Local SGD in Overparameterized Linear Regression

Mike Nguyen, Charly Kirst, Nicole Mücke

We consider distributed learning using constant stepsize SGD (DSGD) over several devices, each sending a final model update to a central server. In a final step, the local estimate…