Fine-tuning language models (LMs) has demonstrated success in a wide array of downstream tasks. However, as LMs are scaled up, the memory requirements for backpropagation become prohibitively high. Zeroth-order (ZO) optimization methods can leverage memory-efficient forward passes to estimate gradients. More recently, MeZO, an adaptation of ZO-SGD, has been shown to consistently outperform zero-shot and in-context learning when combined with suitable task prompts. In this work, we couple ZO methods with variance reduction techniques to enhance stability and convergence for inference-based LM fine-tuning. We introduce Memory-Efficient Zeroth-Order Stochastic Variance-Reduced Gradient (MeZO-SVRG) and demonstrate its efficacy across multiple LM fine-tuning tasks, eliminating the reliance on task-specific prompts. Evaluated across a range of both masked and autoregressive LMs on benchmark GLUE tasks, MeZO-SVRG outperforms MeZO with up to 20% increase in test accuracies in both full- and partial-parameter fine-tuning settings. MeZO-SVRG benefits from reduced computation time as it often surpasses MeZO's peak test accuracy with a $2\times$ reduction in GPU-hours. MeZO-SVRG significantly reduces the required memory footprint compared to first-order SGD, i.e. by $2\times$ for autoregressive models. Our experiments highlight that MeZO-SVRG's memory savings progressively improve compared to SGD with larger batch sizes.
Zeroth-order (derivative-free) optimization attracts a lot of attention in machine learning, because explicit gradient calculations may be computationally expensive or infeasible. To handle large scale problems both in volume and dimension, recently asynchronous doubly stochastic zeroth-order algorithms were proposed. The convergence rate of existing asynchronous doubly stochastic zeroth order algorithms is $O(\frac{1}{\sqrt{T}})$ (also for the sequential stochastic zeroth-order optimization algorithms). In this paper, we focus on the finite sums of smooth but not necessarily convex functions,…
A major challenge of applying zeroth-order (ZO) methods is the high query complexity, especially when queries are costly. We propose a novel gradient estimation technique for ZO methods based on adaptive lazy queries that we term as LAZO. Different from the classic one-point or two-point gradient estimation methods, LAZO develops two alternative ways to check the usefulness of old queries from previous iterations, and then adaptively reuses them to construct the low-variance gradient estimates. We rigorously establish that through judiciously reusing the old queries, LAZO can reduce the varian…
The University of Sydney（悉尼大学）；University of California, Davis（加州大学戴维斯分校）
Derivative-free optimization has become an important technique used in machine learning for optimizing black-box models. To conduct updates without explicitly computing gradient, most current approaches iteratively sample a random search direction from Gaussian distribution and compute the estimated gradient along that direction. However, due to the variance in the search direction, the convergence rates and query complexities of existing methods suffer from a factor of $d$, where $d$ is the problem dimension. In this paper, we introduce a novel Stochastic Zeroth-order method with Variance Red…
The Ohio State University（俄亥俄州立大学）；Duke University（杜克大学）
Two types of zeroth-order stochastic algorithms have recently been designed for nonconvex optimization respectively based on the first-order techniques SVRG and SARAH/SPIDER. This paper addresses several important issues that are still open in these methods. First, all existing SVRG-type zeroth-order algorithms suffer from worse function query complexities than either zeroth-order gradient descent (ZO-GD) or stochastic gradient descent (ZO-SGD). In this paper, we propose a new algorithm ZO-SVRG-Coord-Rand and develop a new analysis for an existing ZO-SVRG-Coord algorithm proposed in Liu et al.…
MIT-IBM Watson AI Lab（麻省理工学院-IBM Watson人工智能实验室）；IBM Research（IBM研究）；Lawrence Livermore National Laboratory（劳伦斯利弗莫尔国家实验室）；University of Michigan, Ann Arbor（密歇根大学安娜堡分校）
As application demands for zeroth-order (gradient-free) optimization accelerate, the need for variance reduced and faster converging approaches is also intensifying. This paper addresses these challenges by presenting: a) a comprehensive theoretical analysis of variance reduced zeroth-order (ZO) optimization, b) a novel variance reduced ZO algorithm, called ZO-SVRG, and c) an experimental evaluation of our approach in the context of two compelling applications, black-box chemical material classification and generation of adversarial examples from black-box deep neural network models. Our theor…
Differentially Private Zeroth-Order Methods for Scalable Large Language Model Finetuning
Zhejiang University（浙江大学）；Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security（杭州高新区（滨江）区块链与数据安全研究院）；University of Chicago（芝加哥大学）
Zeroth-Order Methods for Nonconvex Stochastic Problems with Decision-Dependent Distributions
Graduate School of Information Science and Technology, University of Tokyo（东京大学信息科学与技术研究生院）；Center for Advanced Intelligence Project, RIKEN（理化学研究所先进智能项目中心）
ZorBA: Zeroth-order Federated Fine-tuning of LLMs with Heterogeneous Block Activation
Department of Electrical and Computer Engineering, The University of British Columbia, Vancouver, Canada（电气与计算机工程系，不列颠哥伦比亚大学）；Department of Computer Science and Engineering, Southern University of Science and Technology, China（计算机科学与工程系，南方科技大学）
Zeroth-Order Fine-Tuning of LLMs in Random Subspaces
Beijing Normal University（北京师范大学）；Singapore Management University（新加坡国立大学）；TAL Education Group（TAL教育集团）；Engineering Research Center of Intelligent Technology and Educational Application (MOE)（智能技术与教育应用工程研究中心（教育部））
Hi-ZFO: Hierarchical Zeroth- and First-Order LLM Fine-Tuning via Importance-Guided Tensor Selection
School of Intelligence Science and Technology, Peking University（智能科学与技术学院，北京大学）；State Key Laboratory of General Artificial Intelligence（通用人工智能国家重点实验室）
Double Variance Reduction: A Smoothing Trick for Composite Optimization Problems without First-Order Gradient
School of Management, Xi’an Jiaotong University（西安交通大学管理学院）；State Grid Corporation of China（中国国家电网有限公司）；International Business School, Beijing Foreign Studies University（北京外国语大学国际商学院）；Agency for Science, Technology and Research (A*STAR)（新加坡科技研究局）；College of Computing and Data Science, NTU（南洋理工大学计算机与数据科学学院）
Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark
Michigan State University（密歇根州立大学）；The University of North Carolina at Chapel Hill（北卡罗来纳大学教堂山分校）；UT Austin（德克萨斯大学奥斯汀分校）；University of Minnesota Twin Cities（明尼苏达大学双城分校）；IBM Research（IBM研究院）；Princeton University（普林斯顿大学）；DAMO Academy, Alibaba Group（阿里巴巴达摩院）；MIT（麻省理工学院）；Harvard University（哈佛大学）
The Ohio State University（俄亥俄州立大学）；Princeton University（普林斯顿大学）
Many important machine learning applications amount to solving minimax optimization problems, and in many cases there is no access to the gradient information, but only the function values. In this paper, we focus on such a gradient-free setting, and consider the nonconvex-strongly-concave minimax stochastic optimization problem. In the literature, various zeroth-order (i.e., gradient-free) minimax methods have been proposed, but none of them achieve the potentially feasible computational complexity of $\mathcal{O}(ε^{-3})$ suggested by the stochastic nonconvex minimization theorem. In this pa…
TeZO: Empowering the Low-Rankness on the Temporal Dimension in the Zeroth-Order Optimization for Fine-tuning LLMs
Faculty of Engineering, The University of Sydney（悉尼大学工程学院）；School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University（中山大学深圳校区网络与空间安全学院）
In this paper, we propose and analyze algorithms for zeroth-order optimization of non-convex composite objectives, focusing on reducing the complexity dependence on dimensionality. This is achieved by exploiting the low dimensional structure of the decision set using the stochastic mirror descent method with an entropy alike function, which performs gradient descent in the space equipped with the maximum norm. To improve the gradient estimation, we replace the classic Gaussian smoothing method with a sampling method based on the Rademacher distribution and show that the mini-batch method copes…
Criteo AI Lab；Univ. Grenoble Alpes（格勒诺布尔阿尔卑斯大学）；CNRS（法国国家科学研究中心）；Inria（法国国家信息与自动化研究所）；Grenoble INP（格勒诺布尔理工学院）
We propose a hierarchical version of dual averaging for zeroth-order online non-convex optimization - i.e., learning processes where, at each stage, the optimizer is facing an unknown non-convex loss function and only receives the incurred loss as feedback. The proposed class of policies relies on the construction of an online model that aggregates loss information as it arrives, and it consists of two principal components: (a) a regularizer adapted to the Fisher information metric (as opposed to the metric norm of the ambient space); and (b) a principled exploration of the problem's state spa…
Zeroth-Order Alternating Gradient Descent Ascent Algorithms for a Class of Nonconvex-Nonconcave Minimax Problems
Shanghai University（上海大学）；Newtouch Center for Mathematics of Shanghai University（上海大学纽崔克数学中心）；Academy of Mathematics and Systems Science, Chinese Academy of Sciences（中国科学院数学与系统科学研究院）
University of California, Davis（加州大学戴维斯分校）；University of Southern California（南加州大学）
In this paper, we study zeroth-order algorithms for minimax optimization problems that are nonconvex in one variable and strongly-concave in the other variable. Such minimax optimization problems have attracted significant attention lately due to their applications in modern machine learning tasks. We first consider a deterministic version of the problem. We design and analyze the Zeroth-Order Gradient Descent Ascent (\texttt{ZO-GDA}) algorithm, and provide improved results compared to existing works, in terms of oracle complexity. We also propose the Zeroth-Order Gradient Descent Multi-Step A…
Stochastic Zeroth-Order Optimization under Strongly Convexity and Lipschitz Hessian: Minimax Sample Complexity
University of California, Santa Barbara（加利福尼亚大学圣巴巴拉分校）；University of Texas at Dallas（德克萨斯大学达拉斯分校）；University of California, Berkeley（加利福尼亚大学伯克利分校）；New York University（纽约大学）；Princeton University（普林斯顿大学）
In the learning to learn (L2L) framework, we cast the design of optimization algorithms as a machine learning problem and use deep neural networks to learn the update rules. In this paper, we extend the L2L framework to zeroth-order (ZO) optimization setting, where no explicit gradient information is available. Our learned optimizer, modeled as recurrent neural network (RNN), first approximates gradient by ZO gradient estimator and then produces parameter update utilizing the knowledge of previous iterations. To reduce high variance effect due to ZO gradient estimator, we further introduce ano…
Nanjing University of Aeronautics and Astronautics（南京航空航天大学）；University of Pittsburgh（匹兹堡大学）
Proximal gradient method has been playing an important role to solve many machine learning tasks, especially for the nonsmooth problems. However, in some machine learning problems such as the bandit model and the black-box learning problem, proximal gradient method could fail because the explicit gradients of these problems are difficult or infeasible to obtain. The gradient-free (zeroth-order) method can address these problems because only the objective function values are required in the optimization. Recently, the first zeroth-order proximal stochastic algorithm was proposed to solve the no…
Syracuse University（雪城大学）；Northeastern University（东北大学）；MIT-IBM Watson AI Lab（MIT-IBM沃森人工智能实验室）；IBM Research（IBM研究院）
In this work, we focus on the study of stochastic zeroth-order (ZO) optimization which does not require first-order gradient information and uses only function evaluations. The problem of ZO optimization has emerged in many recent machine learning applications, where the gradient of the objective function is either unavailable or difficult to compute. In such cases, we can approximate the full gradients or stochastic gradients through function value based gradient estimates. Here, we propose a novel hybrid gradient estimator (HGE), which takes advantage of the query-efficiency of random gradie…
Obtaining Lower Query Complexities through Lightweight Zeroth-Order Proximal Gradient Algorithms
School of Artificial Intelligence, Jilin University（吉林大学人工智能学院）；Mohamed bin Zayed University of Artificial Intelligence（穆罕默德·本·扎耶德人工智能大学）；Texas A&M University（得克萨斯农工大学）；University Maryland College Park（马里兰大学学院公园分校）
Graph Convolutional Networks (GCNs) have achieved impressive empirical advancement across a wide variety of semi-supervised node classification tasks. Despite their great success, training GCNs on large graphs suffers from computational and memory issues. A potential path to circumvent these obstacles is sampling-based methods, where at each layer a subset of nodes is sampled. Although recent studies have empirically demonstrated the effectiveness of sampling-based methods, these works lack theoretical convergence guarantees under realistic settings and cannot fully leverage the information of…
Nanjing University of Information Science & Technology（南京信息工程大学）；MBZUAI（穆罕默德·本·扎耶德人工智能大学）；JD Finance America Corporation（京东金融美国公司）；University of Pittsburgh（匹兹堡大学）
The conditional gradient algorithm (also known as the Frank-Wolfe algorithm) has recently regained popularity in the machine learning community due to its projection-free property to solve constrained problems. Although many variants of the conditional gradient algorithm have been proposed to improve performance, they depend on first-order information (gradient) to optimize. Naturally, these algorithms are unable to function properly in the field of increasingly popular zeroth-order optimization, where only zeroth-order information (function value) is available. To fill in this gap, we propose…
Magnitude pruning is a widely used strategy for reducing model size in pure supervised learning; however, it is less effective in the transfer learning regime that has become standard for state-of-the-art natural language processing applications. We propose the use of movement pruning, a simple, deterministic first-order weight pruning method that is more adaptive to pretrained model fine-tuning. We give mathematical foundations to the method and compare it to existing zeroth- and first-order pruning methods. Experiments show that when pruning large pretrained language models, movement pruning…
We study convergence rates of AdaGrad-Norm as an exemplar of adaptive stochastic gradient methods (SGD), where the step sizes change based on observed stochastic gradients, for minimizing non-convex, smooth objectives. Despite their popularity, the analysis of adaptive SGD lags behind that of non adaptive methods in this setting. Specifically, all prior works rely on some subset of the following assumptions: (i) uniformly-bounded gradient norms, (ii) uniformly-bounded stochastic gradient variance (or even noise support), (iii) conditional independence between the step size and stochastic gradi…
The Multi-Query Paradox in Zeroth-Order Optimization
Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong（香港中文大学计算机科学与工程系）；The Chinese University of Hong Kong, Hong Kong（香港中文大学）；Xiamen University, China（厦门大学）
