SlowMo: Improving Communication-Efficient Distributed SGD with Slow Momentum

Abstract

Distributed optimization is essential for training large models on large datasets. Multiple approaches have been proposed to reduce the communication overhead in distributed training, such as synchronizing only after performing multiple local SGD steps, and decentralized methods (e.g., using gossip algorithms) to decouple communications among workers. Although these methods run faster than AllReduce-based methods, which use blocking communication before every update, the resulting models may be less accurate after the same number of updates. Inspired by the BMUF method of Chen & Huo (2016), we propose a slow momentum (SlowMo) framework, where workers periodically synchronize and perform a momentum update, after multiple iterations of a base optimization algorithm. Experiments on image classification and machine translation tasks demonstrate that SlowMo consistently yields improvements in optimization and generalization performance relative to the base optimizer, even when the additional overhead is amortized over many updates so that the SlowMo runtime is on par with that of the base optimizer. We provide theoretical convergence guarantees showing that SlowMo converges to a stationary point of smooth non-convex losses. Since BMUF can be expressed through the SlowMo framework, our results also correspond to the first theoretical convergence guarantees for BMUF.

Accepted to ICLR 2020

References in corpus (17)

• Communication-Efficient Learning of Deep Networks from Decentralized Data
• Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour
• QSGD: Communication-Efficient SGD via Gradient Quantization and Encoding
• Revisiting Distributed Synchronous SGD
• Lookahead Optimizer: k steps forward, 1 step back
• Don't Use Large Mini-Batches, Use Local SGD
• Decentralized Stochastic Optimization and Gossip Algorithms with Compressed Communication
• Cooperative SGD: A unified Framework for the Design and Analysis of Communication-Efficient SGD Algorithms
• Error Feedback Fixes SignSGD and other Gradient Compression Schemes
• On the Linear Speedup Analysis of Communication Efficient Momentum SGD for Distributed Non-Convex Optimization
• Stochastic Gradient Push for Distributed Deep Learning
• signSGD with Majority Vote is Communication Efficient And Fault Tolerant
• Understanding the Role of Momentum in Stochastic Gradient Methods
• MATCHA: Speeding Up Decentralized SGD via Matching Decomposition Sampling
• Decentralized Deep Learning with Arbitrary Communication Compression
• Anytime MiniBatch: Exploiting Stragglers in Online Distributed Optimization
• Accelerated Linear Convergence of Stochastic Momentum Methods in Wasserstein Distances

Cited by in corpus (21)

• PyTorch Distributed: Experiences on Accelerating Data Parallel Training
• An improved convergence analysis for decentralized online stochastic non-convex optimization
• Throughput-Optimal Topology Design for Cross-Silo Federated Learning
• Tighter Theory for Local SGD on Identical and Heterogeneous Data
• SemiFL: Semi-Supervised Federated Learning for Unlabeled Clients with Alternate Training
• Federated Continual Learning: Concepts, Challenges, and Solutions
• FedCM: Federated Learning with Client-level Momentum
• STEM: A Stochastic Two-Sided Momentum Algorithm Achieving Near-Optimal Sample and Communication Complexities for Federated Learning
• Byzantines can also Learn from History: Fall of Centered Clipping in Federated Learning
• Achieving Linear Speedup with Partial Worker Participation in Non-IID Federated Learning
• Joint Local Relational Augmentation and Global Nash Equilibrium for Federated Learning with Non-IID Data
• Faster Non-Convex Federated Learning via Global and Local Momentum
• Overlap Local-SGD: An Algorithmic Approach to Hide Communication Delays in Distributed SGD
• Accelerating Gossip SGD with Periodic Global Averaging
• Communication-Efficient Zeroth-Order Distributed Online Optimization: Algorithm, Theory, and Applications
• Cross-Gradient Aggregation for Decentralized Learning from Non-IID data
• CFedAvg: Achieving Efficient Communication and Fast Convergence in Non-IID Federated Learning
• Local Methods with Adaptivity via Scaling
• Multi-task Federated Edge Learning (MtFEEL) in Wireless Networks
• Local SGD for Near-Quadratic Problems: Improving Convergence under Unconstrained Noise Conditions
• Compositional federated learning: Applications in distributionally robust averaging and meta learning