Full-batch GD outperforms one-pass SGD in learning a single-index model with quadratic activation.
problem Learning a single-index model with quadratic activation using gradient descent.
method Full-batch gradient descent compared to one-pass stochastic gradient descent (SGD) on a correlation loss.
result Full-batch GD requires only n≃d samples for strong recovery, while one-pass SGD requires n≳dlogd samples. Adam's bias shifts from full-batch to max-margin of different norms for separable data.
problem Understanding Adam's implicit bias in the incremental batch setting.
method Analyzing incremental Adam on linearly separable data, constructing datasets, and using a proxy algorithm.
result Incremental Adam can converge to different max-margin classifiers depending on the dataset and batching scheme.
Full-batch GD achieves generalization close to any stationary point with fewer assumptions.
problem Generalization and excess risk bounds for smooth losses, including non-Lipschitz and nonconvex cases.
method Path-dependent analysis of GD's generalization error, focusing on optimization error and stability.
result Generalization error is tightly bound in terms of optimization error and iteration count, bypassing common assumptions.
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.
New research shows many batch selection methods for training work just as well as full batch training.
problem Finding optimal batch selection methods for training.
method Analysis of mini-batch Gradient Descent (GD) and Stochastic GD (SGD) with various batch selection rules.
result All mini-batch schedules, including deterministic ones, generalize optimally for smooth Lipschitz-convex/nonconvex/strongly-convex loss functions.
SGD generalizes better than GD with similar number of iterations.
problem Generalization performance of SGD vs GD in convex optimization.
method Comparison of SGD and GD in stochastic convex optimization model.
result GD requires more iterations to match SGD's generalization performance.
The standard L-BFGS method relies on gradient approximations that are not dominated by noise, so that search directions are descent directions, the line search is reliable, and quasi-Newton updating yields useful quadratic models of the objective function. All of this appears to call for a full batch approach, but sinc…
New SGDA method speeds up nonconvex minimax optimization.
problem Improving convergence of nonconvex minimax optimization.
method SGDA with random reshuffling for nonconvex-PŁ objectives.
result Convergence rates faster than with-replacement SGDA.
New findings show mini-batch SGD operates in a 'Edge of Stochastic Stability' regime.
problem Understanding the stability and convergence of mini-batch SGD.
method Analyzing the mini-batch Hessian and its directional curvature.
result Mini-batch SGD operates in a different stability regime (Edge of Stochastic Stability) compared to full-batch GD.
Adam's generalization performance is improved by batch size and weight decay in neural networks.
problem Understanding how batch size and weight decay affect Adam's generalization in neural networks.
method Theoretical analysis of two-layer over-parameterized CNNs on image data.
result Adam's mini-batch variants can achieve near-zero test error, unlike full-batch Adam.
Paper explains contrastive learning using cosine similarity and proposes mitigations for batch size effects.
problem Understanding and improving contrastive learning through batch size effects.
method Unified framework of cosine similarity, theoretical insights, and auxiliary loss.
result Performance improvement in small-batch settings through proposed auxiliary loss.
Analysis of SGD+M convergence rates in high dimensions with batch size considerations.
problem Understanding convergence rates of SGD+M in high-dimensional settings.
method Analyzing the dynamics of SGD+M on least squares problems with large batch sizes and dimensions.
result Identifies the implicit conditioning ratio (ICR) that regulates SGD+M's acceleration and convergence rates.
Paper analyzes gradient descent with noisy data copies for linear regression, showing regularization and acceleration effects.
problem Improving generalization in machine learning through data augmentation with noise.
method Gradient descent with on-line noisy copies for linear regression analysis.
result Training with on-line noisy copies is equivalent to ridge regularization with a specific regularization parameter.
Choosing appropriate step sizes is critical for reducing the computational cost of training large-scale neural network models. Mini-batch sub-sampling (MBSS) is often employed for computational tractability. However, MBSS introduces a sampling error, that can manifest as a bias or variance in a line search. This is bec…
SGD batch size affects autoencoder global minima sparsity and sharpness.
problem Investigating how batch size impacts autoencoder learning.
method Non-convex autoencoder training with SGD, varying batch sizes.
result SGD batch size influences global minimum sparsity and sharpness.
New bounds for KANs trained with DP-SGD, addressing correlated noise.
problem Risk bounds for Kolmogorov-Arnold Networks trained by DP-SGD with correlated noise.
method Established new optimization and population risk analysis for KANs trained with DP-SGD, addressing correlated noise.
result First optimization and population risk analysis of correlated-noise mechanisms for DP training in non-convex settings, including neural networks.
LADIES improves GCN training efficiency and accuracy for large graphs.
problem Training large graph convolutional networks (GCNs) is computationally expensive.
method LADIES uses layer-dependent importance sampling to select nodes for training.
result LADIES outperforms previous methods in both time and memory efficiency.
New algorithm reduces overfitting in neural networks.
problem Overfitting in neural networks.
method Integrates SMC with SGHMC for mini-batch sampling.
result SMCSGHMC outperforms SGD and deep ensembles.
Improved linear regression with privacy and robustness guarantees.
problem Private and robust linear regression with adversarial corruption.
method Differentially private stochastic gradient descent with full-batch gradient descent and adaptive clipping.
result Near optimal sample complexity for both private and robust linear regression.
A higher-order Runge-Kutta optimizer performs poorly compared to Adam when evaluated fairly.
problem Evaluating the performance of adaptive Runge-Kutta optimizers under strict conditions.
method Built and evaluated a representative Adam variant using a Bogacki-Shampine 3(2) RK pair, FSAL reuse, and local-error step control.
result The adaptive nature of the RK optimizer is illusory; it behaves like a fixed-step optimizer with gradient averaging.
Bayesian model averaging fails under covariate shift, affecting neural networks' performance.
problem Bayesian model averaging's failure in neural networks under covariate shift.
method Explained the issue and proposed novel priors to improve robustness.
result Bayesian model averaging is problematic under covariate shift, especially with linear feature dependencies.
We analyze training dynamics in Gaussian mixture models using a comparison theorem.
problem Analyzing training algorithms with Gaussian mixture data.
method Applying a Gaussian comparison theorem to a specific family of training algorithms.
result Validated dynamic mean-field expressions and provided iterative refinement schemes.
BLISS optimizes GNN training by adaptively sampling nodes.
problem High computational costs in training GNNs on large graphs.
method Uses Bandit Layer Importance Sampling to dynamically select nodes.
result Improves GNN performance with reduced computational cost.
Gradient descent on neural nets often operates at the Edge of Stability, where loss behavior is complex but loss decreases over time.
problem Understanding the optimization dynamics of neural networks at the Edge of Stability.
method Empirical demonstration of gradient descent behavior in neural network training.
result Gradient descent on neural networks typically occurs at the Edge of Stability, where loss behavior is non-monotonic but loss decreases over time.
Nowadays stochastic approximation methods are one of the major research direction to deal with the large-scale machine learning problems. From stochastic first order methods, now the focus is shifting to stochastic second order methods due to their faster convergence and availability of computing resources. In this pap…
Posterior refinement improves sample efficiency in Bayesian neural networks.
problem Bayesian neural networks suffer from poor predictive performance due to inaccurate posterior approximations.
method Propose refining Gaussian approximate posteriors with normalizing flows to improve predictive distributions.
result Posterior refinement yields competitive predictive performance with minimal computational overhead.
Analysis of SGD for Gaussian mixture classification using dynamical mean-field theory.
problem Learning dynamics of SGD for a neural network classifying Gaussian mixture.
method Applying dynamical mean-field theory to track SGD dynamics in high dimensions.
result Reveals how SGD navigates the non-convex loss landscape.
Study on MC dropout in wide neural networks and its convergence to Gaussian processes.
problem Understanding the behavior of Monte Carlo dropout in wide neural networks.
method Rigorously studied the limiting distribution of wide untrained NNs under dropout, proving convergence to Gaussian processes. Investigated correlations and non-Gaussian behavior in finite width NNs.
result Wide untrained neural networks under dropout converge to Gaussian processes for fixed sets of weights and biases.
The paper analyzes early stopping in linear regression and shows it's equivalent to ridge regularization.
problem Understanding the effect of early stopping on linear regression models.
method Characterization of gradient descent dynamics and analysis of excess risk.
result Early stopped solution is equivalent to minimum norm solution for a generalized ridge regularized problem.
Paper proposes a KGE framework that reduces training time and carbon footprint.
problem Efficient KGE learning with reduced computational cost and environmental impact.
method Full batch learning, Orthogonal Procrustes Analysis, non-negative-sampling training.
result Significant reduction in training time and carbon footprint compared to state-of-the-art approaches.
AMAGOLD improves stochastic gradient MCMC by infrequent Metropolis-Hastings corrections.
problem Bias in stochastic gradient Hamiltonian Monte Carlo (SGHMC).
method AMAGOLD infrequently uses Metropolis-Hastings corrections to remove bias, with a fixed step size schedule.
result AMAGOLD converges to the target distribution with a fixed, rather than a diminishing, step size, and at most a constant factor slower convergence rate.
New algorithms optimize risk for large datasets, improving efficiency.
problem Optimizing risk for large datasets with robust methods.
method Proposed algorithms for distributionally robust optimization with CVaR and χ² divergence uncertainty sets.
result Algorithms require independent gradient evaluations of training set size and parameters, suitable for large-scale applications.
Revisits online Laplace methods for neural networks, showing they are sound under certain conditions.
problem Online Laplace methods violate the Laplace approximation's critical assumption.
method Re-derives online Laplace methods, showing they target a variational bound on a mode-corrected variant of the Laplace evidence.
result Online Laplace and its mode-corrected counterpart share stationary points that satisfy the Laplace method's assumption.
In practice it is often found that large over-parameterized neural networks generalize better than their smaller counterparts, an observation that appears to conflict with classical notions of function complexity, which typically favor smaller models. In this work, we investigate this tension between complexity and gen…
Momentum affects optimization differently at small vs large batch sizes near instability.
problem Understanding how momentum impacts optimization near the edge of stability.
method Demonstrated through batch-size dependent behavior of SGD with momentum.
result Momentum operates in two distinct regimes: amplifying stochastic fluctuations at small batch sizes and stabilizing at large batch sizes.
This paper examines the convergence of adaptive sampling methods for Bayesian neural networks.
problem Uncertainty quantification in deep neural networks, especially for medical applications.
method Locally adaptive and scalable diffusion-based sampling methods.
result These methods can have a substantial bias in the distribution they sample, even in the limit of vanishing step sizes.
Bayesian Predictive Coding improves deep learning uncertainty quantification.
problem Limitations of maximum a posteriori and maximum likelihood estimates in predictive coding.
method Developed Bayesian Predictive Coding (BPC) that estimates a posterior distribution over network parameters.
result BPC offers comparable uncertainty quantification to existing methods in Bayesian deep learning and improves convergence properties.
A new model uses attention and Gaussian processes for efficient time-series generation.
problem Computational inefficiency and uncertainty underestimation in sequence transduction.
method Attention-based Gaussian process network for real-valued sequence generation.
result The model improves training efficiency and learns factorized generative distribution.
SGD with large learning rates can achieve better test accuracy than expected.
problem SGD with large learning rates often outperforms expected convergence bounds.
method Proved that SGD with small learning rates stays close to gradient flow path on modified loss.
result Explicitly adding an implicit regularizer to the loss improves test accuracy.
Study shows how mini-batch GD with random reshuffling affects least squares regression dynamics.
problem Analyzing the error dynamics of mini-batch GD with random reshuffling for least squares regression.
method Represented training and generalization errors through a sample cross-covariance matrix Z, compared with sample covariance matrix of original features X, and used linear scaling rule for analysis.
result Mini-batch GD with random reshuffling exhibits subtle step-size dependence not detectable by gradient flow analysis, converging to a limit dependent on the step size.
DD algorithm tracks test error from train error without validation data.
problem Systematic generalization gap between train and test errors in modern model training.
method Decoupled descent (DD) algorithm that cancels data reuse biases via approximate message passing.
result DD algorithm rigorously demonstrates zero-cost validation and 100% data utilization.
AdamW optimizes a constrained loss with ℓ∞ norm constraint.
problem Understanding the optimization behavior of AdamW with ℓ∞ norm constraint. method Analyzing AdamW as a smoothed version of SignGD and connecting it to Frank-Wolfe optimization.
result AdamW implicitly performs constrained optimization with ℓ∞ norm constraint. New analysis for black-box learning without gradients, improving generalization bounds.
problem Generalization error analysis for derivative-free optimization.
method Zeroth-order Stochastic Search (ZoSS) algorithm for Lipschitz and smooth losses.
result Generalization bounds independent of model dimension, batch size, and number of perturbed evaluations.
Expectation-Maximization (EM) is a prominent approach for parameter estimation of hidden (aka latent) variable models. Given the full batch of data, EM forms an upper-bound of the negative log-likelihood of the model at each iteration and updates to the minimizer of this upper-bound. We first provide a "model level" in…
Neural networks fit fewer samples than their parameters suggest in practice.
problem Understanding the practical limitations of neural network flexibility.
method Examination of neural network optimization, parameter efficiency, and loss surfaces.
result Neural networks can only fit training sets with significantly fewer samples than their parameters suggest.
Fine-grained analysis of gradient descent with momentum provides modified loss equations.
problem Understanding the dynamics of gradient descent with momentum.
method Fine-grained analysis and derivation of modified loss equations.
result Global approximation bounds and continuous modified equations for HB.
SAM improves neural network generalization by penalizing sharpness, clarifying its exact notion and mechanism.
problem Improving deep neural network generalization for various settings.
method Sharpness-Aware Minimization (SAM) technique that penalizes a notion of sharpness of the model.
result SAM regularizes the third notion of sharpness, most likely preferred for practical performance.
Gradient descent forces neural network eigenvalues to a specific threshold.
problem Understanding why gradient descent drives eigenvalues to a specific threshold.
method Introduced edge coupling, a functional on consecutive iterate pairs, to explain the trajectory towards the eigenvalue threshold.
result Gradient descent forces the Hessian eigenvalue to the threshold 2/η from arbitrary initialization.