Tight analyses of first-order methods with error feedback

Aymeric Dieuleveut (Ecole Polytechnique, IPParis) · Daniel Berg Thomsen (INRIA) · Adrien Taylor (Inria)
communication overheadcompressed gradient descentcomputational bottleneckconvergence degradationconvergence ratedistributed learningef methodef21 methoderror feedback schemesinformation compressionlyapunov functionperformance guaranteesrigorous comparisonsingle-agent settingtheoretical insightsunderlying mechanisms

Communication between agents often constitutes a major computational bottleneck in distributed learning. One of the most common mitigation strategies is to compress the information exchanged, thereby reducing communication overhead. To counteract the degradation in convergence associated with compressed communication, error feedback schemes