Subsampling is not Magic: Why Large Batch Sizes Work for Differentially Private Stochastic Optimisation
This repository contains the code and Mathematica notebooks for the "Subsampling is not Magic: Why Large Batch Sizes Work for Differentially Private Stochastic Optimisation" (ICML 2024) paper.
Install Poetry following the instructions here.
Run poetry install --without dev to install the dependencies.
cd ./src
Run sh ./plot.sh