Status: Archive (code is provided as-is, no updates expected)
MC-dropout estimates uncertainty at test time using the variance statistics extracted from several dropout-enabled forward passes. Unfortunately, the prediction cost of an effective MC-dropout can reach hundreds of feed-forward iterations for each prediction. In this repository, I model MC-dropout in a Deep Reinforcement Learning (DRL) framework, to find the optimial passes needed for producing predefined confidence level.
- For creating the Open-AI gym environment, I followed the code from Making a custom environment in gym
- For MC-dropout expriments, I used the code from Yarin Gal - DropoutUncertaintyExps