Speaker

Carolina Musso

Carolina Musso

Statistitian at the Instituto de Pesquisa e Estatística do Distrito Federal

Brasília, Brazil

Actions

Biologist and Statistician, with a phD in Ecology and specialization in Data Science. Working as data analysis in public service for the past seven years.

Area of Expertise

  • Environment & Cleantech
  • Government, Social Sector & Education
  • Physical & Life Sciences

Topics

  • Epidemiology
  • Public Health
  • Biostatistics
  • Biological Invasions
  • Machine Leaning
  • Computational Statistics

Recalibration of Gaussian Neural Network regression models: the recalibratiNN package

Artificial Neural Networks (ANNs) are powerful models in representation learning, known for their high prediction performance. However, their predictions often lack proper calibration, leading to unreliable uncertainty estimation. Various methods have been proposed to address this issue, but implementing and selecting the appropriate technique can be challenging. Therefore, this work presents an R package that offers a computational implementation of a quantile-based post-processing technique for recalibration. We hope this package will encourage researchers to pursue reliable uncertainty estimates from their models. The current version of the package provides functions to recalibrate Gaussian models, such as neural networks adjusted with the MSE loss function. The method leverages information from cumulative probabilities, enabling the generation of Monte Carlo samples from the recalibrated predictive distribution and performing recalibration locally and globally (https://arxiv.org/abs/2403.05756). The so-called recalibratriNN package also includes diagnostic functions to visualize miscalibration. It is readily available on GitHub (https://cmusso86.github.io/recalibratiNN/).

useR! Sessionize Event

July 2024 Salzburg, Austria

Carolina Musso

Statistitian at the Instituto de Pesquisa e Estatística do Distrito Federal

Brasília, Brazil

Actions

Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.

Jump to top