[ABE-L] Ciclo de Palestras - PPGE-IM/UFRJ - 20 de maio

[Carlos Zanini]

 Caros colegas,

A próxima palestra do Ciclo de Palestras do Programa de Pós-graduação em
Estatística (PPGE) da UFRJ ocorrerá na quarta-feira, 20 de maio, no horário
das 15h00 às 16h30. A palestra será transmitida a partir do Laboratório de
Sistemas Estocásticos (LSE), Sala I-044-B, Centro de Tecnologia - UFRJ, com
acesso remoto através do link:
meet.google.com/hno-grzn-ujb

*Palestrante:* Rafael Izbicki

*Título:* Calibrated, Epistemic-Aware, and Interpretable Uncertainty in
Machine Learning

*Resumo:* Machine learning has become a modern oracle: it produces
predictions across several disciplines with remarkable accuracy. But
perfect predictions are not always possible. Reliable decision-making,
therefore, requires more than a single answer. A good model should also
know what it does not know. This talk discusses uncertainty quantification
as a central ingredient of trustworthy machine learning. I will start from
conformal prediction, a distribution-free framework that turns point
predictions into prediction sets with finite-sample coverage guarantees
under exchangeability. I will then introduce CREDO, a method that combines
conformal calibration with credal uncertainty modeling for regression.
CREDO constructs epistemic-aware prediction intervals that widen in
data-sparse regions, remain calibrated after conformalization, and admit an
interpretable decomposition into aleatoric uncertainty, epistemic
uncertainty, and calibration slack. Finally, I will discuss how the same
calibration principle extends beyond prediction to statistical inference.

Att
Carlos.


-- 
----------------------------------------------------------------------------------------
Carlos Tadeu Pagani Zanini
Professor Adjunto
Departamento de Métodos Estatísticos - UFRJ
SIAPE: 1146101
carloszanini em dme.ufrj.br
carloszanini em im.ufrj.br
https://sites.google.com/view/ctpzanini
-------------------------------------------------------------------------------------------
-------------- Próxima Parte ----------
Um anexo em HTML foi limpo...
URL: <http://lists.ime.usp.br/pipermail/abe/attachments/20260517/1b40e16e/attachment.htm>