MC Dropout Notebooks

Source code for the ICML2021 workshop paper Notes on the behavior of MC Dropout

Authors: Francesco Verdoja
Maintainer: Francesco Verdoja, francesco.verdoja@aalto.fi
Affiliation: Intelligent Robotics Lab, Aalto University

Notebooks

This repository contains all Jupyter notebooks needed to replicate the experiments in the original paper.

Each notebook is named according to the following convention:

f_{shape}_{bias}_{dropout_rate}.ipynb

• shape describes the function used in the test (among the ones presented in the paper)
• bias can be either bias or nobias if the network has bias on the last layer or not.
• dropout_rate is either 02p if $p_d = 0.2$, or 05p if $p_d = 0.5$

Reference

If you end up using parts or the entirety of the code in this repository for academic research, please reference the original paper:

F. Verdoja and V. Kyrki, “Notes on the behavior of MC Dropout,” presented at the ICML 2021 Workshop on Uncertainty and Robustness in Deep Learning, Jul. 23, 2021 [Online]. Available: https://arxiv.org/abs/2008.02627