Closed bischtob closed 5 years ago
Here's a better paper (I think) regarding this: https://publications.aston.ac.uk/373/1/NCRG_94_004.pdf
It's a high priority item to me (independent of AI-611). Dropout alone provides crude uncertainty estimate but there's nothing in our current implementation that evaluates the uncertainty or enforces that we will have the right uncertainty.
Nhat, my intern, is implementing this for our RL project. You can ask him for tips to get it right.
Cool!
We need to have a custom loss function similar to the ones from this paper: https://arxiv.org/abs/1703.04977