On Gradient Coding With Partial Recovery

We consider a generalization of the gradient coding framework where dataset is divided across $n$ workers and each worker transmits to master node one or more linear combinations gradients over its assigned data subsets. Unlike conventional which requires recover sum all subsets in presence straggler workers, we relax goal computing at least some notation="LaTeX">$\alpha $ fraction gradients. begin by deriving lower bound on computation load any scheme also propose two strategies achieve this bound, albeit cost high communication number partitions can be polynomial . then schemes based cyclic assignment utilize have load. When single combination, prove bounds using partitions. Finally, describe class different intermediate operating points for provide simulation results demonstrate empirical performance our schemes.