A multilevel Monte Carlo framework debiases biased gradient compressors, preserving SGD convergence guarantees while reducing communication cost, with adaptive variance-minimizing level selection.
1 (pl t,i)2 (∆l t,i)2 # (47) amd by writing the expectation w.r.t pl explicitly, we have: E[∥˜gt,i∥2] = LX l=1
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Beyond Communication Overhead: A Multilevel Monte Carlo Approach for Mitigating Compression Bias in Distributed Learning
A multilevel Monte Carlo framework debiases biased gradient compressors, preserving SGD convergence guarantees while reducing communication cost, with adaptive variance-minimizing level selection.