SGD with memory: fundamental properties and stochastic acceleration

• URL: http://arxiv.org/abs/2410.04228v1
• Date: Sat, 5 Oct 2024 16:57:40 GMT
• Title: SGD with memory: fundamental properties and stochastic acceleration
• Authors: Dmitry Yarotsky, Maksim Velikanov,
• Abstract summary: In non-stochastic setting, the optimal exponent $xi$ in the loss convergence $L_tsim C_Lt-xi is double that in plain GD. We show that in memory-1 algorithms we can make $C_L$ arbitrarily small while maintaining stability.
• Score: 15.25021403154845
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: An important open problem is the theoretically feasible acceleration of mini-batch SGD-type algorithms on quadratic problems with power-law spectrum. In the non-stochastic setting, the optimal exponent $\xi$ in the loss convergence $L_t\sim C_Lt^{-\xi}$ is double that in plain GD and is achievable using Heavy Ball (HB) with a suitable schedule; this no longer works in the presence of mini-batch noise. We address this challenge by considering first-order methods with an arbitrary fixed number $M$ of auxiliary velocity vectors (*memory-$M$ algorithms*). We first prove an equivalence between two forms of such algorithms and describe them in terms of suitable characteristic polynomials. Then we develop a general expansion of the loss in terms of signal and noise propagators. Using it, we show that losses of stationary stable memory-$M$ algorithms always retain the exponent $\xi$ of plain GD, but can have different constants $C_L$ depending on their effective learning rate that generalizes that of HB. We prove that in memory-1 algorithms we can make $C_L$ arbitrarily small while maintaining stability. As a consequence, we propose a memory-1 algorithm with a time-dependent schedule that we show heuristically and experimentally to improve the exponent $\xi$ of plain SGD.

Related papers

• MGDA Converges under Generalized Smoothness, Provably [27.87166415148172]
  Multi-objective optimization (MOO) is receiving more attention in various fields such as multi-task learning. Recent works provide some effective algorithms with theoretical analysis but they are limited by the standard $L$-smooth or bounded-gradient assumptions. We study a more general and realistic class of generalized $ell$-smooth loss functions, where $ell$ is a general non-decreasing function of gradient norm.
  arXiv  Detail & Related papers  (2024-05-29T18:36:59Z)
• Fast Minimization of Expected Logarithmic Loss via Stochastic Dual Averaging [8.990961435218544]
  We propose a first-order algorithm named $B$-sample dual averaging with the logarithmic barrier. For the Poisson inverse problem, our algorithm attains an $varepsilon$ solution in $smashtildeO(d3/varepsilon2)$ time. When computing the maximum-likelihood estimate for quantum state tomography, our algorithm yields an $varepsilon$-optimal solution in $smashtildeO(d3/varepsilon2)$ time.
  arXiv  Detail & Related papers  (2023-11-05T03:33:44Z)
• A Doubly Robust Approach to Sparse Reinforcement Learning [19.68978899041642]
  We propose a new regret algorithm for episodic sparse linear Markov decision process (SMDP) The proposed algorithm is $tildeO(sigma-1_min s_star H sqrtN)$, where $sigma_min$ denotes the restrictive minimum eigenvalue of the average Gram matrix of feature vectors.
  arXiv  Detail & Related papers  (2023-10-23T18:52:17Z)
• Weak Schur sampling with logarithmic quantum memory [0.0]
  We introduce a new algorithm for the task of weak Schur sampling. Our algorithm efficiently determines both the Young label which indexes the irreducible representations and the multiplicity label of the symmetric group.
  arXiv  Detail & Related papers  (2023-09-21T10:02:46Z)
• An Oblivious Stochastic Composite Optimization Algorithm for Eigenvalue Optimization Problems [76.2042837251496]
  We introduce two oblivious mirror descent algorithms based on a complementary composite setting. Remarkably, both algorithms work without prior knowledge of the Lipschitz constant or smoothness of the objective function. We show how to extend our framework to scale and demonstrate the efficiency and robustness of our methods on large scale semidefinite programs.
  arXiv  Detail & Related papers  (2023-06-30T08:34:29Z)
• Scalable Differentially Private Clustering via Hierarchically Separated Trees [82.69664595378869]
  We show that our method computes a solution with cost at most $O(d3/2log n)cdot OPT + O(k d2 log2 n / epsilon2)$, where $epsilon$ is the privacy guarantee. Although the worst-case guarantee is worse than that of state of the art private clustering methods, the algorithm we propose is practical.
  arXiv  Detail & Related papers  (2022-06-17T09:24:41Z)
• Gaussian Process Bandit Optimization with Few Batches [49.896920704012395]
  We introduce a batch algorithm inspired by finite-arm bandit algorithms. We show that it achieves the cumulative regret upper bound $Oast(sqrtTgamma_T)$ using $O(loglog T)$ batches within time horizon $T$. In addition, we propose a modified version of our algorithm, and characterize how the regret is impacted by the number of batches.
  arXiv  Detail & Related papers  (2021-10-15T00:54:04Z)
• Streaming Complexity of SVMs [110.63976030971106]
  We study the space complexity of solving the bias-regularized SVM problem in the streaming model. We show that for both problems, for dimensions of $frac1lambdaepsilon$, one can obtain streaming algorithms with spacely smaller than $frac1lambdaepsilon$.
  arXiv  Detail & Related papers  (2020-07-07T17:10:00Z)
• On the Almost Sure Convergence of Stochastic Gradient Descent in Non-Convex Problems [75.58134963501094]
  This paper analyzes the trajectories of gradient descent (SGD) We show that SGD avoids saddle points/manifolds with $1$ for strict step-size policies.
  arXiv  Detail & Related papers  (2020-06-19T14:11:26Z)
• Private Stochastic Convex Optimization: Optimal Rates in Linear Time [74.47681868973598]
  We study the problem of minimizing the population loss given i.i.d. samples from a distribution over convex loss functions. A recent work of Bassily et al. has established the optimal bound on the excess population loss achievable given $n$ samples. We describe two new techniques for deriving convex optimization algorithms both achieving the optimal bound on excess loss and using $O(minn, n2/d)$ gradient computations.
  arXiv  Detail & Related papers  (2020-05-10T19:52:03Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.