Decentralized optimization method improves communication efficiency and accuracy
ZO-COSMO: Index-Free One-Hop Mixing for Decentralized Zeroth-Order Optimization
Machine Learning
Summary
Decentralized machine learning involves many agents working together without a central controller, but they need to communicate efficiently to learn well. This paper introduces a method called ZO-COSMO that allows agents to share only minimal information with their direct neighbors while still improving learning accuracy. The method carefully mixes information using a technique that preserves important signals and reduces errors. Experiments show that this approach improves performance on tasks like language understanding and works well on different network structures.
What this means in practice
- •For distributed machine learning teams: Enable communication-efficient decentralized learning with minimal coordination across nodes while maintaining accuracy in large-scale models.
- •For edge computing operators: Reduce data exchange overhead between edge devices performing joint optimization by sharing minimal, carefully combined updates.
Authors
Shengjun Zhang, Tingyi Liu, Heng Zhang, Dong Xie
Abstract
Sparse communication in decentralized zeroth-order learning requires compatible peer-state coordinates. We characterize this one-hop condition and develop \textsf{ZO-COSMO}, coupling two-query estimation with average-preserving masked consensus using $q$ values per active link. Global supports serve all-neighbor mixing; matching updates require agreement only within each pair. We derive a sharp contraction-per-scalar bound within the matching class and convergence guarantees for the core and sparse-momentum updates. At fixed matching, exact moment identities characterize how shared directions preserve gradient-heterogeneity cancellation and redistribute estimation error and disagreement. Mechanism experiments cover unequal curvatures, noise, and sparse momentum. Further tests span $64$ synthetic agents and eight logical Qwen LoRA workers. At matched payload budgets, Qwen2-7B QNLI gains $3.65$ accuracy points over explicit-index Rand-$k$; edge-local updates gain $3.42$ and $2.53$ points over all-neighbor mixing on eight-worker complete and ring graphs. A matched-first-step ablation gives a $3.92$-point momentum benefit. Seed-aware and same-matching controls distinguish encoding, scheduling, and query correlation.