tagas/vcnn

Retrospective Uncertainties for Deep Models using Vine Copulas

This repository contains an implementation of the VCNN, as descibed in ["Paper submission 1289 - Retrospective Uncertainties for Deep Models using Vine Copulas"]

Visual abstract: We propose a plug in vine-copula module that can complement any neural network with a pre- diction interval, any time after the model has been trained, without requiring any modifications to it. Additionally, our prediction intervals capture both - aleatoric and epistemic uncertainty.

User guide

Dependencies

Download or clone this repository. VCNN relys on both R and Python packages. Hence, base R and the connecting package "rpy2" should be installed. Then within R, the following libraries are required: rvinecopulib, eecop, and stats.

Example

An example demo can be found in the notebook demo.ipynb.