Variance Reduced EXTRA and DIGing and Their Optimal Acceleration for Strongly Convex Decentralized Optimization
This work addresses the problem of efficient decentralized training for large-scale distributed data in machine learning, offering incremental improvements over existing methods.
The authors tackled decentralized optimization for training machine learning models with distributed data by extending EXTRA and DIGing methods with variance reduction, achieving improved complexities in stochastic gradient evaluations and communication rounds, with VR-EXTRA requiring O((κ_s+n)log(1/ε)) stochastic gradient evaluations and O((κ_b+κ_c)log(1/ε)) communication rounds, and accelerated versions achieving optimal O((√(nκ_s)+n)log(1/ε)) and O(√(κ_bκ_c)log(1/ε)) complexities.
We study stochastic decentralized optimization for the problem of training machine learning models with large-scale distributed data. We extend the widely used EXTRA and DIGing methods with variance reduction (VR), and propose two methods: VR-EXTRA and VR-DIGing. The proposed VR-EXTRA requires the time of $O((κ_s+n)\log\frac{1}ε)$ stochastic gradient evaluations and $O((κ_b+κ_c)\log\frac{1}ε)$ communication rounds to reach precision $ε$, which are the best complexities among the non-accelerated gradient-type methods, where $κ_s$ and $κ_b$ are the stochastic condition number and batch condition number for strongly convex and smooth problems, respectively, $κ_c$ is the condition number of the communication network, and $n$ is the sample size on each distributed node. The proposed VR-DIGing has a little higher communication cost of $O((κ_b+κ_c^2)\log\frac{1}ε)$. Our stochastic gradient computation complexities are the same as the ones of single-machine VR methods, such as SAG, SAGA, and SVRG, and our communication complexities keep the same as those of EXTRA and DIGing, respectively. To further speed up the convergence, we also propose the accelerated VR-EXTRA and VR-DIGing with both the optimal $O((\sqrt{nκ_s}+n)\log\frac{1}ε)$ stochastic gradient computation complexity and $O(\sqrt{κ_bκ_c}\log\frac{1}ε)$ communication complexity. Our stochastic gradient computation complexity is also the same as the ones of single-machine accelerated VR methods, such as Katyusha, and our communication complexity keeps the same as those of accelerated full batch decentralized methods, such as MSDA.