Subsampling is not Magic: Why Large Batch Sizes Work for Differentially Private Stochastic Optimisation

Open in new window