Automatic hyperparameter selection for Lasso-like models solving inverse problems

Summary

We propose a hyperparameter selection technique based on SURE to automatically calibrate Lasso-like models solving inverse problems.

This repository contains the automatic calibrator and is used to demonstrate the superiority of this method on sparse models solving the M/EEG inverse problem. As part of our benchmark, we provide an implementation of two competitors: Lambda-MAP and temporal cross-validation.

An in-depth explanation can be found here: https://arxiv.org/abs/2112.12178

This work was accepted at the Medical Imaging Meets NeurIPS 2021 workshop.

Note that the default solver in MNE-Python for inverse problems is automatically calibrated using Monte Carlo Finite Difference (MCFD) SURE.

Installation

Start by installing the necessary requirements. We recommend creating a new venv or conda environment. Once created and activated, run

pip install -r requirements.txt

Then to install our package:

pip install -e .