universidade Federal de Minas Gerais*
-------------- Próxima Parte ----------
Um anexo em HTML foi limpo...
URL: <http://lists.ime.usp.br/pipermail/abe/attachments/20221014/72085309/attachment.htm>