Scalable Marginal Likelihood Estimation for Model Selection in Deep Learning

  Alexander Immer 1 2 Matthias Bauer  3 4 Vincent Fortuin 1 Gunnar Rtsch 1 2 Mohammad Emtiyaz Khan 5

             Abstract                   3 layers, 5221 params     1 layer, 151 params
   Marginal-likelihood based model-selection, even           MargLik = 88        MargLik = 110
   though promising, is rarely used in deep learning
   due to estimation difficulties. Instead, most ap-
   proaches rely on validatio ...