Preconditioned non-convex gradient descent improves noisy matrix estimation.
problem Estimating low-rank matrices from noisy measurements.
method Preconditioned non-convex gradient descent for noisy measurements.
result Preconditioned method converges to minimax optimal estimate at a linear rate.
PolarGrad optimizes deep learning models by considering matrix structure, outperforming Adam and Muon.
problem Efficient optimization of large-scale neural networks and language models.
method A unifying framework for analyzing matrix-aware preconditioned methods, including PolarGrad.
result PolarGrad outperforms Adam and Muon in various tasks.
Optimal preconditioning improves Langevin sampling efficiency.
problem Improving sampling efficiency in high-dimensional target distributions.
method Optimal preconditioning using Fisher information, applied to MALA.
result Adaptive MCMC scheme significantly outperforms other methods.
Preconditioned gradient methods are among the most general and powerful tools in optimization. However, preconditioning requires storing and manipulating prohibitively large matrices. We describe and analyze a new structure-aware preconditioning algorithm, called Shampoo, for stochastic optimization over tensor spaces.…
Bias correction improves language model training performance.
problem Stochastic update bias in preconditioned optimizers.
method Cross-fitted preconditioning and variance-corrected inversion.
result Reduces held-out pretraining loss by 0.15 nats.
Two methods solve kernel ridge regression problems efficiently.
problem Solving kernel ridge regression problems with large datasets.
method RPCholesky and KRILL preconditioning techniques.
result Efficient solutions to KRR problems with strong guarantees.
Gradient-based MCMC for discrete spaces improves sampling performance.
problem Sampling in discrete spaces using traditional methods is challenging.
method Introduced new discrete Metropolis-Hastings samplers inspired by MALA, with a novel preconditioning technique.
result Demonstrated strong empirical performance across various challenging sampling problems.
Improved convergence for overparameterized low-rank matrix sensing.
problem Overparameterized low-rank matrix sensing with unknown rank and ill-conditioning.
method ScaledGD(λ) - preconditioned gradient descent method. result ScaledGD(λ) converges at a constant linear rate after a logarithmic number of iterations. We provide an online convex optimization algorithm with regret that interpolates between the regret of an algorithm using an optimal preconditioning matrix and one using a diagonal preconditioning matrix. Our regret bound is never worse than that obtained by diagonal preconditioning, and in certain setting even surpass…
Stochastic Gradient Langevin Dynamics infuses isotropic gradient noise to SGD to help navigate pathological curvature in the loss landscape for deep networks. Isotropic nature of the noise leads to poor scaling, and adaptive methods based on higher order curvature information such as Fisher Scoring have been proposed t…
Dual Space Preconditioning speeds up gradient descent in overparameterized models.
problem Improving convergence of gradient descent in overparameterized linear models.
method Introducing a novel preconditioner of the form ablaK for convex K and applying it to overparameterized linear models. result The iterates of the preconditioned gradient descent converge to a solution W∞ satisfying XW∞=Y. Adaptive learning rate algorithms such as RMSProp are widely used for training deep neural networks. RMSProp offers efficient training since it uses first order gradients to approximate Hessian-based preconditioning. However, since the first order gradients include noise caused by stochastic optimization, the approxima…
Stochastic gradient descent (SGD) still is the workhorse for many practical problems. However, it converges slow, and can be difficult to tune. It is possible to precondition SGD to accelerate its convergence remarkably. But many attempts in this direction either aim at solving specialized problems, or result in signif…
The computational and storage complexity of kernel machines presents the primary barrier to their scaling to large, modern, datasets. A common way to tackle the scalability issue is to use the conjugate gradient algorithm, which relieves the constraints on both storage (the kernel matrix need not be stored) and computa…
Stochastic Gradient Descent improved for various Hilbert scales and misspecified models.
problem Understanding and optimizing SGD in Hilbert scales for machine learning.
method Extending SGD analysis to Hilbert scales, including Sobolev and Diffusion spaces, and showing the effects of smoothness and preconditioning.
result Violation of smoothness assumption affects learning rate; preconditioning in Hilbert scales reduces the number of iterations for misspecified models.
We propose a novel Riemannian manifold preconditioning approach for the tensor completion problem with rank constraint. A novel Riemannian metric or inner product is proposed that exploits the least-squares structure of the cost function and takes into account the structured symmetry that exists in Tucker decomposition…
Bayesian sparse learning method improves deep neural network efficiency.
problem Sparse learning in deep neural networks with complex geometry.
method Preconditioned stochastic gradient Langevin Dynamics (PSGLD) for sampling and adaptive optimization of hyperparameters.
result The proposed algorithm achieves asymptotic convergence with controlled bias.
TDprop uses Jacobi preconditioning to improve adaptive optimizers in Deep RL.
problem Improving performance of adaptive optimizers in Deep RL.
method TDprop computes per-parameter learning rates based on Jacobi preconditioning of the TD update rule.
result TDprop matches or exceeds Adam's performance in Deep RL experiments, suggesting Jacobi preconditioning can improve adaptive methods.
Standard gradient descent methods are susceptible to a range of issues that can impede training, such as high correlations and different scaling in parameter space.These difficulties can be addressed by second-order approaches that apply a pre-conditioning matrix to the gradient to improve convergence. Unfortunately, s…
This paper studies the performance of a recently proposed preconditioned stochastic gradient descent (PSGD) algorithm on recurrent neural network (RNN) training. PSGD adaptively estimates a preconditioner to accelerate gradient descent, and is designed to be simple, general and easy to use, as stochastic gradient desce…
Randomized block-diagonal preconditioning improves parallel learning convergence.
problem Improving convergence of gradient-based optimization methods in parallel settings.
method Randomization of coordinates during optimization to repartition tasks.
result Randomization significantly improves convergence of block-diagonal preconditioned methods.
Effective training of deep neural networks suffers from two main issues. The first is that the parameter spaces of these models exhibit pathological curvature. Recent methods address this problem by using adaptive preconditioning for Stochastic Gradient Descent (SGD). These methods improve convergence by adapting to th…
Gradient descent with preconditioning finds global optima in overparameterized nonconvex factorization.
problem Finding global optima in nonconvex Burer-Monteiro factorization.
method Preconditioned gradient descent for overparameterized nonconvex function minimization.
result Gradient descent with preconditioning achieves linear convergence in the overparameterized case.
Unified framework for understanding and optimizing training acceleration.
problem Challenges in optimizing training with regularization and acceleration techniques.
method Explains how AdaGrad, RMSProp, and Adam accelerate training, and derives a generalization for L1-regularization. result Derives a unified mathematical framework for understanding and optimizing training acceleration.
APGD algorithm efficiently recovers over-parameterized matrices from noisy measurements.
problem Matrix sensing problem with over-parameterization and noisy measurements.
method Alternating preconditioned gradient descent (APGD) algorithm incorporating preconditioning terms.
result APGD converges to a near-optimal error at a linear rate.
New stability analysis improves generalization of multipass SGD.
problem Improper preconditioning affects generalization in multipass SGD.
method Developed on-average stability analysis for multipass SGD.
result Proper preconditioning yields optimal effective dimension dependence.
Stein variational gradient descent (SVGD) is a particle-based inference algorithm that leverages gradient information for efficient approximate inference. In this work, we enhance SVGD by leveraging preconditioning matrices, such as the Hessian and Fisher information matrix, to incorporate geometric information into SV…
Polyak-Ruppert CLT for SA-Adam with momentum and non-convergent adaptive preconditioning
problem Adaptive optimizers combining momentum and non-convergent preconditioning
method Proving positive drift stability and a non-autonomous Polyak-Ruppert CLT for SA-Adam
result The iterate-marginal covariance is exactly the plain stochastic gradient descent (SGD) sandwich
Increasing the batch size is a popular way to speed up neural network training, but beyond some critical batch size, larger batch sizes yield diminishing returns. In this work, we study how the critical batch size changes based on properties of the optimization algorithm, including acceleration and preconditioning, thr…
SignSGD analysis quantifies its effects in high dimensions.
problem Understanding signSGD's effects in high-dimensional settings.
method High-dimensional analysis of signSGD, deriving SDE and ODE for risk.
result Quantification of signSGD's effects: effective learning rate, noise compression, diagonal preconditioning, gradient noise reshaping.
In this paper we introduce a parameter dependent class of Krylov-based methods, namely CD, for the solution of symmetric linear systems. We give evidence that in our proposal we generate sequences of conjugate directions, extending some properties of the standard Conjugate Gradient (CG) method, in order to preserve the…
New analysis shows SNG's effectiveness in small samples.
problem Limited understanding of SNG in small data settings.
method Sketch-and-project analysis of SNG.
result Global convergence and rate characterization for SNG.
State-of-the-art models are now trained with billions of parameters, reaching hardware limits in terms of memory consumption. This has created a recent demand for memory-efficient optimizers. To this end, we investigate the limits and performance tradeoffs of memory-efficient adaptively preconditioned gradient methods.…
This work analyzes Adam's preconditioning effect on quadratic functions and quantifies its impact on condition number.
problem Understanding and quantifying the preconditioning effect of Adam to alleviate ill-conditioning in gradient descent.
method Detailed analysis of Adam's preconditioning effect for quadratic functions, including empirical evidence.
result Adam can mitigate the condition number but at a dimension-dependent cost, with specific bounds for different types of Hessians.
Higher-order ODE solvers improve deep learning performance.
problem Improving deep learning performance using higher-order ODE solvers.
method Evaluation and improvement of Runge-Kutta (RK) methods for deep learning.
result Higher-order RK solvers can improve deep learning performance by incorporating key ingredients of optimizers.
Gradient descent converges geometrically to optimal self-attention parameters.
problem Training softmax self-attention layers for linear regression.
method Structure-aware gradient descent with preconditioner and regularizer.
result Gradient descent converges geometrically to global minima.
NeuralIF uses neural networks to improve preconditioning for faster CG convergence.
problem Improving convergence of conjugate gradient method for large-scale sparse systems.
method Data-driven approach using graph neural networks to generate incomplete factorization.
result Data-driven preconditioners accelerate convergence of conjugate gradient method.
Nystrom approximation speeds up kernel model training.
problem Slow convergence in kernel models due to poor conditioning.
method Spectral preconditioning with Nystrom approximation for scalability.
result Nystrom approximation accelerates gradient descent nearly as well as exact preconditioner.
PDHAMS improves sampling for discrete distributions with quadratic potential functions.
problem Sampling discrete distributions efficiently and accurately.
method Integrates a second-order approximation of the potential function and uses Gaussian integral trick.
result PDHAMS yields superior performance compared to other methods.
SAPPHIRE tackles ill-conditioned rERM problems with faster convergence.
problem Ill-conditioned objectives and non-smooth regularizers in rERM.
method Sketch-based preconditioning and scaled proximal mapping.
result Achieves condition-number-free linear convergence.
New theory explains why normalization is preferred in SGD under heavy-tailed noise.
problem Understanding why normalization is preferred in stochastic gradient descent (SGD) under heavy-tailed noise.
method Developed a worst-case complexity theory for stochastically preconditioned SGD and its variants.
result Normalization guarantees convergence at optimal rates, while clipping may fail in the worst case.
Uncertainty sampling, a popular active learning algorithm, is used to reduce the amount of data required to learn a classifier, but it has been observed in practice to converge to different parameters depending on the initialization and sometimes to even better parameters than standard training on all the data. In this…
New particle-based VI algorithm expands function class and improves scalability.
problem Limited function class in particle-based VI algorithms restricts flexibility and scalability.
method Introduces a functional regularization term to expand the function class and proposes PFG algorithm.
result Proposed PFG algorithm has larger function class, improved scalability, better adaptation to ill-conditioned distributions, and provable convergence.
The paper explores how dynamic preconditioning affects the CLT in online averaging.
problem When does dynamic preconditioning preserve the Polyak-Ruppert CLT?
method The authors decompose the averaged error and identify a stabilization-rate threshold for the CLT to hold.
result The CLT holds if the dynamic remainder vanishes in L2 and the stabilization rate exceeds a threshold. Novel probabilistic solver speeds up solving related linear systems.
problem Efficiently solving multiple related linear systems.
method Probabilistic linear solver over the parameter space, leveraging solved systems.
result Faster and more efficient solution of related linear systems.
This work analyzes how preconditioning affects generalization in machine learning models.
problem The impact of preconditioning on the generalization of machine learning models.
method An asymptotic bias-variance decomposition of the generalization error for ridgeless regression under various preconditioners.
result The optimal preconditioner depends on label noise, model specification, and signal alignment, with NGD potentially better under certain conditions.
We describe a framework for deriving and analyzing online optimization algorithms that incorporate adaptive, data-dependent regularization, also termed preconditioning. Such algorithms have been proven useful in stochastic optimization by reshaping the gradients according to the geometry of the data. Our framework capt…
Muon optimizer simplifies matrix optimization with spectral orthogonalization.
problem Matrix optimization challenges, especially with large condition numbers.
method Simplified Muon optimizer using spectral orthogonalization of gradients.
result Simplified Muon converges linearly with independent scalar sequences, outperforming gradient descent and Adam.