More Like this

1. On the Convergence of Communication-Efficient Local SGD for Federated Learning

   Hongchang Gao, An Xu (February 2021, 35th AAAI Conference on Artificial Intelligence (AAAI 2021))

   Federated Learning (FL) has attracted increasing attention in recent years. A leading training algorithm in FL is local SGD, which updates the model parameter on each worker and averages model parameters across different workers only once in a while. Although it has fewer communication rounds than the classical parallel SGD, local SGD still has large communication overhead in each communication round for large machine learning models, such as deep neural networks. To address this issue, we propose a new communicationefficient distributed SGD method, which can significantly reduce the communication cost by the error-compensated double compression mechanism. Under the non-convex setting, our theoretical results show that our approach has better communication complexity than existing methods and enjoys the same linear speedup regarding the number of workers as the full-precision local SGD. Moreover, we propose a communication-efficient distributed SGD with momentum, which also has better communication complexity than existing methods and enjoys a linear speedup with respect to the number of workers. At last, extensive experiments are conducted to verify the performance of our proposed methods. 

   more » « less
2. Byzantine-Resilient SGD in High Dimensions on Heterogeneous Data

   https://doi.org/10.1109/ISIT45174.2021.9518248  

   Data, Deepesh; Diggavi, Suhas (July 2021, IEEE International Symposium on Information Theory (ISIT))

   We study distributed stochastic gradient descent (SGD) in the master-worker architecture under Byzantine attacks. We consider the heterogeneous data model, where different workers may have different local datasets, and we do not make any probabilistic assumptions on data generation. At the core of our algorithm, we use the polynomial-time outlier-filtering procedure for robust mean estimation proposed by Steinhardt et al. (ITCS 2018) to filter-out corrupt gradients. In order to be able to apply their filtering procedure in our heterogeneous data setting where workers compute stochastic gradients, we derive a new matrix concentration result, which may be of independent interest. We provide convergence analyses for smooth strongly-convex and non-convex objectives and show that our convergence rates match that of vanilla SGD in the Byzantine-free setting. In order to bound the heterogeneity, we assume that the gradients at different workers have bounded deviation from each other, and we also provide concrete bounds on this deviation in the statistical heterogeneous data model. 

   more » « less
3. Byzantine-Resilient SGD in High Dimensions on Heterogeneous Data

   Data, Deepesh; Diggavi, Suhas N. (July 2021, IEEE International Symposium on Information Theory (ISIT))

   We study distributed stochastic gradient descent (SGD) in the master-worker architecture under Byzantine at- tacks. We consider the heterogeneous data model, where different workers may have different local datasets, and we do not make any probabilistic assumptions on data generation. At the core of our algorithm, we use the polynomial-time outlier-filtering procedure for robust mean estimation proposed by Steinhardt et al. (ITCS 2018) to filter-out corrupt gradients. In order to be able to apply their filtering procedure in our heterogeneous data setting where workers compute stochastic gradients, we derive a new matrix concentration result, which may be of independent interest. We provide convergence analyses for smooth strongly- convex and non-convex objectives and show that our convergence rates match that of vanilla SGD in the Byzantine-free setting. In order to bound the heterogeneity, we assume that the gradients at different workers have bounded deviation from each other, and we also provide concrete bounds on this deviation in the statistical heterogeneous data model. 

   more » « less
4. Symmetry Teleportation for Accelerated Optimization

   Bo Zhao; Nima Dehmamy; Robin Walters; Rose Yu (January 2022, Advances in neural information processing systems)

   Existing gradient-based optimization methods update parameters locally, in a direction that minimizes the loss function. We study a different approach, symmetry teleportation, that allows parameters to travel a large distance on the loss level set, in order to improve the convergence speed in subsequent steps. Teleportation exploits symmetries in the loss landscape of optimization problems. We derive loss-invariant group actions for test functions in optimization and multi-layer neural networks, and prove a necessary condition for teleportation to improve convergence rate. We also show that our algorithm is closely related to second order methods. Experimentally, we show that teleportation improves the convergence speed of gradient descent and AdaGrad for several optimization problems including test functions, multi-layer regressions, and MNIST classification. 

   more » « less
5. Information Theoretic Limits of Data Shuffling for Distributed Learning

   https://doi.org/10.1109/GLOCOM.2016.7841903  

   Attia, Mohamed Adel; Tandon, Ravi (December 2016, IEEE Globecom)

   Data shuffling is one of the fundamental building blocks for distributed learning algorithms, that increases the statistical gain for each step of the learning process. In each iteration, different shuffled data points are assigned by a central node to a distributed set of workers to perform local computations, which leads to communication bottlenecks. The focus of this paper is on formalizing and understanding the fundamental information-theoretic tradeoff between storage (per worker) and the worst-case communication overhead for the data shuffling problem. We completely characterize the information theoretic tradeoff for K = 2, and K = 3 workers, for any value of storage capacity, and show that increasing the storage across workers can reduce the communication overhead by leveraging coding. We propose a novel and systematic data delivery and storage update strategy for each data shuffle iteration, which preserves the structural properties of the storage across the workers, and aids in minimizing the communication overhead in subsequent data shuffling iterations. 

   more » « less