ICML 2024 Papers

In this study, we consider the realm of covariance matrices in machine learning, particularly focusing on computing Fréchet means on the manifold of symmetric positive definite matrices, commonly referred to as Karcher or geometric means. Such means are leveraged in numerous machine learning tasks. Relying on advanced statistical tools, we introduce a random matrix theory based method that estimates Fréchet means, which is particularly beneficial when dealing with low sample support and a high number of matrices to average. Our experimental evaluation, involving both synthetic and real-world EEG and hyperspectral datasets, shows that we largely outperform state-of-the-art methods.

In this paper we theoretically show that interior-point methods based on self-concordant barriers possess favorable global complexity beyond their standard application area of convex optimization. To do that we propose first- and second-order methods for non-convex optimization problems with general convex set constraints and linear constraints. Our methods attain a suitably defined class of approximate first- or second-order KKT points with the worst-case iteration complexity similar to unconstrained problems, namely $O(\varepsilon^{-2})$ (first-order) and $O(\varepsilon^{-3/2})$ (second-order), respectively.

Dynamic Sparse Training (DST) is an effective approach for addressing the substantial training resource requirements posed by the ever-increasing size of the Deep Neural Networks (DNNs). Characterized by its dynamic "train-prune-grow'' schedule during training, DST implicitly develops a bi-level structure for training the weights while discovering a subnetwork topology. However, such a structure is consistently overlooked by the current DST algorithms for further optimization opportunities, and these algorithms, on the other hand, solely optimize the weights while determining masks heuristically. In this paper, we extensively study DST algorithms and argue that the training scheme of DST naturally forms a bi-level problem in which the updating of weight and mask is interdependent. Based on this observation, we introduce a novel efficient training framework called BiDST, which for the first time, introduces bi-level optimization methodology into dynamic sparse training domain. Unlike traditional partial-heuristic DST schemes, which suffer from sub-optimal search efficiency for masks and miss the opportunity to fully explore the topological space of neural networks, BiDST excels at discovering excellent sparse patterns by optimizing mask and weight simultaneously, resulting in maximum 2.62% higher accuracy, 2.1$\times$ faster execution speed, and 25$\times$ reduced overhead. Code available at https://github.com/jjsrf/BiDST-ICML2024.

Bayesian optimization is a powerful framework for optimizing functions that are expensive or time-consuming to evaluate. Recent work has considered Bayesian optimization of function networks (BOFN), where the objective function is given by a network of functions, each taking as input the output of previous nodes in the network as well as additional parameters. Leveraging this network structure has been shown to yield significant performance improvements. Existing BOFN algorithms for general-purpose networks evaluate the full network at each iteration. However, many real-world applications allow for evaluating nodes individually. To exploit this, we propose a novel knowledge gradient acquisition function that chooses which node and corresponding inputs to evaluate in a cost-aware manner, thereby reducing query costs by evaluating only on a part of the network at each step. We provide an efficient approach to optimizing our acquisition function and show that it outperforms existing BOFN methods and other benchmarks across several synthetic and real-world problems. Our acquisition function is the first to enable cost-aware optimization of a broad class of function networks.

The Gradient Descent-Ascent (GDA) algorithm, designed to solve minimax optimization problems, takes the descent and ascent steps either simultaneously (Sim-GDA) or alternately (Alt-GDA). While Alt-GDA is commonly observed to converge faster, the performance gap between the two is not yet well understood theoretically, especially in terms of global convergence rates. To address this theory-practice gap, we present fine-grained convergence analyses of both algorithms for strongly-convex-strongly-concave and Lipschitz-gradient objectives. Our new iteration complexity upper bound of Alt-GDA is strictly smaller than the lower bound of Sim-GDA; i.e., Alt-GDA is provably faster. Moreover, we propose Alternating-Extrapolation GDA (Alex-GDA), a general algorithmic framework that subsumes Sim-GDA and Alt-GDA, for which the main idea is to alternately take gradients from extrapolations of the iterates. We show that Alex-GDA satisfies a smaller iteration complexity bound, identical to that of the Extra-gradient method, while requiring less gradient computations. We also prove that Alex-GDA enjoys linear convergence for bilinear problems, for which both Sim-GDA and Alt-GDA fail to converge at all.

In federated learning (FL), data heterogeneity is the main reason that existing theoretical analyses are pessimistic about the convergence rate. In particular, for many FL algorithms, the convergence rate grows dramatically when the number of local updates becomes large, especially when the product of the gradient divergence and local Lipschitz constant is large. However, empirical studies can show that more local updates can improve the convergence rate even when these two parameters are large, which is inconsistent with the theoretical findings. This paper aims to bridge this gap between theoretical understanding and practical performance by providing a theoretical analysis from a new perspective on data heterogeneity. In particular, we propose a new and weaker assumption compared to the local Lipschitz gradient assumption, named the heterogeneity-driven pseudo-Lipschitz assumption. We show that this and the gradient divergence assumptions can jointly characterize the effect of data heterogeneity. By deriving a convergence upper bound for FedAvg and its extensions, we show that, compared to the existing works, local Lipschitz constant is replaced by the much smaller heterogeneity-driven pseudo-Lipschitz constant and the corresponding convergence upper bound can be significantly reduced for the same number of local updates, although its order stays the same. In addition, when the local objective function is quadratic, more insights on the impact of data heterogeneity can be obtained using the heterogeneity-driven pseudo-Lipschitz constant. For example, we can identify a region where FedAvg can outperform mini-batch SGD even when the gradient divergence can be arbitrarily large. Our findings are validated using experiments.

Foundation Models (FMs) have demonstrated remarkable insights into the relational dynamics of the world, leading to the crucial question: *how do these models acquire an understanding of world hybrid relations?* Traditional statistical learning, particularly for prediction problems, may overlook the rich and inherently structured information from the data, especially regarding the relationships between objects. We introduce a mathematical model that formalizes relational learning as hypergraph recovery to study pre-training of FMs. In our framework, the world is represented as a hypergraph, with data abstracted as random samples from hyperedges. We theoretically examine the feasibility of a Pre-Trained Model (PTM) to recover this hypergraph and analyze the data efficiency in a minimax near-optimal style. By integrating rich graph theories into the realm of PTMs, our mathematical framework offers powerful tools for an in-depth understanding of pre-training from a unique perspective and can be used under various scenarios. As an example, we extend the framework to entity alignment in multimodal learning.

This paper considers stochastic weakly convex optimization without the standard Lipschitz continuity assumption. Based on new adaptive regularization (stepsize) strategies, we show that a wide class of stochastic algorithms, including the stochastic subgradient method, preserve the $\mathcal{O} ( 1 / \sqrt{K})$ convergence rate with constant failure rate. Our analyses rest on rather weak assumptions: the Lipschitz parameter can be either bounded by a general growth function of $\\|x\\|$ or locally estimated through independent random samples. Numerical experiments demonstrate the efficiency and robustness of our proposed stepsize policies.

Developing machine learning models that account for potential faults encountered in real-world environments presents a fundamental challenge for mission-critical applications. In this paper, we introduce a novel theoretical framework grounded in learning theory for dealing with faults. In particular, we propose a framework called *fault-tolerant PAC learning*, aimed at identifying the most fault-tolerant models from a given hypothesis class (such as neural networks). We show that if faults occur randomly, fault-tolerant learning is equivalent to regular PAC learning. However, for *adversarial* faults, we show that the sample complexity of fault-tolerant PAC learning can grow linearly w.r.t. the number of perturbing functions induced by the faults, even for a hypothesis class with VC-dimension 1. We then provide a matching upper bound by restricting the number of perturbing functions. Finally, we show that the linear dependency on the number of perturbing functions can be substantially improved for *deletion faults* in neural networks. Our work provides a powerful formal framework and avenues for a number of future investigations on the precise characterization of fault-tolerant learning.

This paper considers the optimization problem of the form $\min_{{\bf x}\in{\mathbb R}^d} f({\bf x})\triangleq \frac{1}{n}\sum_{i=1}^n f_i({\bf x})$, where $f(\cdot)$ satisfies the Polyak–Łojasiewicz (PL) condition with parameter $\mu$ and $\{f_i(\cdot)\}_{i=1}^n$ is $L$-mean-squared smooth. We show that any gradient method requires at least $\Omega(n+\kappa\sqrt{n}\log(1/\epsilon))$ incremental first-order oracle (IFO) calls to find an $\epsilon$-suboptimal solution, where $\kappa\triangleq L/\mu$ is the condition number of the problem. This result nearly matches upper bounds of IFO complexity for best-known first-order methods. We also study the problem of minimizing the PL function in the distributed setting such that the individuals $f_1(\cdot),\dots,f_n(\cdot)$ are located on a connected network of $n$ agents. We provide lower bounds of $\Omega(\kappa/\sqrt{\gamma}\log(1/\epsilon))$, $\Omega((\kappa+\tau\kappa/\sqrt{\gamma})\log(1/\epsilon))$ and $\Omega\big(n+\kappa\sqrt{n}\log(1/\epsilon)\big)$ for communication rounds, time cost and local first-order oracle calls respectively, where $\gamma\in(0,1]$ is the spectral gap of the mixing matrix associated with the network and $\tau>0$ is the time cost of per communication round. Furthermore, we propose a decentralized first-order method that nearly matches above lower bounds in expectation.

Recently, Cutkosky et al. introduce the online-to-non-convex framework, which utilizes online learning methods to solve non-smooth non-convex optimization problems, and achieves an $\mathcal{O}(\epsilon^{-3}\delta^{-1})$ gradient complexity for finding $(\delta,\epsilon)$-stationary points. However, their results rely on the bounded variance assumption of stochastic gradients and only hold in expectation. To address these limitations, we investigate the case that stochastic gradients obey heavy-tailed distributions with finite $\mathfrak{p}$-th moments for some $\mathfrak{p}\in(1,2]$, and propose a novel algorithm which is able to identify a $(\delta,\epsilon)$-stationary point with high probability, after consuming $\tilde{\mathcal{O}}(\epsilon^{-\frac{2\mathfrak{p}-1}{\mathfrak{p}-1}}\delta^{-1})$ stochastic gradients. The key idea is first incorporating the gradient clipping technique into the online-to-non-convex framework to produce a sequence of points, the averaged gradient norms of which is no greater than $\epsilon$. Then, we propose a validation method to select one $(\delta,\epsilon)$-stationary point among the candidates. When gradient distributions have bounded variance, i.e., $\mathfrak{p}=2$, our result turns into $\tilde{\mathcal{O}}(\epsilon^{-3}\delta^{-1})$, which improves the existing $\tilde{\mathcal{O}}(\epsilon^{-4}\delta^{-1})$ high-probability bound. When the objective is smooth, our algorithm can also find an $\epsilon$-stationary point with $\tilde{\mathcal{O}}(\epsilon^{-\frac{3\mathfrak{p}-2}{\mathfrak{p}-1}})$ gradient queries.

While supervised federated learning approaches have enjoyed significant success, the domain of unsupervised federated learning remains relatively underexplored. Several federated EM algorithms have gained popularity in practice, however, their theoretical foundations are often lacking. In this paper, we first introduce a federated gradient EM algorithm (FedGrEM) designed for the unsupervised learning of mixture models, which supplements the existing federated EM algorithms by considering task heterogeneity and potential adversarial attacks. We present a comprehensive finite-sample theory that holds for general mixture models, then apply this general theory on specific statistical models to characterize the explicit estimation error of model parameters and mixture proportions. Our theory elucidates when and how FedGrEM outperforms local single-task learning with insights extending to existing federated EM algorithms. This bridges the gap between their practical success and theoretical understanding. Our numerical results validate our theory, and demonstrate FedGrEM's superiority over existing unsupervised federated learning benchmarks.

This paper rethinks Sharpness-Aware Minimization (SAM), which is originally formulated as a zero-sum game where the weights of a network and a bounded perturbation try to minimize/maximize, respectively, the same differentiable loss. To fundamentally improve this design, we argue that SAM should instead be reformulated using the 0-1 loss. As a continuous relaxation, we follow the simple conventional approach where the minimizing (maximizing) player uses an upper bound (lower bound) surrogate to the 0-1 loss. This leads to a novel formulation of SAM as a bilevel optimization problem, dubbed as BiSAM. BiSAM with newly designed lower-bound surrogate loss indeed constructs stronger perturbation. Through numerical evidence, we show that BiSAM consistently results in improved performance when compared to the original SAM and variants, while enjoying similar computational complexity. Our code is available at https://github.com/LIONS-EPFL/BiSAM.

When predictions are performative, the choice of which predictor to deploy influences the distribution of future observations. The overarching goal in learning under performativity is to find a predictor that has low performative risk, that is, good performance on its induced distribution. One family of solutions for optimizing the performative risk, including bandits and other derivative-free methods, is agnostic to any structure in the performative feedback, leading to exceedingly slow convergence rates. A complementary family of solutions makes use of explicit models for the feedback, such as best-response models in strategic classification, enabling faster rates. However, these rates critically rely on the feedback model being correct. In this work we study a general protocol for making use of possibly misspecified models in performative prediction, called plug-in performative optimization. We show this solution can be far superior to model-agnostic strategies, as long as the misspecification is not too extreme. Our results support the hypothesis that models, even if misspecified, can indeed help with learning in performative settings.

Real-valued functions on geometric data---such as node attributes on a graph---can be optimized using descriptors from persistent homology, allowing the user to incorporate topological terms in the loss function. When optimizing a single real-valued function (the one-parameter setting), there is a canonical choice of descriptor for persistent homology: the barcode. The operation mapping a real-valued function to its barcode is differentiable almost everywhere, and the convergence of gradient descent for losses using barcodes is relatively well understood. When optimizing a vector-valued function (the multiparameter setting), there is no unique choice of descriptor for multiparameter persistent homology, and many distinct descriptors have been proposed. This calls for the development of a general framework for differentiability and optimization that applies to a wide range of multiparameter homological descriptors. In this article, we develop such a framework and show that it encompasses well-known descriptors of different flavors, such as signed barcodes and the multiparameter persistence landscape. We complement the theory with numerical experiments supporting the idea that optimizing multiparameter homological descriptors can lead to improved performances compared to optimizing one-parameter descriptors, even when using the simplest and most efficiently computable multiparameter descriptors.

Kahneman & Tversky's $\textit{prospect theory}$ tells us that humans perceive random variables in a biased but well-defined manner (1992); for example, humans are famously loss-averse. We show that objectives for aligning LLMs with human feedback implicitly incorporate many of these biases---the success of these objectives (e.g., DPO) over cross-entropy minimization can partly be ascribed to them belonging to a family of loss functions that we call $\textit{human-aware losses}$ (HALOs). However, the utility functions these methods attribute to humans still differ from those in the prospect theory literature. Using a Kahneman-Tversky model of human utility, we propose a HALO that directly maximizes the utility of generations instead of maximizing the log-likelihood of preferences, as current methods do. We call this approach KTO, and it matches or exceeds the performance of preference-based methods at scales from 1B to 30B, despite only learning from a binary signal of whether an output is desirable. More broadly, our work suggests that there is no one HALO that is universally superior; the best loss depends on the inductive biases most appropriate for a given setting, an oft-overlooked consideration.

Offline preference optimization allows fine-tuning large models directly from offline data, and has proved effective in recent alignment practices. We propose generalized preference optimization (GPO), a family of offline losses parameterized by a general class of convex functions. GPO enables a unified view over preference optimization, encompassing existing algorithms such as DPO, IPO and SLiC as special cases, while naturally introducing new variants. The GPO framework also sheds light on how offline algorithms enforce regularization, through the design of the convex function that defines the loss. Our analysis and experiments reveal the connections and subtle differences between the offline regularization and the KL divergence regularization intended by the canonical RLHF formulation. In a controlled setting akin to Gao et al 2023, we also show that different GPO variants achieve similar trade-offs between regularization and performance, though the optimal values of hyper-parameter might differ as predicted by theory. In all, our results present new algorithmic toolkits and empirical insights to alignment practitioners.

Bayesian optimization has been successfully applied to optimize black-box functions where the number of evaluations is severely limited. However, in many real-world applications, it is hard or impossible to know in advance which designs are feasible due to some physical or system limitations. These issues lead to an even more challenging problem of optimizing an unknown function with unknown constraints. In this paper, we observe that in such scenarios optimal solution typically lies on the boundary between feasible and infeasible regions of the design space, making it considerably more difficult than that with interior optima. Inspired by this observation, we propose BE-CBO, a new Bayesian optimization method that efficiently explores the boundary between feasible and infeasible designs. To identify the boundary, we learn the constraints with an ensemble of neural networks that outperform the standard Gaussian Processes for capturing complex boundaries. Our method demonstrates superior performance against state-of-the-art methods through comprehensive experiments on synthetic and real-world benchmarks. Code available at: https://github.com/yunshengtian/BE-CBO

In recent years, interest in gradient-based optimization over Riemannian manifolds has surged. However, a significant challenge lies in the reliance on hyperparameters, especially the learning rate, which requires meticulous tuning by practitioners to ensure convergence at a suitable rate. In this work, we introduce innovative learning-rate-free algorithms for stochastic optimization over Riemannian manifolds, eliminating the need for hand-tuning and providing a more robust and user-friendly approach. We establish high probability convergence guarantees that are optimal, up to logarithmic factors, compared to the best-known optimally tuned rate in the deterministic setting. Our approach is validated through numerical experiments, demonstrating competitive performance against learning-rate-dependent algorithms.