• English 
  • Spanish 

The importance of inductive bias in convolutional models for statistical downscaling

Conferencia: 9th International Workshop on Climate Informatics
Año: 2019
Tipo de contribución: Oral
Archivo PDF: 2019_Bano-Medina_CI.pdf
Poster: 2019_Bano-Medina_CI_b.pdf

Statistical downscaling is routinely used to produce regional climate change projections from coarse global model outputs. Along with the success of deep learning in multiple disciplines, recent studies outline the capability of deep neural models as a statistical downscaling technique. In this work, we analyze this problem from a multi-site perspective and highlight the benefits of a deep learning model. We argue that their merits are due to the existence of an inductive bias in multi-site architectures that prevents overfitting in overparameterized models with no need for dimensionality reduction techniques. We frame the experiment in the largest to date downscaling intercomparison study, called VALUE. The result is a better local reproducibility of multi-site deep neural models in comparison with single-site neural models and VALUE's benchmark methods.

Palabras clave: