The paper analyzes dynamics of momentum in high dimensions with sparse updates.
problem Theoretical analysis of momentum dynamics in high-dimensional sparse settings.
method Theoretical analysis of two models: least squares with sparse inputs and logistic regression with a rare class.
result Characterization of high-dimensional limits of momentum dynamics and phase structure.
Optimization algorithms with momentum, e.g., (ADAM), have been widely used for building deep learning models due to the faster convergence rates compared with stochastic gradient descent (SGD). Momentum helps accelerate SGD in the relevant directions in parameter updating, which can minify the oscillations of parameter…
SQuARM-SGD improves decentralized SGD efficiency with momentum.
problem Efficient decentralized training of large-scale models over networks.
method Fixed local SGD steps with Nesterov's momentum, sparsified and quantized updates, locally computed triggering criterion.
result Convergence rate matches vanilla SGD, momentum improves test performance.
AuON is a linear-time optimizer that improves upon Muon's performance without approximate orthogonal matrices.
problem High memory and computational costs of orthogonal momentum updates.
method AuON uses normalized nonlinear scaling and a 'emergency brake' to handle exploding attention logits.
result AuON achieves strong performance without approximate orthogonal matrices, preserving structural alignment and reconditioning.
Games generalize the single-objective optimization paradigm by introducing different objective functions for different players. Differentiable games often proceed by simultaneous or alternating gradient updates. In machine learning, games are gaining new importance through formulations like generative adversarial netwo…
A new principle for optimizer selection improves training speed and performance.
problem Finding the best optimizer hyperparameters for faster training.
method Formulate optimizer selection as maximizing the expected drop rate in loss, treating gradients and updates as signals and an optimizer as a causal filter.
result Greedy optimizer selection yields stable and effective momentum rules.
We present a unifying framework for adapting the update direction in gradient-based iterative optimization methods. As natural special cases we re-derive classical momentum and Nesterov's accelerated gradient method, lending a new intuitive interpretation to the latter algorithm. We show that a new algorithm, which we …
A new accelerated method with simpler momentum update rules.
problem Optimizing parameters in machine learning models.
method Proposes a novel accelerated stochastic gradient method with simpler momentum update rules.
result The method outperforms Sgdm and Adam in practical problems.
CAdam optimizes online learning by adapting to distribution shifts and noise.
problem Challenges in online learning data, including distribution shifts and noise, affect Adam's performance.
method CAdam uses a confidence-based approach to assess the consistency between momentum and gradients before updating parameters.
result CAdam outperforms other optimizers in various settings with distribution shift or noise.
In this paper we study the problem of convergence and generalization error bound of stochastic momentum for deep learning from the perspective of regularization. To do so, we first interpret momentum as solving an ℓ2-regularized minimization problem to learn the offsets between arbitrary two successive model para…
New algorithm accelerates single-pass SGD for generalized linear prediction.
problem Improving single-pass non-quadratic stochastic optimization.
method Data-dependent proximal method incorporating dual-momentum acceleration.
result Momentum acceleration resolves open problem in streaming setting.
This paper analyzes the skewness of momentum trading strategies.
problem Understanding the skewness of momentum trading strategies.
method Examined linear and nonlinear momentum trading strategies, focusing on skewness.
result Skewness is generally positive and has a term structure.
Federated learning (FL) provides a communication-efficient approach to solve machine learning problems concerning distributed data, without sending raw data to a central server. However, existing works on FL only utilize first-order gradient descent (GD) and do not consider the preceding iterations to gradient update w…
Distributed optimization is essential for training large models on large datasets. Multiple approaches have been proposed to reduce the communication overhead in distributed training, such as synchronizing only after performing multiple local SGD steps, and decentralized methods (e.g., using gossip algorithms) to decou…
New algorithm reduces FL sample and communication costs.
problem Optimizing FL for minimal samples and rounds.
method Stochastic two-sided momentum algorithm.
result Achieves near-optimal sample and communication complexities.
SMG combines shuffling and momentum for non-convex optimization.
problem Non-convex finite-sum optimization problems.
method Shuffling Gradient-based method with momentum.
result Established state-of-the-art convergence rates for SMG.
AB dynamically scales gradients to mitigate asynchronous training delays.
problem Gradient delay in asynchronous training reduces model performance.
method Adaptive Braking (AB) dynamically scales gradients based on alignment.
result AB enables training with up to 32 update steps of delay without accuracy loss.
A new Bayesian filtering method speeds up stochastic Newton optimization.
problem Minimizing log-convex functions using stochastic methods.
method Contextualizes the problem as Bayesian inference, applying Bayesian filtering to update estimates.
result Establishes conditions for diminishing effect of older observations, akin to momentum.
Adaptive momentum method solves non-convex min-max problems.
problem Non-convex min-max optimization problems in training generative adversarial networks.
method Proposes an adaptive momentum algorithm for non-convex min-max optimization.
result Establishes non-asymptotic convergence rates for the proposed algorithm.
Unified framework for analyzing batch updating methods with noisy gradients.
problem Analyzing convergence of batch updating methods with noisy gradients and approximations.
method Unified framework using convergence of stochastic processes.
result Establishes a general theorem for most known convergence results.
Improved sample complexity for actor-critic algorithms in MDPs.
problem Achieving optimal policies with limited data in reinforcement learning.
method Single-timescale actor-critic with STORM (STOchastic Recursive Momentum) and a sample buffer.
result Optimal sample complexity of O(ε−2) for ε-optimal policies. New loss function connects learning rate and momentum.
problem Finding optimal learning rate and momentum empirically.
method Proposes a new information-theoretical loss function.
result Loss, learning rate, and momentum are closely connected.
Two major momentum-based techniques that have achieved tremendous success in optimization are Polyak's heavy ball method and Nesterov's accelerated gradient. A crucial step in all momentum-based methods is the choice of the momentum parameter m which is always suggested to be set to less than 1. Although the choice…
FedGLOMO accelerates FL convergence for non-convex functions.
problem Efficiently solving non-convex optimization problems in federated learning with client heterogeneity.
method Combines global and local momentum updates to reduce variance and improve convergence rate.
result Achieves O(ε−1.5) convergence to ε-stationary point, compared to O(ε−2). It is common practice to decay the learning rate. Here we show one can usually obtain the same learning curve on both training and test sets by instead increasing the batch size during training. This procedure is successful for stochastic gradient descent (SGD), SGD with momentum, Nesterov momentum, and Adam. It reache…
FetchSGD reduces communication in federated learning with sketching.
problem Communication bottlenecks and convergence issues in federated learning.
method FetchSGD uses Count Sketch to compress and merge model updates efficiently.
result FetchSGD achieves high compression rates and good convergence without sparse client participation.
STORM-PG uses momentum for faster policy gradient updates.
problem Improving policy gradient methods for reinforcement learning.
method Introduces STORM-PG, a SARAH-based algorithm with exponential moving average.
result Achieves O(1/ε3) sample complexity, matching best-known rate. A new correction term improves sample efficiency in deep reinforcement learning.
problem Momentum accumulation in TD learning leads to doubly stale gradients.
method Proposed a correction term to address the issue of doubly stale gradients.
result Improves sample efficiency in policy evaluation.
Mime algorithm improves federated learning by adapting centralized methods.
problem Mitigating client drift in federated learning.
method Combines control variates and server-level statistics to adapt centralized algorithms to federated learning.
result Mime outperforms any centralized method in federated learning.
ASVGD accelerates SVGD for efficient sampling.
problem Slow SVGD in high-dimensional sampling.
method Accelerated gradient flow in a metric space of probability densities, using Nesterov's method and momentum-based updates.
result ASVGD outperforms SVGD and other methods in sampling efficiency.
Proposes EDM algorithm to accelerate model training in distributed networks.
problem Hindered effectiveness of distributed stochastic optimization algorithms due to data heterogeneity and network sparsity.
method Introduces Exact-Diffusion with Momentum (EDM) algorithm, incorporating momentum techniques to mitigate bias and enhance convergence rate.
result EDM algorithm converges sub-linearly to the optimal solution, radius independent of data heterogeneity, for non-convex objective functions.
We introduce novel variants of momentum by incorporating the variance of the stochastic loss function. The variance characterizes the confidence or uncertainty of the local features of the averaged loss surface across the i.i.d. subsets of the training data defined by the mini-batches. We show two applications of the g…
New methods solve optimization problems with heavy-tailed noise, improving upon existing complexity bounds.
problem Optimization problems with heavy-tailed noise and weakly average smoothness.
method Normalized stochastic first-order methods with Polyak, multi-extrapolated, and recursive momentum.
result First-order oracle complexity results for finding approximate stochastic stationary points under heavy-tailed noise.
Adam performs better with equal momentum parameters, revealing a gradient scale invariance principle.
problem Why Adam performs better with β1=β2. method Formalized gradient scale invariance and proved it for Adam with equal β1 and β2. result Adam becomes gradient scale invariant of first order if and only if β1=β2. While momentum-based accelerated variants of stochastic gradient descent (SGD) are widely used when training machine learning models, there is little theoretical understanding on the generalization error of such methods. In this work, we first show that there exists a convex loss function for which the stability gap fo…
New insights into training machine learning models with momentum.
problem Lack of theoretical understanding on the generalization error of momentum-based methods.
method Analyzed modified momentum-based update rule (SGDEM) for smooth Lipschitz loss functions.
result SGDEM admits an upper-bound on the generalization error for smooth Lipschitz loss functions.
FedNNNN improves FL by adjusting model update vector norms.
problem Slow convergence and low prediction accuracy in FL.
method Norm-Normalized Neural Network Aggregation (FedNNNN) with momentum control.
result Up to 5.4% accuracy improvement on multiple datasets.
Ringmaster LMO accelerates training in distributed systems by asynchronously updating neural networks.
problem Asynchronous training in distributed systems where workers compute gradients at different speeds.
method Introduces an asynchronous LMO-based momentum method for unconstrained stochastic nonconvex optimization.
result Establishes convergence guarantees and time complexity bounds for asynchronous LMO-based updates.
SUSTAIN algorithm tackles stochastic bilevel optimization with near-optimal complexity.
problem Stochastic bilevel optimization problems with specific convexity and smoothness properties.
method SUSTAIN algorithm using single-timescale double-momentum stochastic approximation.
result SUSTAIN achieves near-optimal complexity for finding ε-stationary solutions.
Omega method mitigates noise in stochastic game optimization.
problem Noise sensitivity and convergence issues in stochastic game optimization.
method Omega method incorporates EMA of historic gradients in its update rule.
result Omega method outperforms optimistic gradient method in stochastic games.
This work studies the implicit bias of mini-batch SGD in classification.
problem Understanding the implicit bias of mini-batch SGD in multi-class classification.
method Characterizes how batch size, momentum, and variance reduction affect convergence and max-margin behavior under different norms.
result Momentum enables small-batch convergence to an approximate max-margin solution, while variance reduction recovers the exact full-batch bias.
DFedAvgM is a decentralized FedAvg with momentum for privacy and communication efficiency.
problem Efficiently train models with privacy and communication efficiency in federated learning.
method Decentralized Federated Averaging with Momentum (DFedAvgM) on clients connected by an undirected graph, using stochastic gradient descent with momentum and quantization.
result DFedAvgM converges under trivial assumptions and can be improved with the PŁ property, numerically verified.
We accelerate PMD algorithms for reinforcement learning using functional methods.
problem Improving reinforcement learning algorithms for large-scale optimization.
method Proposed a momentum-based update for PMD algorithms leveraging duality.
result The proposed method accelerates reinforcement learning algorithms.
CoolMomentum combines momentum and Simulated Annealing for deep learning optimization.
problem Global optimization of non-convex functions in deep learning.
method Discretized Langevin dynamics with Simulated Annealing.
result CoolMomentum achieves high accuracy on Resnet-20 on Cifar-10 and Efficientnet-B0 on Imagenet.
Deep neural networks are vulnerable to adversarial examples, which poses security concerns on these algorithms due to the potentially severe consequences. Adversarial attacks serve as an important surrogate to evaluate the robustness of deep learning models before they are deployed. However, most of existing adversaria…
Recently, research on accelerated stochastic gradient descent methods (e.g., SVRG) has made exciting progress (e.g., linear convergence for strongly convex problems). However, the best-known methods (e.g., Katyusha) requires at least two auxiliary variables and two momentum parameters. In this paper, we propose a fast …
Paper proves suboptimal convergence rate of last iterate for SGDM.
problem Proves suboptimal convergence rate of last iterate for SGDM.
method Focuses on convergence rate of last iterate of SGDM, introduces Follow-The-Regularized-Leader-based algorithms.
result Shows optimal convergence rate of last iterate for unconstrained convex stochastic optimization problems.
This paper studies accelerations in Q-learning algorithms. We propose an accelerated target update scheme by incorporating the historical iterates of Q functions. The idea is conceptually inspired by the momentum-based accelerated methods in the optimization theory. Conditions under which the proposed accelerated algor…