Stochastic Sign Descent Methods: New Algorithms and Better Theory

                       Mher Safaryan 1 Peter Richtárik 1 2

              Abstract               hence the training data is typically split and stored across
                                 a number of compute nodes capable of working in parallel.
    Various gradient compression schemes have been      Training such models then amounts to solving optimization
    proposed to mitigate the communication cost        problems of the ...