Search

We introduce the CRONOS algorithm for convex optimization of two-layer neural networks. CRONOS is the first algorithm capable of scaling to high-dimensional datasets such as ImageNet, which are ubiquitous in modern deep learning. This significantly improves upon prior work, which has been restricted to downsampled versions of MNIST and CIFAR-10. Taking CRONOS as a primitive, we then develop a new algorithm called CRONOS-AM, which combines CRONOS with alternating minimization, to obtain an algorithm capable of training multi-layer networks with arbitrary architectures. Our theoretical analysis proves that CRONOS converges to the global minimum of the convex reformulation under mild assumptions. In addition, we validate the efficacy of CRONOS and CRONOS-AM through extensive large-scale numerical experiments with GPU acceleration in JAX. Our results show that CRONOS-AM can obtain comparable or better validation accuracy than predominant tuned deep learning optimizers on vision and language tasks with benchmark datasets such as ImageNet and IMDb. To the best of our knowledge, CRONOS is the first algorithm which utilizes the convex reformulation to enhance performance on large-scale learning tasks.

Kernel ridge regression (KRR) is a fundamental computational tool, appearing in problems that range from computational chemistry to health analytics, with a particular interest due to its starring role in Gaussian process regression. However, it is challenging to scale KRR solvers to large datasets: with $n$ training points, a direct solver (i.e., Cholesky decomposition) uses $O(n^2)$ storage and $O(n^3)$ flops. Iterative methods for KRR, such as preconditioned conjugate gradient (PCG), avoid the cubic scaling of direct solvers and often use low-rank preconditioners; a rank $r$ preconditioner uses $O(rn)$ storage and each iteration requires $O(n^2)$ flops. To reduce the storage and iteration complexity of iterative solvers for KRR, we propose ASkotch ($\textbf{A}$ccelerated $\textbf{s}$calable $\textbf{k}$ernel $\textbf{o}$p$\textbf{t}$imization using block $\textbf{c}$oordinate descent with $\textbf{H}$essian preconditioning). For a given block size $|b| << n$, each iteration of ASkotch uses $O(r|b| + n)$ storage and $O(n|b|)$ flops, so ASkotch scales better than Cholesky decomposition and PCG. We prove that ASkotch obtains linear convergence to the optimum, with the convergence rate depending on the square roots of the $\textit{preconditioned}$ block condition numbers. Furthermore, we solve KRR problems that were considered to be impossibly large while using limited computational resources: we show that ASkotch outperforms PCG methods with respect to generalization error on large-scale KRR (up to $n = 10^8$) and KRR classification tasks (up to $n = 10^7$) while running each of our experiments on $\textit{a single 12 GB Titan V GPU}$. Our work opens up the possibility of as-yet-unimagined applications of KRR across a wide range of disciplines., Comment: 40 pages (including appendices), 14 figures, 3 tables

This paper explores challenges in training Physics-Informed Neural Networks (PINNs), emphasizing the role of the loss landscape in the training process. We examine difficulties in minimizing the PINN loss function, particularly due to ill-conditioning caused by differential operators in the residual term. We compare gradient-based optimizers Adam, L-BFGS, and their combination Adam+L-BFGS, showing the superiority of Adam+L-BFGS, and introduce a novel second-order optimizer, NysNewton-CG (NNCG), which significantly improves PINN performance. Theoretically, our work elucidates the connection between ill-conditioned differential operators and ill-conditioning in the PINN loss and shows the benefits of combining first- and second-order optimization methods. Our work presents valuable insights and more powerful optimization strategies for training PINNs, which could improve the utility of PINNs for solving difficult partial differential equations., Comment: ICML 2024 Oral; 33 pages (including appendices), 10 figures, 3 tables

We introduce the GEneralized Newton Inexact Operator Splitting solver (GeNIOS) for large-scale convex optimization. GeNIOS speeds up ADMM by approximately solving approximate subproblems: it uses a second-order approximation to the most challenging ADMM subproblem and solves it inexactly with a fast randomized solver. Despite these approximations, GeNIOS retains the convergence rate of classic ADMM and can detect primal and dual infeasibility from the algorithm iterates. At each iteration, the algorithm solves a positive-definite linear system that arises from a second-order approximation of the first subproblem and computes an approximate proximal operator. GeNIOS solves the linear system using an indirect solver with a randomized preconditioner, making it particularly useful for large-scale problems with dense data. Our high-performance open-source implementation in Julia allows users to specify convex optimization problems directly (with or without conic reformulation) and allows extensive customization. We illustrate GeNIOS's performance on a variety of problem types. Notably, GeNIOS is more than ten times faster than existing solvers on large-scale, dense problems.

This paper introduces PROMISE ($\textbf{Pr}$econditioned Stochastic $\textbf{O}$ptimization $\textbf{M}$ethods by $\textbf{I}$ncorporating $\textbf{S}$calable Curvature $\textbf{E}$stimates), a suite of sketching-based preconditioned stochastic gradient algorithms for solving large-scale convex optimization problems arising in machine learning. PROMISE includes preconditioned versions of SVRG, SAGA, and Katyusha; each algorithm comes with a strong theoretical analysis and effective default hyperparameter values. In contrast, traditional stochastic gradient methods require careful hyperparameter tuning to succeed, and degrade in the presence of ill-conditioning, a ubiquitous phenomenon in machine learning. Empirically, we verify the superiority of the proposed algorithms by showing that, using default hyperparameter values, they outperform or match popular tuned stochastic gradient optimizers on a test bed of $51$ ridge and logistic regression problems assembled from benchmark machine learning repositories. On the theoretical side, this paper introduces the notion of quadratic regularity in order to establish linear convergence of all proposed methods even when the preconditioner is updated infrequently. The speed of linear convergence is determined by the quadratic regularity ratio, which often provides a tighter bound on the convergence rate compared to the condition number, both in theory and in practice, and explains the fast global linear convergence of the proposed methods., Comment: 52 pages, 9 Figures

This paper investigates two randomized preconditioning techniques for solving kernel ridge regression (KRR) problems with a medium to large number of data points ($10^4 \leq N \leq 10^7$), and it introduces two new methods with state-of-the-art performance. The first method, RPCholesky preconditioning, accurately solves the full-data KRR problem in $O(N^2)$ arithmetic operations, assuming sufficiently rapid polynomial decay of the kernel matrix eigenvalues. The second method, KRILL preconditioning, offers an accurate solution to a restricted version of the KRR problem involving $k \ll N$ selected data centers at a cost of $O((N + k^2) k \log k)$ operations. The proposed methods solve a broad range of KRR problems, making them ideal for practical applications., Comment: 29 pages, 11 figures

This paper presents GeNI-ADMM, a framework for large-scale composite convex optimization, that facilitates theoretical analysis of both existing and new approximate ADMM schemes. GeNI-ADMM encompasses any ADMM algorithm that solves a first- or second-order approximation to the ADMM subproblem inexactly. GeNI-ADMM exhibits the usual $\mathcal O(1/t)$-convergence rate under standard hypotheses and converges linearly under additional hypotheses such as strong convexity. Further, the GeNI-ADMM framework provides explicit convergence rates for ADMM variants accelerated with randomized linear algebra, such as NysADMM and sketch-and-solve ADMM, resolving an important open question on the convergence of these methods. This analysis quantifies the benefit of improved approximations and can inspire the design of new ADMM variants with faster convergence., Comment: 37 pages, 4 figures, 2 tables

SketchySGD improves upon existing stochastic gradient methods in machine learning by using randomized low-rank approximations to the subsampled Hessian and by introducing an automated stepsize that works well across a wide range of convex machine learning problems. We show theoretically that SketchySGD with a fixed stepsize converges linearly to a small ball around the optimum. Further, in the ill-conditioned setting we show SketchySGD converges at a faster rate than SGD for least-squares problems. We validate this improvement empirically with ridge regression experiments on real data. Numerical experiments on both ridge and logistic regression problems with dense and sparse data, show that SketchySGD equipped with its default hyperparameters can achieve comparable or better results than popular stochastic gradient methods, even when they have been tuned to yield their best performance. In particular, SketchySGD is able to solve an ill-conditioned logistic regression problem with a data matrix that takes more than $840$GB RAM to store, while its competitors, even when tuned, are unable to make any progress. SketchySGD's ability to work out-of-the box with its default hyperparameters and excel on ill-conditioned problems is an advantage over other stochastic gradient methods, most of which require careful hyperparameter tuning (especially of the learning rate) to obtain good performance and degrade in the presence of ill-conditioning., Comment: 65 pages, 43 figures, 8 tables

This paper develops a scalable new algorithm, called NysADMM, to minimize a smooth convex loss function with a convex regularizer. NysADMM accelerates the inexact Alternating Direction Method of Multipliers (ADMM) by constructing a preconditioner for the ADMM subproblem from a randomized low-rank Nystr\"om approximation. NysADMM comes with strong theoretical guarantees: it solves the ADMM subproblem in a constant number of iterations when the rank of the Nystr\"om approximation is the effective dimension of the subproblem regularized Gram matrix. In practice, ranks much smaller than the effective dimension can succeed, so NysADMM uses an adaptive strategy to choose the rank that enjoys analogous guarantees. Numerical experiments on real-world datasets demonstrate that NysADMM can solve important applications, such as the lasso, logistic regression, and support vector machines, in half the time (or less) required by standard solvers. The breadth of problems on which NysADMM beats standard solvers is a surprise: it suggests that ADMM is a dominant paradigm for numerical optimization across a wide range of statistical learning problems that are usually solved with bespoke methods.

This paper introduces the Nystr\"om PCG algorithm for solving a symmetric positive-definite linear system. The algorithm applies the randomized Nystr\"om method to form a low-rank approximation of the matrix, which leads to an efficient preconditioner that can be deployed with the conjugate gradient algorithm. Theoretical analysis shows that preconditioned system has constant condition number as soon as the rank of the approximation is comparable with the number of effective degrees of freedom in the matrix. The paper also develops adaptive methods that provably achieve similar performance without knowledge of the effective dimension. Numerical tests show that Nystr\"om PCG can rapidly solve large linear systems that arise in data analysis problems, and it surpasses several competing methods from the literature., Comment: 37 pages, 3 figures

Models like LASSO and ridge regression are extensively used in practice due to their interpretability, ease of use, and strong theoretical guarantees. Cross-validation (CV) is widely used for hyperparameter tuning in these models, but do practical optimization methods minimize the true out-of-sample loss? A recent line of research promises to show that the optimum of the CV loss matches the optimum of the out-of-sample loss (possibly after simple corrections). It remains to show how tractable it is to minimize the CV loss. In the present paper, we show that, in the case of ridge regression, the CV loss may fail to be quasiconvex and thus may have multiple local optima. We can guarantee that the CV loss is quasiconvex in at least one case: when the spectrum of the covariate matrix is nearly flat and the noise in the observed responses is not too high. More generally, we show that quasiconvexity status is independent of many properties of the observed data (response norm, covariate-matrix right singular vectors and singular-value scaling) and has a complex dependence on the few that remain. We empirically confirm our theory using simulated experiments., Comment: Published in NeurIPS 2021