GNC smooths loss function for large-batch SGD, improving generalization.
problem Extremely large-batch SGD leads to poor generalization and converges to sharp minima.
method Gradient noise convolution (GNC) smooths loss function by convolving gradient noise with the loss function.
result GNC achieves state-of-the-art generalization performance for large-scale deep neural networks.
A method predicts GNS of transformer layers using normalization layer norms.
problem Estimating gradient noise scale with minimal variance.
method Simultaneously compute per-example gradient norms and parameter gradients.
result Total GNS is predicted well by normalization layer GNS.
New findings show noisy gradient descent can generalize well, even with non-SGD noise.
problem The role of noise in gradient descent's generalization ability.
method Analyzed the structure of SGD noise and proposed a new noisy gradient descent algorithm.
result Noises in classes different from SGD can also effectively regularize gradient descent.
Adaptive algorithms improve performance in non-convex optimization across various scenarios.
problem Improper handling of noise scales, gradient magnitudes, and smoothness in non-convex optimization.
method Design and analysis of noise-adaptive, scale-free, and generalized algorithms.
result Adaptive algorithms achieve optimal rates and performance in diverse optimization settings.
Study shows how network width affects SGD hyper-parameters and generalization.
problem Understanding how network width impacts SGD hyper-parameters and generalization.
method Generated model families by increasing network width, performed hyper-parameter search.
result Wider networks achieve higher test accuracy and optimal normalized noise scale.
Stochastic Gradient Langevin Dynamics (SGLD) is a sampling scheme for Bayesian modeling adapted to large datasets and models. SGLD relies on the injection of Gaussian Noise at each step of a Stochastic Gradient Descent (SGD) update. In this scheme, every component in the noise vector is independent and has the same sca…
SGD handles label noise with bounds improving over SGLD.
problem Label noise in non-convex optimization.
method Stochastic gradient descent with uniform dissipativity and smoothness conditions, using Wasserstein distance and algorithmic stability.
result Generalization error bounds with a rate of n − 2 / 3 n^{-2/3} n − 2/3 , better than SGLD's n − 1 / 2 n^{-1/2} n − 1/2 . Gradient descent with noise converges to a unique optimum in nonconvex matrix factorization.
problem Gradient descent with noise converges to a unique optimum in nonconvex matrix factorization.
method A perturbed form of gradient descent with arbitrary initialization.
result Gradient descent with noise converges to a unique optimum.
Continuous-time analysis shows SGD with noise prefers flat minima.
problem Optimizing neural networks using SGD with noise.
method Continuous-time model for SGD with noise analysis.
result Optimization prefers flat minima in certain noise regimes.
Improved saliency maps for deep neural networks with reduced noise.
problem Noisy explanations in Integrated Gradients for deep neural networks.
method SmoothTaylor, adaptive noising, and SmoothGrad techniques.
result SmoothTaylor and adaptive noising generate better quality saliency maps.
Adaptively preconditions SGLD for faster convergence and better generalization.
problem Pathological curvature in deep network loss landscapes.
method Adaptive estimation of noise parameters to precondition isotropic gradient noise.
result Adaptively preconditioned SGLD achieves faster convergence and generalization equivalent of SGD.
Scaled sparse linear regression jointly estimates the regression coefficients and noise level in a linear model. It chooses an equilibrium with a sparse regression method by iteratively estimating the noise level via the mean residual square and scaling the penalty in proportion to the estimated noise level. The iterat…
Study mini-batch SGD noise and its limits, proving complexity guarantees.
problem Analyzing the noise in mini-batch SGD and its impact on optimization.
method Examined the conditional covariance and diffusion limits of SGD under different sampling designs.
result Proved mean-square upper bounds and Fisher van Trees lower bounds for SGD, linking them to effective dimension and condition number.
Derives scaling limits and fluctuations for SGD in high dimensions.
problem Understanding SGD behavior in high-dimensional settings with varying noise levels.
method Interacting particle system approach, treating SGD iterates as such, with covariance structure considered.
result Precise three-step phase transition observed in SGD behavior: ballistic, diffusive, then random.
Muon outperforms GD in associative memory learning by balancing frequency components.
problem Training dynamics and scaling behavior of Muon in associative memory learning.
method Study of Muon in a linear associative memory model with softmax retrieval and hierarchical frequency spectrum over query-answer pairs.
result Muon achieves exponential speedup over GD in noiseless case and superior scaling efficiency in noisy case.
Algorithm selects public datasets for private machine learning.
problem Choosing the most suitable public dataset for private machine learning.
method Measures gradient subspace distance between public and private datasets.
result Excess risk scales with the subspace distance between gradients.
DiSK improves DP optimizers by simplifying Kalman filtering for better performance.
problem Performance drop of DP optimizers in large-scale training due to noise injection.
method DiSK uses Kalman filtering to denoise privatized gradients and refine gradient estimations.
result DiSK achieves significant performance improvements over standard DP optimizers in large-scale training.
This research explains why SGD generalizes better than ADAM in deep learning.
problem Understanding the generalization gap between SGD and ADAM in deep learning.
method Analyzing local convergence behaviors through Levy-driven stochastic differential equations (SDEs).
result SGD is more locally unstable and better escapes from sharp minima to flatter ones, leading to better generalization.
One way to avoid overfitting in machine learning is to use model parameters distributed according to a Bayesian posterior given the data, rather than the maximum likelihood estimator. Stochastic gradient Langevin dynamics (SGLD) is one algorithm to approximate such Bayesian posteriors for large models and datasets. SGL…
DP-SGD can update fewer coordinates while maintaining privacy.
problem How to update fewer coordinates in DP-SGD without losing optimization signal.
method TP-TopK (Two-Phase TopK DP-SGD), a two-phase method for coordinate-sparse private training.
result Private training can update fewer coordinates without losing optimization signal, scaling noise with active dimension \(k\) instead of full dimension \(d\).
Proposes a method to improve Byzantine-robustness in compressed federated learning.
problem Byzantine-robustness in compressed federated learning.
method Gradient difference compression and stochastic average gradient algorithm (SAGA).
result The proposed method reaches a neighborhood of the optimal solution at a linear convergence rate.
Improved DP-SGD for variational inference reduces noise and variance.
problem Poor convergence and high variance in variational parameter outputs due to gradient noise in DP-SGD.
method Introduced aligned gradients and iterate averaging to reduce DP-induced noise, and noise-aware posteriors.
result Less noisy gradient estimator and improved parameter estimates for variational inference.
Improved stochastic gradient estimation for deep learning in high dimensions.
problem Inadmissibility of mini-batch gradients in high-dimensional settings.
method Stein-rule shrinkage applied to gradient computation.
result The proposed SR-Adam outperforms Adam in large-batch settings.
FDS tackles long horizon hyperparameter optimization issues.
problem Memory scaling and gradient degradation in long horizon tasks.
method Forward-mode differentiation with sharing (FDS).
result Significantly outperforms greedy gradient-based alternatives.
A new method for deep learning imbalance or noise, ABSGD, improves efficiency and effectiveness.
problem Data imbalance or label noise in deep learning.
method A modification of momentum SGD with individual-level weights proportional to loss values.
result Guaranteed convergence to stationary points of DRO problems, capturing class diversity.
This work improves texture segmentation by automatically tuning hyperparameters for Total-Variation.
problem The challenge is to automatically select hyperparameters for Total-Variation texture segmentation.
method The approach involves extending Stein's unbiased gradient estimator to handle correlated Gaussian noise, leading to an automatic tuning method.
result The method provides an automatic way to select hyperparameters for Total-Variation texture segmentation.
LEGO estimates tangent spaces more robustly than LPCA in noisy data.
problem Estimating tangent spaces in high-noise settings.
method Spectral method using graph Laplacian eigenvectors and gradient orthogonization.
result LEGO yields more robust tangent space estimates than LPCA.
SGD transitions between maxima and minima with varying time scales.
problem Understanding SGD's behavior near critical points in noisy landscapes.
method Analyzing SGD convergence and escape dynamics in 1D landscapes with infinite- and finite-variance noise.
result SGD reliably moves to the basin's minimum unless close to a local maximum, where it can linger.
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.
Derives effective continuous dynamics for adaptive SGD methods.
problem Analyzing noise in adaptive SGD methods.
method Stochastic modified equations framework and Malladi's scaling rules.
result Sampling-induced noise in SGD limits to independent Brownian motions.
New analysis reveals batch size effects on stochastic conditional gradient methods.
problem Understanding the role of batch size in stochastic conditional gradient methods.
method Deriving a new analysis focusing on momentum-based stochastic conditional gradient algorithms (e.g., Scion).
result Increasing batch size initially improves optimization accuracy but can degrade performance beyond a critical threshold.
SpecGD mitigates misalignment in phase retrieval models with anisotropic inputs.
problem Misalignment during gradient descent in phase retrieval models with anisotropic inputs.
method Spectral gradient descent modifies gradient updates to preserve directional information and remove spike amplification.
result SpecGD removes spike amplification, leading to stable alignment and accelerated noise contraction.
AdaGrad-Norm achieves optimal convergence rates for non-convex objectives without tuning.
problem Optimal convergence rates for non-convex, smooth objectives with adaptive step sizes.
method Adaptive SGD (AdaGrad-Norm) with self-tuning step sizes, analyzing under unbounded gradients and affine variance scaling.
result AdaGrad-Norm achieves order optimal convergence rate of $\mathcal{O}\left(\frac{\mathrm{poly}\log(T)}{\sqrt{T}}
ight)$ under optimal assumptions.
Large batch sizes improve model performance but have limits; we found a simple predictor.
problem Understanding the limits of large batch sizes across different domains.
method Empirical model using gradient noise scale to predict optimal batch size.
result Gradient noise scale predicts the largest useful batch size across various domains.
New phases identified in neural scaling laws with compute limits.
problem Understanding neural scaling laws under compute constraints.
method Solved neural scaling model with stochastic gradient descent, derived loss curves, analyzed model-parameter-count phases.
result Identified 4 phases (+3 subphases) in data-complexity/target-complexity phase-plane, derived exponents.
New methods accelerate distributed optimization in noisy networks.
problem Optimizing distributed stochastic gradient methods for noisy, connected networks.
method Developed a framework for choosing stepsize and momentum parameters, proving acceleration and providing performance bounds.
result Distributed accelerated methods achieve acceleration with optimal complexity, reducing bias and variance.
A new method for efficient distributed optimization using trajectory-based normalized gradients.
problem Efficient communication in large-scale distributed optimization.
method A bijective mapping between gradient distributions, using normalized gradients and dynamically extracted references.
result Trajectory-based normalized gradients (TNG) improves communication efficiency in distributed optimization.
Wavelet-based online learning adapts to noisy Besov spaces with high probability.
problem Minimizing integrated squared error in Besov spaces with noisy observations.
method Adaptive wavelet-based online learning algorithm that dynamically adjusts to gradient noise.
result Achieves minimax-optimal integrated squared error with high probability.
DAIS improves AIS for differentiable marginal likelihood estimation.
problem Differentiable marginal likelihood estimation for complex models.
method Proposes Differentiable Annealed Importance Sampling (DAIS) to make AIS differentiable.
result DAIS achieves convergence and consistency in Bayesian linear regression.
Study on neural network dynamics in high dimensions with quadratic activation.
problem Understanding training dynamics in overparameterized neural networks.
method Derivation of gradient flow equations and analysis under l2-regularization.
result Characterization of estimator performance and spectral properties in the high-dimensional limit.
New method reduces neural network memorization of noisy labels.
problem Neural networks memorize label noise, leading to poor generalization.
method Proposes an auxiliary network that predicts gradients without labels.
result Reduces memorization of label-noise, improving generalization.
Study on the noise in SGD minibatches near local minima.
problem Understanding the noise in SGD minibatches near local minima.
method Detailed analysis of SGD noise in linear regression and derivation of a general formula for different types of minima.
result Provides insight into the stability of training neural networks and suggests large learning rates can help generalization.
Classical scaling is shown to be optimal under various noisy conditions.
problem Consistency of classical scaling under general noise conditions.
method Established using finite fourth moments of noise, derived convergence rates, and matching minimax lower bounds.
result Classical scaling achieves minimax optimality in recovering true configuration from noisy dissimilarities.
New bounds for M-SGD show its error distribution is nearly Gaussian.
problem Understanding the error distribution of M-SGD.
method Proved non-asymptotic bounds for M-SGD in Wasserstein distance.
result Error distribution of M-SGD is approximately Gaussian.
Paper improves REINFORCE for VI without restrictive assumptions.
problem Improves REINFORCE for VI without restrictive assumptions.
method Introduces VIMCO- ⋆ \star ⋆ gradient estimator to overcome SNR collapse. result VIMCO- ⋆ \star ⋆ achieves N \sqrt{N} N SNR scaling, superior to existing VIMCO. New guarantees for SGD in non-convex optimization without strict noise bounds.
problem Efficiently escaping saddle points in non-convex optimization.
method Mean-square arguments and relaxed gradient noise variance bounds.
result Gradient descent can efficiently escape saddle points with a more relaxed gradient noise variance bound.
Stochastic gradient algorithms have been the main focus of large-scale learning problems and they led to important successes in machine learning. The convergence of SGD depends on the careful choice of learning rate and the amount of the noise in stochastic estimates of the gradients. In this paper, we propose a new ad…
New methods speed up fitting for large datasets with noisy observations.
problem Fitting large datasets with Gaussian noise and known covariance.
method Two minibatch variants of extreme deconvolution, online EM algorithm, and gradient-based optimisation.
result Methods can scale to larger models and fit larger datasets faster.