Convex Optimization using Sparsified Stochastic Gradient Descent with Memory

The interest for distributed stochastic optimization has raised to train complex Machine Learning models with more data on distributed systems. Increasing the computation power speeds up the training but it faces a communication bottleneck between workers which hurts the scale-up of these distributed algorithms. Previous work tried to address this issue through quantization by broadcasting low precision updates (Seide et al., 2014; Alistarh et al., 2017) or through sparsification by sharing only the most important coordinates to update (Aji and Heafield, 2017; Lin et al., 2018). Even though the sparsifica- tion method works well in practice, it lacked a theoretical proof until now. We propose a sparsification scheme for SGD where only a small constant number of coordinates are applied at each iteration. The amount of data to be communicated is drastically reduced while the O(1/T ) convergence rate of vanilla SGD is preserved. Our concise proof extends to parallel setting and gains a linear speed up in the number of workers. We exper- iment with sparsified SGD with memory in C++ and Python, and show the excellent convergence properties for top-k and random-k operators. Our scheme outperforms QSGD in progress per number of bits sent. It also opens the path to using lock-free asynchronous parallelization (e.g. Hogwild! Niu et al. (2011)) on dense problems as the sparsity of the gradient updates is enforced.

Adaptation in Stochastic Algorithms: From Nonsmooth Optimization to Min-Max Problems and Beyond

Ahmet Alacaoglu

Stochastic gradient descent (SGD) and randomized coordinate descent (RCD) are two of the workhorses for training modern automated decision systems. Intriguingly, convergence properties of these methods are not well-established as we move away from the specific case of smooth minimization. In this dissertation, we focus on related problems of nonsmooth optimization and min-max optimization to improve the theoretical understanding of stochastic algorithms.First, we study SGD-based adaptive algorithms and propose a regret analysis framework overcoming the limitations of the existing ones in the convex case. In the nonconvex case, we prove convergence of an adaptive gradient algoritm for solving constrained weakly convex optimization, generalizing the previously known results on unconstrained smooth optimization. We also propose an algorithm combining Nesterov's smoothing with SGD to solve convex problems with infinitely many linear constraints, with optimal rates.Then, we move on to convex-concave min-max problems with bilinear coupling and analyze primal-dual coordinate descent (PDCD) algorithms. We obtain the first PDCD methods with the optimal

\mathcal{O}(1/k)

rate on the the standard optimality measure expected primal-dual gap, which was an open question since 2014. Our analysis also aims to explain the practical behavior of these algorithms by showing that the last iterate enjoys adaptive linear convergence without altering the parameters, depending on a certain error bound condition. Furthermore, we propose an algorithm combining the favorable properties of two branches of PDCD methods: the new method uses large step sizes with dense data and its per-iteration cost depends on the number of nonzeros of the data matrix. Thanks to these unique properties, this method enjoys compelling practical performance complementing its rigorous theoretical guarantees.Next, we consider monotone variational inequalities that generalize convex-concave min-max problems with nonbilinear coupling. We introduce variance reduced algorithms that converge under the same set of assumptions as their deterministic counterparts and improve the best-known complexities for solving convex-concave min-max problems with finite-sum structure. Optimality of our algorithms for this problem class is established in a recent work via matching lower bounds. Finally, we show our preliminary results on policy optimization methods for solving two player zero-sum Markov games for competitive reinforcement learning (RL). Even though this is a nonconvex-nonconcave min-max problem in general, thanks to the special structure, it is tractable to find an approximate Nash equilibrium. We introduce an algorithm that improves the best-known sample complexity of policy gradient methods. This development combines tools from RL and stochastic primal-dual optimization, showing the importance of techniques from convex-concave optimization.

EPFL2021

An Accelerated First-Order Method for Non-convex Optimization on Manifolds

Nicolas Boumal, Christopher Arnold Criscitiello

We describe the first gradient methods on Riemannian manifolds to achieve accelerated rates in the non-convex case. Under Lipschitz assumptions on the Riemannian gradient and Hessian of the cost function, these methods find approximate first-order critical points faster than regular gradient descent. A randomized version also finds approximate second-order critical points. Both the algorithms and their analyses build extensively on existing work in the Euclidean case. The basic operation consists in running the Euclidean accelerated gradient descent method (appropriately safe-guarded against non-convexity) in the current tangent space, then moving back to the manifold and repeating. This requires lifting the cost function from the manifold to the tangent space, which can be done for example through the Riemannian exponential map. For this approach to succeed, the lifted cost function (called the pullback) must retain certain Lipschitz properties. As a contribution of independent interest, we prove precise claims to that effect, with explicit constants. Those claims are affected by the Riemannian curvature of the manifold, which in turn affects the worst-case complexity bounds for our optimization algorithms.

SPRINGER2022

Stochastic distributed learning with gradient quantization and double-variance reduction

Konstantin Mishchenko, Sebastian Urban Stich

We consider distributed optimization over several devices, each sending incremental model updates to a central server. This setting is considered, for instance, in federated learning. Various schemes have been designed to compress the model updates in order to reduce the overall communication cost. However, existing methods suffer from a significant slowdown due to additional variance omega > 0 coming from the compression operator and as a result, only converge sublinearly. What is needed is a variance reduction technique for taming the variance introduced by compression. We propose the first methods that achieve linear convergence for arbitrary compression operators. For strongly convex functions with condition number kappa, distributed among n machines with a finite-sum structure, each worker having less than in components, we also (i) give analysis for the weakly convex and the non-convex cases and (ii) verify in experiments that our novel variance reduced schemes are more efficient than the baselines. Moreover, we show theoretically that as the number of devices increases, higher compression levels are possible without this affecting the overall number of communications in comparison with methods that do not perform any compression. This leads to a significant reduction in communication cost. Our general analysis allows to pick the most suitable compression for each problem, finding the rig ht balance between additional variance and communication savings. Finally, we also (iii) give analysis for arbitrary quantized updates.

TAYLOR & FRANCIS LTD2022

Adaptation in Stochastic Algorithms: From Nonsmooth Optimization to Min-Max Problems and Beyond

Ahmet Alacaoglu

\mathcal{O}(1/k)

EPFL2021