PyTorchKR Papers

최신 논문

총 1,287편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2501.04443

Revisiting LocalSGD and SCAFFOLD: Improved Rates and Missing Analysis

LocalSGD and SCAFFOLD are widely used methods in distributed stochastic optimization, with numerous applications in machine learning, large-scale data processing, and federated learning. However, rigorously establishing their theoretical advantages over simpler methods, such as minibatch SGD (MbSGD), has proven challenging, as existing analyses often rely on strong assumptions, unrealistic premises, or overly restrictive scenarios. In this work, we revisit the convergence properties of LocalSGD and SCAFFOLD under a variety of existing or weaker conditions, including gradient similarity, Hessian similarity, weak convexity, and Lipschitz continuity of the Hessian. Our analysis shows that (i) LocalSGD achieves faster convergence compared to MbSGD for weakly convex functions without requiring stronger gradient similarity assumptions; (ii) LocalSGD benefits significantly from higher-order similarity and smoothness; and (iii) SCAFFOLD demonstrates faster convergence than MbSGD for a broader class of non-quadratic functions. These theoretical insights provide a clearer understanding of the conditions under which LocalSGD and SCAFFOLD outperform MbSGD.

arXiv:2412.20017공식 코드 · Papers with Code

A Nearly Optimal Single Loop Algorithm for Stochastic Bilevel Optimization under Unbounded Smoothness

This paper studies the problem of stochastic bilevel optimization where the upper-level function is nonconvex with potentially unbounded smoothness and the lower-level function is strongly convex. This problem is motivated by meta-learning applied to sequential data, such as text classification using recurrent neural networks, where the smoothness constant of the upper-level loss function scales linearly with the gradient norm and can be potentially unbounded. Existing algorithm crucially relies on the nested loop design, which requires significant tuning efforts and is not practical. In this paper, we address this issue by proposing a Single Loop bIlevel oPtimizer (SLIP). The proposed algorithm first updates the lower-level variable by a few steps of stochastic gradient descent, and then simultaneously updates the upper-level variable by normalized stochastic gradient descent with momentum and the lower-level variable by stochastic gradient descent. Under standard assumptions, we show that our algorithm finds an $\epsilon$-stationary point within $\widetilde{O}(1/\epsilon^4)$\footnote{Here $\widetilde{O}(\cdot)$ compresses logarithmic factors of $1/\epsilon$ and $1/\delta$, where $\delta\in(0,1)$ denotes the failure probability.} oracle calls of stochastic gradient or Hessian-vector product, both in expectation and with high probability. This complexity result is nearly optimal up to logarithmic factors without mean-square smoothness of the stochastic gradient oracle. Our proof relies on (i) a refined characterization and control of the lower-level variable and (ii) establishing a novel connection between bilevel optimization and stochastic optimization under distributional drift. Our experiments on various tasks show that our algorithm significantly outperforms strong baselines in bilevel optimization.

arXiv:2412.19529

Nonconvex Stochastic Optimization under Heavy-Tailed Noises: Optimal Convergence without Gradient Clipping

Recently, the study of heavy-tailed noises in first-order nonconvex stochastic optimization has gotten a lot of attention since it was recognized as a more realistic condition as suggested by many empirical observations. Specifically, the stochastic noise (the difference between the stochastic and true gradient) is considered to have only a finite $\mathfrak{p}$-th moment where $\mathfrak{p}\in\left(1,2\right]$ instead of assuming it always satisfies the classical finite variance assumption. To deal with this more challenging setting, people have proposed different algorithms and proved them to converge at an optimal $\mathcal{O}(T^{\frac{1-\mathfrak{p}}{3\mathfrak{p}-2}})$ rate for smooth objectives after $T$ iterations. Notably, all these new-designed algorithms are based on the same technique - gradient clipping. Naturally, one may want to know whether the clipping method is a necessary ingredient and the only way to guarantee convergence under heavy-tailed noises. In this work, by revisiting the existing Batched Normalized Stochastic Gradient Descent with Momentum (Batched NSGDM) algorithm, we provide the first convergence result under heavy-tailed noises but without gradient clipping. Concretely, we prove that Batched NSGDM can achieve the optimal $\mathcal{O}(T^{\frac{1-\mathfrak{p}}{3\mathfrak{p}-2}})$ rate even under the relaxed smooth condition. More interestingly, we also establish the first $\mathcal{O}(T^{\frac{1-\mathfrak{p}}{2\mathfrak{p}}})$ convergence rate in the case where the tail index $\mathfrak{p}$ is unknown in advance, which is arguably the common scenario in practice.

arXiv:2412.14291

Projected gradient methods for nonconvex and stochastic optimization: new complexities and auto-conditioned stepsizes

We present a novel class of projected gradient (PG) methods for minimizing a smooth but not necessarily convex function over a convex compact set. We first provide a novel analysis of the "vanilla" PG method, achieving the best-known iteration complexity for finding an approximate stationary point of the problem. We then develop an "auto-conditioned" projected gradient (AC-PG) variant that achieves the same iteration complexity without requiring the input of the Lipschitz constant of the gradient or any line search procedure. The key idea is to estimate the Lipschitz constant using first-order information gathered from the previous iterations, and to show that the error caused by underestimating the Lipschitz constant can be properly controlled. We then generalize the PG methods to the stochastic setting, by proposing a stochastic projected gradient (SPG) method and a variance-reduced stochastic gradient (VR-SPG) method, achieving new complexity bounds in different oracle settings. We also present auto-conditioned stepsize policies for both stochastic PG methods and establish comparable convergence guarantees.

arXiv:2412.14488

A stochastic first-order method with multi-extrapolated momentum for highly smooth unconstrained optimization

In this paper, we consider an unconstrained stochastic optimization problem where the objective function exhibits high-order smoothness. Specifically, we propose a new stochastic first-order method (SFOM) with multi-extrapolated momentum, in which multiple extrapolations are performed in each iteration, followed by a momentum update based on these extrapolations. We demonstrate that the proposed SFOM can accelerate optimization by exploiting the high-order smoothness of the objective function $f$. Assuming that the $p$th-order derivative of $f$ is Lipschitz continuous for some $p\ge2$, and under additional mild assumptions, we establish that our method achieves a sample complexity of $\widetilde{\mathcal{O}}(\epsilon^{-(3p+1)/p})$ for finding a point $x$ such that $\mathbb{E}[\|\nabla f(x)\|]\le\epsilon$. To the best of our knowledge, this is the first SFOM to leverage arbitrary-order smoothness of the objective function for acceleration, resulting in a sample complexity that improves upon the best-known results without assuming the mean-squared smoothness condition. Preliminary numerical experiments validate the practical performance of our method and support our theoretical findings.

arXiv:2412.12049공식 코드 · Papers with Code

Bilevel Learning with Inexact Stochastic Gradients

Bilevel learning has gained prominence in machine learning, inverse problems, and imaging applications, including hyperparameter optimization, learning data-adaptive regularizers, and optimizing forward operators. The large-scale nature of these problems has led to the development of inexact and computationally efficient methods. Existing adaptive methods predominantly rely on deterministic formulations, while stochastic approaches often adopt a doubly-stochastic framework with impractical variance assumptions, enforces a fixed number of lower-level iterations, and requires extensive tuning. In this work, we focus on bilevel learning with strongly convex lower-level problems and a nonconvex sum-of-functions in the upper-level. Stochasticity arises from data sampling in the upper-level which leads to inexact stochastic hypergradients. We establish their connection to state-of-the-art stochastic optimization theory for nonconvex objectives. Furthermore, we prove the convergence of inexact stochastic bilevel optimization under mild assumptions. Our empirical results highlight significant speed-ups and improved generalization in imaging tasks such as image denoising and deblurring in comparison with adaptive deterministic bilevel methods.

arXiv:2412.10663

Memory-Efficient 4-bit Preconditioned Stochastic Optimization

Preconditioned stochastic optimization algorithms, exemplified by Shampoo, outperform first-order optimizers by offering theoretical convergence benefits and practical gains in large-scale neural network training. However, they incur substantial memory overhead due to the storage demands of non-diagonal preconditioning matrices. To address this, we introduce 4-bit quantization for Shampoo's preconditioners. We introduce two key methods: First, we apply Cholesky decomposition followed by quantization of the Cholesky factors, reducing memory usage by leveraging their lower triangular structure while better preserving spectral properties to minimize information loss. To our knowledge, this is the first quantization approach applied to Cholesky factors of preconditioners. Second, we incorporate error feedback in the quantization process, efficiently storing Cholesky factor and error state in the lower and upper triangular parts of the same matrix. Through extensive experiments, we demonstrate that combining Cholesky quantization with error feedback enhances memory efficiency and algorithm performance in large-scale deep-learning tasks. Theoretically, we also provide convergence proofs for quantized Shampoo under both smooth and non-smooth stochastic optimization settings.

arXiv:2412.07979공식 코드 · Papers with Code

AmCLR: Unified Augmented Learning for Cross-Modal Representations

Contrastive learning has emerged as a pivotal framework for representation learning, underpinning advances in both unimodal and bimodal applications like SimCLR and CLIP. To address fundamental limitations like large batch size dependency and bimodality, methods such as SogCLR leverage stochastic optimization for the global contrastive objective. Inspired by SogCLR's efficiency and adaptability, we introduce AmCLR and xAmCLR objective functions tailored for bimodal vision-language models to further enhance the robustness of contrastive learning. AmCLR integrates diverse augmentations, including text paraphrasing and image transformations, to reinforce the alignment of contrastive representations, keeping batch size limited to a few hundred samples unlike CLIP which needs batch size of 32,768 to produce reasonable results. xAmCLR further extends this paradigm by incorporating intra-modal alignments between original and augmented modalities for richer feature learning. These advancements yield a more resilient and generalizable contrastive learning process, aimed at overcoming bottlenecks in scaling and augmentative diversity. Since we have built our framework on the existing SogCLR, we are able to demonstrate improved representation quality with fewer computational resources, establishing a foundation for scalable and robust multi-modal learning.

arXiv:2412.06070

Stochastic Gradient Descent Revisited

Stochastic gradient descent (SGD) has been a go-to algorithm for nonconvex stochastic optimization problems arising in machine learning. Its theory however often requires a strong framework to guarantee convergence properties. We hereby present a full scope convergence study of biased nonconvex SGD, including weak convergence, function-value convergence and global convergence, and also provide subsequent convergence rates and complexities, all under relatively mild conditions in comparison with literature.

arXiv:2412.05343

Equivariant Denoisers for Image Restoration

One key ingredient of image restoration is to define a realistic prior on clean images to complete the missing information in the observation. State-of-the-art restoration methods rely on a neural network to encode this prior. Moreover, typical image distributions are invariant to some set of transformations, such as rotations or flips. However, most deep architectures are not designed to represent an invariant image distribution. Recent works have proposed to overcome this difficulty by including equivariance properties within a Plug-and-Play paradigm. In this work, we propose a unified framework named Equivariant Regularization by Denoising (ERED) based on equivariant denoisers and stochastic optimization. We analyze the convergence of this algorithm and discuss its practical benefit.

arXiv:2412.04177공식 코드 · Papers with Code

Fixed-Mean Gaussian Processes for Post-hoc Bayesian Deep Learning

Recently, there has been an increasing interest in performing post-hoc uncertainty estimation about the predictions of pre-trained deep neural networks (DNNs). Given a pre-trained DNN via back-propagation, these methods enhance the original network by adding output confidence measures, such as error bars, without compromising its initial accuracy. In this context, we introduce a novel family of sparse variational Gaussian processes (GPs), where the posterior mean is fixed to any continuous function when using a universal kernel. Specifically, we fix the mean of this GP to the output of the pre-trained DNN, allowing our approach to effectively fit the GP's predictive variances to estimate the DNN prediction uncertainty. Our approach leverages variational inference (VI) for efficient stochastic optimization, with training costs that remain independent of the number of training points, scaling efficiently to large datasets such as ImageNet. The proposed method, called fixed mean GP (FMGP), is architecture-agnostic, relying solely on the pre-trained model's outputs to adjust the predictive variances. Experimental results demonstrate that FMGP improves both uncertainty estimation and computational efficiency when compared to state-of-the-art methods.

arXiv:2412.02089

Offline Stochastic Optimization of Black-Box Objective Functions

Many challenges in science and engineering, such as drug discovery and communication network design, involve optimizing complex and expensive black-box functions across vast search spaces. Thus, it is essential to leverage existing data to avoid costly active queries of these black-box functions. To this end, while Offline Black-Box Optimization (BBO) is effective for deterministic problems, it may fall short in capturing the stochasticity of real-world scenarios. To address this, we introduce Stochastic Offline BBO (SOBBO), which tackles both black-box objectives and uncontrolled uncertainties. We propose two solutions: for large-data regimes, a differentiable surrogate allows for gradient-based optimization, while for scarce-data regimes, we directly estimate gradients under conservative field constraints, improving robustness, convergence, and data efficiency. Numerical experiments demonstrate the effectiveness of our approach on both synthetic and real-world tasks.

arXiv:2412.01050

Resilience-oriented Planning and Cost Allocation of Energy Storage Integrated with Soft Open Point Based on Resilience Insurance

In recent years, frequent extreme events have put forward higher requirements for improving the resilience of distribution networks (DNs). Introducing energy storage integrated with soft open point (E-SOP) is one of the effective ways to improve resilience. However, the widespread application of E-SOP is limited by its high investment cost. Based on this, we propose a cost allocation framework and optimal planning method of E-SOP in resilient DN. Firstly, a cost allocation mechanism for E-SOP based on resilience insurance service is designed; the probability of power users purchasing resilience insurance service is determined based on the expected utility theory. Then, a four-layer stochastic distributionally robust optimization (SDRO) model is developed for E-SOP planning and insurance pricing strategy, where the uncertainty in the intensity of contingent extreme events is addressed by a stochastic optimization approach, while the uncertainty in the occurrence of outages and resilience insurance purchases resulting from a specific extreme event is addressed via a distributionally robust optimization approach. Finally, The effectiveness of the proposed model is verified on the modified IEEE 33-bus DN.

arXiv:2412.02291공식 코드 · Papers with Code

Conformal Symplectic Optimization for Stable Reinforcement Learning

Training deep reinforcement learning (RL) agents necessitates overcoming the highly unstable nonconvex stochastic optimization inherent in the trial-and-error mechanism. To tackle this challenge, we propose a physics-inspired optimization algorithm called relativistic adaptive gradient descent (RAD), which enhances long-term training stability. By conceptualizing neural network (NN) training as the evolution of a conformal Hamiltonian system, we present a universal framework for transferring long-term stability from conformal symplectic integrators to iterative NN updating rules, where the choice of kinetic energy governs the dynamical properties of resulting optimization algorithms. By utilizing relativistic kinetic energy, RAD incorporates principles from special relativity and limits parameter updates below a finite speed, effectively mitigating abnormal gradient influences. Additionally, RAD models NN optimization as the evolution of a multi-particle system where each trainable parameter acts as an independent particle with an individual adaptive learning rate. We prove RAD's sublinear convergence under general nonconvex settings, where smaller gradient variance and larger batch sizes contribute to tighter convergence. Notably, RAD degrades to the well-known adaptive moment estimation (ADAM) algorithm when its speed coefficient is chosen as one and symplectic factor as a small positive value. Experimental results show RAD outperforming nine baseline optimizers with five RL algorithms across twelve environments, including standard benchmarks and challenging scenarios. Notably, RAD achieves up to a 155.1% performance improvement over ADAM in Atari games, showcasing its efficacy in stabilizing and accelerating RL training.

arXiv:2411.15151

Memory-Driven Metaheuristics: Improving Optimization Performance

Metaheuristics are stochastic optimization algorithms that mimic natural processes to find optimal solutions to complex problems. The success of metaheuristics largely depends on the ability to effectively explore and exploit the search space. Memory mechanisms have been introduced in several popular metaheuristic algorithms to enhance their performance. This chapter explores the significance of memory in metaheuristic algorithms and provides insights from well-known algorithms. The chapter begins by introducing the concept of memory, and its role in metaheuristic algorithms. The key factors influencing the effectiveness of memory mechanisms are discussed, such as the size of the memory, the information stored in memory, and the rate of information decay. A comprehensive analysis of how memory mechanisms are incorporated into popular metaheuristic algorithms is presented and concludes by highlighting the importance of memory in metaheuristic performance and providing future research directions for improving memory mechanisms. The key takeaways are that memory mechanisms can significantly enhance the performance of metaheuristics by enabling them to explore and exploit the search space effectively and efficiently, and that the choice of memory mechanism should be tailored to the problem domain and the characteristics of the search space.

arXiv:2411.13999

Accelerated zero-order SGD under high-order smoothness and overparameterized regime

We present a novel gradient-free algorithm to solve a convex stochastic optimization problem, such as those encountered in medicine, physics, and machine learning (e.g., adversarial multi-armed bandit problem), where the objective function can only be computed through numerical simulation, either as the result of a real experiment or as feedback given by the function evaluations from an adversary. Thus we suppose that only a black-box access to the function values of the objective is available, possibly corrupted by adversarial noise: deterministic or stochastic. The noisy setup can arise naturally from modeling randomness within a simulation or by computer discretization, or when exact values of function are forbidden due to privacy issues, or when solving non-convex problems as convex ones with an inexact function oracle. By exploiting higher-order smoothness, fulfilled, e.g., in logistic regression, we improve the performance of zero-order methods developed under the assumption of classical smoothness (or having a Lipschitz gradient). The proposed algorithm enjoys optimal oracle complexity and is designed under an overparameterization setup, i.e., when the number of model parameters is much larger than the size of the training dataset. Overparametrized models fit to the training data perfectly while also having good generalization and outperforming underparameterized models on unseen data. We provide convergence guarantees for the proposed algorithm under both types of noise. Moreover, we estimate the maximum permissible adversarial noise level that maintains the desired accuracy in the Euclidean setup, and then we extend our results to a non-Euclidean setup. Our theoretical results are verified on the logistic regression problem.

arXiv:2411.12898

Problem-dependent convergence bounds for randomized linear gradient compression

In distributed optimization, the communication of model updates can be a performance bottleneck. Consequently, gradient compression has been proposed as a means of increasing optimization throughput. In general, due to information loss, compression introduces a penalty on the number of iterations needed to reach a solution. In this work, we investigate how the iteration penalty depends on the interaction between compression and problem structure, in the context of non-convex stochastic optimization. We focus on linear compression schemes, where compression and decompression can be modeled as multiplication with a random matrix. We consider several distributions of matrices, among them random orthogonal matrices and matrices with random Gaussian entries. We find that in each case, the impact of compression on convergence can be quantified in terms of the norm of the Hessian of the objective, using a norm defined by the compression scheme. The analysis reveals that in certain cases, compression performance is related to low-rank structure or other spectral properties of the problem. In these cases, our bounds predict that the penalty introduced by compression is significantly reduced compared to worst-case bounds that only consider the compression level, ignoring problem data. We verify the theoretical findings on several optimization problems, including fine-tuning an image classification model.

arXiv:2411.10784

On Reductions and Representations of Learning Problems in Euclidean Spaces

Many practical prediction algorithms represent inputs in Euclidean space and replace the discrete 0/1 classification loss with a real-valued surrogate loss, effectively reducing classification tasks to stochastic optimization. In this paper, we investigate the expressivity of such reductions in terms of key resources, including dimension and the role of randomness. We establish bounds on the minimum Euclidean dimension $D$ needed to reduce a concept class with VC dimension $d$ to a Stochastic Convex Optimization (SCO) problem in $\mathbb{R}^D$, formally addressing the intuitive interpretation of the VC dimension as the number of parameters needed to learn the class. To achieve this, we develop a generalization of the Borsuk-Ulam Theorem that combines the classical topological approach with convexity considerations. Perhaps surprisingly, we show that, in some cases, the number of parameters $D$ must be exponentially larger than the VC dimension $d$, even if the reduction is only slightly non-trivial. We also present natural classification tasks that can be represented in much smaller dimensions by leveraging randomness, as seen in techniques like random initialization. This result resolves an open question posed by Kamath, Montasser, and Srebro (COLT 2020). Our findings introduce new variants of \emph{dimension complexity} (also known as \emph{sign-rank}), a well-studied parameter in learning and complexity theory. Specifically, we define an approximate version of sign-rank and another variant that captures the minimum dimension required for a reduction to SCO. We also propose several open questions and directions for future research.

arXiv:2411.10438공식 코드 · Papers with Code

MARS: Unleashing the Power of Variance Reduction for Training Large Models

Training deep neural networks--and more recently, large models demands efficient and scalable optimizers. Adaptive gradient algorithms like Adam, AdamW, and their variants have been central to this task. Despite the development of numerous variance reduction algorithms in the past decade aimed at accelerating stochastic optimization in both convex and nonconvex settings, variance reduction has not found widespread success in training deep neural networks or large language models. Consequently, it has remained a less favored approach in modern AI. In this paper, to unleash the power of variance reduction for efficient training of large models, we propose a unified optimization framework, MARS (Make vAriance Reduction Shine), which reconciles preconditioned gradient methods with variance reduction via a scaled stochastic recursive momentum technique. Within our framework, we introduce three instances of MARS that leverage preconditioned gradient updates based on AdamW, Lion, and Shampoo, respectively. We also draw a connection between our algorithms and existing optimizers. Experimental results on training GPT-2 models indicate that MARS consistently outperforms AdamW by a large margin. The implementation of MARS is available at https://github.com/AGI-Arena/MARS.

arXiv:2411.09127

Complexity-Aware Training of Deep Neural Networks for Optimal Structure Discovery

We propose a novel algorithm for combined unit/filter and layer pruning of deep neural networks that functions during training and without requiring a pre-trained network to apply. Our algorithm optimally trades-off learning accuracy and pruning levels while balancing layer vs. unit/filter pruning and computational vs. parameter complexity using only three user-defined parameters, which are easy to interpret and tune. The optimal network structure is found as the solution of a stochastic optimization problem over the network weights and the parameters of variational Bernoulli distributions for 0/1 Random Variables scaling the units and layers of the network. Pruning occurs when a variational parameter converges to 0 rendering the corresponding structure permanently inactive, thus saving computations during training and prediction. A key contribution of our approach is to define a cost function that combines the objectives of prediction accuracy and network pruning in a computational/parameter complexity-aware manner and the automatic selection of the many regularization parameters. We show that the solutions of the optimization problem to which the algorithm converges are deterministic networks. We analyze the ODE system that underlies our stochastic optimization algorithm and establish domains of attraction around zero for the dynamics of the network parameters. These results provide theoretical support for safely pruning units/filters and/or layers during training and lead to practical pruning conditions. We evaluate our method on the CIFAR-10/100 and ImageNet datasets using ResNet architectures and demonstrate that our method improves upon layer only or unit only pruning and favorably competes with combined unit/filter and layer pruning algorithms requiring pre-trained networks with respect to pruning ratios and test accuracy.