Timezone: »
We provide the first importance sampling variants of variance reduced algorithms for empirical risk minimization with non-convex loss functions. In particular, we analyze non-convex versions of \texttt{SVRG}, \texttt{SAGA} and \texttt{SARAH}. Our methods have the capacity to speed up the training process by an order of magnitude compared to the state of the art on real datasets. Moreover, we also improve upon current mini-batch analysis of these methods by proposing importance sampling for minibatches in this setting. Surprisingly, our approach can in some regimes lead to superlinear speedup with respect to the minibatch size, which is not usually present in stochastic optimization. All the above results follow from a general analysis of the methods which works with {\em arbitrary sampling}, i.e., fully general randomized strategy for the selection of subsets of examples to be sampled in each iteration. Finally, we also perform a novel importance sampling analysis of \texttt{SARAH} in the convex setting.
Author Information
Samuel Horvath (KAUST)
I am currently an assistant professor of Machine Learning at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI). Prior to that, I completed my MS and Ph. D. in statistics at King Abdullah University of Science and Technology (KAUST), advised by professor Peter Richtárik. Before that, I was an undergraduate student in Financial Mathematics at Comenius University. In my research, I focus on providing a fundamental understanding of how distributed and federated training algorithms work and how they interact with different sources of heterogeneity, such as system-level variability in the computing infrastructure and statistical variability in the training data. Inspired by the theoretical insights, I aim to design efficient and practical distributed/federated training algorithms. I am broadly interested in federated learning, distributed optimization, and efficient deep learning.
Peter Richtarik (KAUST)
Peter Richtarik is an Associate Professor of Computer Science and Mathematics at KAUST and an Associate Professor of Mathematics at the University of Edinburgh. He is an EPSRC Fellow in Mathematical Sciences, Fellow of the Alan Turing Institute, and is affiliated with the Visual Computing Center and the Extreme Computing Research Center at KAUST. Dr. Richtarik received his PhD from Cornell University in 2007, and then worked as a Postdoctoral Fellow in Louvain, Belgium, before joining Edinburgh in 2009, and KAUST in 2017. Dr. Richtarik's research interests lie at the intersection of mathematics, computer science, machine learning, optimization, numerical linear algebra, high performance computing and applied probability. Through his recent work on randomized decomposition algorithms (such as randomized coordinate descent methods, stochastic gradient descent methods and their numerous extensions, improvements and variants), he has contributed to the foundations of the emerging field of big data optimization, randomized numerical linear algebra, and stochastic methods for empirical risk minimization. Several of his papers attracted international awards, including the SIAM SIGEST Best Paper Award, the IMA Leslie Fox Prize (2nd prize, twice), and the INFORMS Computing Society Best Student Paper Award (sole runner up). He is the founder and organizer of the Optimization and Big Data workshop series.
Related Events (a corresponding poster, oral, or spotlight)
-
2019 Poster: Nonconvex Variance Reduced Optimization with Arbitrary Sampling »
Wed. Jun 12th 01:30 -- 04:00 AM Room Pacific Ballroom #95
More from the Same Authors
-
2021 : FjORD: Fair and Accurate Federated Learning under heterogeneous targets with Ordered Dropout »
Samuel Horvath · Stefanos Laskaridis · Ilias Leontiadis -
2021 : EF21: A New, Simpler, Theoretically Better, and Practically Faster Error Feedback »
Peter Richtarik · Peter Richtarik · Ilyas Fatkhullin -
2023 : Improving Accelerated Federated Learning with Compression and Importance Sampling »
Michał Grudzień · Grigory Malinovsky · Peter Richtarik -
2023 : Federated Learning with Regularized Client Participation »
Grigory Malinovsky · Samuel Horváth · Konstantin Burlachenko · Peter Richtarik -
2023 : Federated Optimization Algorithms with Random Reshuffling and Gradient Compression »
Abdurakhmon Sadiev · Grigory Malinovsky · Eduard Gorbunov · Igor Sokolov · Ahmed Khaled · Konstantin Burlachenko · Peter Richtarik -
2023 : Momentum Provably Improves Error Feedback! »
Ilyas Fatkhullin · Alexander Tyurin · Peter Richtarik -
2023 : ELF: Federated Langevin Algorithms with Primal, Dual and Bidirectional Compression »
Avetik Karagulyan · Peter Richtarik -
2023 : Towards a Better Theoretical Understanding of Independent Subnetwork Training »
Egor Shulgin · Peter Richtarik -
2023 : Convergence of First-Order Algorithms for Meta-Learning with Moreau Envelopes »
Konstantin Mishchenko · Slavomír Hanzely · Peter Richtarik -
2023 Poster: High-Probability Bounds for Stochastic Optimization and Variational Inequalities: the Case of Unbounded Variance »
Abdurakhmon Sadiev · Marina Danilova · Eduard Gorbunov · Samuel Horváth · Gauthier Gidel · Pavel Dvurechenskii · Alexander Gasnikov · Peter Richtarik -
2023 Poster: EF21-P and Friends: Improved Theoretical Communication Complexity for Distributed Optimization with Bidirectional Compression »
Kaja Gruntkowska · Alexander Tyurin · Peter Richtarik -
2022 Poster: Proximal and Federated Random Reshuffling »
Konstantin Mishchenko · Ahmed Khaled · Peter Richtarik -
2022 Poster: A Convergence Theory for SVGD in the Population Limit under Talagrand's Inequality T1 »
Adil Salim · Lukang Sun · Peter Richtarik -
2022 Poster: 3PC: Three Point Compressors for Communication-Efficient Distributed Training and a Better Theory for Lazy Aggregation »
Peter Richtarik · Igor Sokolov · Elnur Gasanov · Ilyas Fatkhullin · Zhize Li · Eduard Gorbunov -
2022 Spotlight: A Convergence Theory for SVGD in the Population Limit under Talagrand's Inequality T1 »
Adil Salim · Lukang Sun · Peter Richtarik -
2022 Spotlight: Proximal and Federated Random Reshuffling »
Konstantin Mishchenko · Ahmed Khaled · Peter Richtarik -
2022 Spotlight: 3PC: Three Point Compressors for Communication-Efficient Distributed Training and a Better Theory for Lazy Aggregation »
Peter Richtarik · Igor Sokolov · Elnur Gasanov · Ilyas Fatkhullin · Zhize Li · Eduard Gorbunov -
2022 Poster: ProxSkip: Yes! Local Gradient Steps Provably Lead to Communication Acceleration! Finally! »
Konstantin Mishchenko · Grigory Malinovsky · Sebastian Stich · Peter Richtarik -
2022 Spotlight: ProxSkip: Yes! Local Gradient Steps Provably Lead to Communication Acceleration! Finally! »
Konstantin Mishchenko · Grigory Malinovsky · Sebastian Stich · Peter Richtarik -
2022 Poster: FedNL: Making Newton-Type Methods Applicable to Federated Learning »
Mher Safaryan · Rustem Islamov · Xun Qian · Peter Richtarik -
2022 Spotlight: FedNL: Making Newton-Type Methods Applicable to Federated Learning »
Mher Safaryan · Rustem Islamov · Xun Qian · Peter Richtarik -
2021 : Closing Remarks »
Shiqiang Wang · Nathalie Baracaldo · Olivia Choudhury · Gauri Joshi · Peter Richtarik · Praneeth Vepakomma · Han Yu -
2021 : Opening Remarks »
Shiqiang Wang · Nathalie Baracaldo · Olivia Choudhury · Gauri Joshi · Peter Richtarik · Praneeth Vepakomma · Han Yu -
2021 Poster: ADOM: Accelerated Decentralized Optimization Method for Time-Varying Networks »
Dmitry Kovalev · Egor Shulgin · Peter Richtarik · Alexander Rogozin · Alexander Gasnikov -
2021 Spotlight: ADOM: Accelerated Decentralized Optimization Method for Time-Varying Networks »
Dmitry Kovalev · Egor Shulgin · Peter Richtarik · Alexander Rogozin · Alexander Gasnikov -
2021 Poster: MARINA: Faster Non-Convex Distributed Learning with Compression »
Eduard Gorbunov · Konstantin Burlachenko · Zhize Li · Peter Richtarik -
2021 Spotlight: MARINA: Faster Non-Convex Distributed Learning with Compression »
Eduard Gorbunov · Konstantin Burlachenko · Zhize Li · Peter Richtarik -
2021 Poster: PAGE: A Simple and Optimal Probabilistic Gradient Estimator for Nonconvex Optimization »
Zhize Li · Hongyan Bao · Xiangliang Zhang · Peter Richtarik -
2021 Poster: Stochastic Sign Descent Methods: New Algorithms and Better Theory »
Mher Safaryan · Peter Richtarik -
2021 Poster: Distributed Second Order Methods with Fast Rates and Compressed Communication »
Rustem Islamov · Xun Qian · Peter Richtarik -
2021 Spotlight: Distributed Second Order Methods with Fast Rates and Compressed Communication »
Rustem Islamov · Xun Qian · Peter Richtarik -
2021 Oral: PAGE: A Simple and Optimal Probabilistic Gradient Estimator for Nonconvex Optimization »
Zhize Li · Hongyan Bao · Xiangliang Zhang · Peter Richtarik -
2021 Spotlight: Stochastic Sign Descent Methods: New Algorithms and Better Theory »
Mher Safaryan · Peter Richtarik -
2020 Poster: Stochastic Subspace Cubic Newton Method »
Filip Hanzely · Nikita Doikov · Yurii Nesterov · Peter Richtarik -
2020 Poster: Variance Reduced Coordinate Descent with Acceleration: New Method With a Surprising Application to Finite-Sum Problems »
Filip Hanzely · Dmitry Kovalev · Peter Richtarik -
2020 Poster: Acceleration for Compressed Gradient Descent in Distributed and Federated Optimization »
Zhize Li · Dmitry Kovalev · Xun Qian · Peter Richtarik -
2020 Poster: From Local SGD to Local Fixed-Point Methods for Federated Learning »
Grigory Malinovsky · Dmitry Kovalev · Elnur Gasanov · Laurent CONDAT · Peter Richtarik -
2019 Poster: SAGA with Arbitrary Sampling »
Xun Qian · Zheng Qu · Peter Richtarik -
2019 Poster: SGD: General Analysis and Improved Rates »
Robert Gower · Nicolas Loizou · Xun Qian · Alibek Sailanbayev · Egor Shulgin · Peter Richtarik -
2019 Oral: SAGA with Arbitrary Sampling »
Xun Qian · Zheng Qu · Peter Richtarik -
2019 Oral: SGD: General Analysis and Improved Rates »
Robert Gower · Nicolas Loizou · Xun Qian · Alibek Sailanbayev · Egor Shulgin · Peter Richtarik -
2018 Poster: SGD and Hogwild! Convergence Without the Bounded Gradients Assumption »
Lam Nguyen · PHUONG_HA NGUYEN · Marten van Dijk · Peter Richtarik · Katya Scheinberg · Martin Takac -
2018 Oral: SGD and Hogwild! Convergence Without the Bounded Gradients Assumption »
Lam Nguyen · PHUONG_HA NGUYEN · Marten van Dijk · Peter Richtarik · Katya Scheinberg · Martin Takac