Próximas Charlas 

31 de Mayo, 2024

Alejandra Cabaña (UAB- España)

An attempt to test for feature importance in Neural Networks via their limiting Gaussian Process Regression

Resumen

Neural networks have a tremendous empirical success. But some applications call for an explanation, rather than a black-box prediction.   A not-so-recent paper by Matheus et.al (https://arxiv.org/abs/1804.11271) studies the relationship between Gaussian processes with a recursive kernel deffinition and a random multi-layer-perceptron.   I will show an easy tool for computing the covariance kernel of the limiting Gaussian process, and an attempt to design a test of model comparison.