How to Make the Gradients Small Privately: Improved Rates for Differentially Private Non-Convex Optimization | Synapse