CAFe compresses client updates against the previous aggregate, improving the DCGD convergence bound by (1-omega) without control variates, under equal step sizes and bounded heterogeneity.
The selected datasets are MNIST, EMNIST, and CIFAR-100, and we fol- low [17] to choose models for the three datasets, which are CONV4,CONV4, and ResNet-18, respectively
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Communication Compression for Distributed Learning without Control Variates
CAFe compresses client updates against the previous aggregate, improving the DCGD convergence bound by (1-omega) without control variates, under equal step sizes and bounded heterogeneity.