Simulation-based inference with scattering representations: scattering is all you need

TL;DR

This paper demonstrates that scattering representations can be effectively used for simulation-based inference with images, providing an interpretable, shift-resilient approach that outperforms traditional summary statistics without needing extra simulations.

Contribution

The paper introduces a scattering-based method for SBI that avoids additional simulations and enhances interpretability and robustness, especially in high-dimensional image data.

Findings

Scattering representations outperform traditional second order statistics.

The approach is resilient to covariate shift.

No extra simulations are needed for training or derivatives.

Abstract

We demonstrate the successful use of scattering representations without further compression for simulation-based inference (SBI) with images (i.e. field-level), illustrated with a cosmological case study. Scattering representations provide a highly effective representational space for subsequent learning tasks, although the higher dimensional compressed space introduces challenges. We overcome these through spatial averaging, coupled with more expressive density estimators. Compared to alternative methods, such an approach does not require additional simulations for either training or computing derivatives, is interpretable, and resilient to covariate shift. As expected, we show that a scattering only approach extracts more information than traditional second order summary statistics.