nns-as-gps
Comparison of neural network ensembles and Monte Carlo dropout in their ability to mimic the behavior of a Gaussian process.
The ensemble methods implemented are:
- A vanilla ensemble of neural networks, each predicting a single value, and trained using MSE loss.
- An ensemble of neural networks that each predict a mean and variance, trained with NLL loss, as advised by this paper.
- Monte-Carlo dropout, as described in this paper
See also:
- Another implementation of ensembles with predictive uncertainty here
Author : Kunal Menda