Approximate Cross-Validation for Structured Models

This repository contains code for approximate cross-validation for structred models based on the infinitesimal jackknife as developed in [1].

Requirements

conda env create -f environment.yml
conda activate approxcv

Illustration

To perform LWCV on a AR(0) hidden Markov model trained on synthetic data,

python ./leave_within_structure_hmm.py

To perform LWCV on a discrete MRF,

python ./leave_within_structure_mrf.py

References

[1] Ghosh, Soumya*, William T. Stephenson*, Tin D. Nguyen, Sameer K. Deshpande, and Tamara Broderick. Approximate Cross-Validation for Structured Models. NeurIPS 2020.

^* Equal contribution