I have a regression GAM (General Additive Model) and I want to learn its epistemic uncertainty( the variance of my residuals or predictions as a function of my input).
I have already used a bayesian approach to turn my GAM into a gaussian process so I can construct a covariance matrix but this approach is not scalable due to the high dimension of my problem.
I am trying to use an approach that uses the current model as a black-box and observe only the input and the residual, the closest thing that I found is quantile regression but I was wondering if there is any deep learning approach that learn the variance from the input.
Most of deep learning approach that I found estimate the mean and the variance simultaneously (Deep bayes, MVE, MC dropout ...)
A naive approach that I am implementing currently is a neural network that learn the variance as function of my input by minimizing the likelihood of my residuals as a centred gaussian but I didn't find any paper or ressources on this approach.
Do you have any idea on the problem, any possible ressources or an opinion on my current approach ?


1 Answer 1


Take a look on Google trust score - https://github.com/google/TrustScore

Based on the following paper - https://papers.nips.cc/paper/9547-can-you-trust-your-models-uncertainty-evaluating-predictive-uncertainty-under-dataset-shift.pdf

A relatively naive but useful approach


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.