This paper explains why Adam generalizes worse than SGD by analyzing its components.
problem Understanding why Adam generalizes worse than Stochastic Gradient Descent (SGD).
method Diffusion theoretical framework to disentangle the effects of Adaptive Learning Rate and Momentum.
result Adaptive Learning Rate helps escape saddle points but not select flat minima, while Momentum provides a drift effect to help pass through saddle points.
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.
Untuned SGD converges but with an exponential dependence on smoothness, adaptive methods prevent this.
problem The exponential dependence on smoothness in untuned SGD's convergence rate.
method Untuned SGD with arbitrary stepsize η, adaptive methods like NSGD, AMSGrad, and AdaGrad.
result Adaptive methods prevent the exponential dependence on smoothness in SGD.
Adaptive step-size method improves compressed SGD performance in machine learning.
problem Communication bottleneck in distributed and decentralized optimization.
method Developed an adaptive step-size method for compressed SGD.
result Order-optimal convergence rates for various objective functions.
This paper analyzes convergence of DP-SGD with adaptive quantile clipping.
problem Empirical success of adaptive clipping methods lacks theoretical understanding.
method Comprehensive convergence analysis of SGD with quantile clipping (QC-SGD).
result Establishes theoretical guarantees for DP-QC-SGD, revealing relationships between quantile selection, step size, and convergence.
New insights into SGD and SGD-M in high dimensions.
problem Understanding and comparing SGD and SGD-M in high-dimensional settings.
method Developed high-dimensional scaling limits for SGD-M and online SGD, examining their dynamics and performance.
result SGD-M amplifies high-dimensional effects, potentially degrading performance compared to online SGD.
Adaptive distributed SGD reduces delay in slow workers.
problem Minimizing delay in distributed SGD with stragglers.
method Adaptive policy for varying k k k to optimize error-runtime trade-off. result Numerical simulations confirm the effectiveness of the adaptive approach.
Introduces a new stochastic optimization method for deep learning.
problem Minimizing loss functions in deep neural networks.
method Introduces a second-order stochastic Runge-Kutta method and an adaptive SGD-G2.
result The method yields consistent minimization of loss functions and automatically adjusts learning rates.
LAGS-SGD optimizes deep learning training by sparsifying gradients layer-wise.
problem Reduces long training times in large deep neural networks with distributed S-SGD.
method Layer-wise adaptive gradient sparsification combined with S-SGD.
result LAGS-SGD achieves convergence guarantees and outperforms vanilla S-SGD.
Paper analyzes high probability convergence of adaptive SGD with momentum.
problem Theoretical understanding of adaptive SGD with momentum in nonconvex settings is incomplete.
method High probability analysis under weak assumptions.
result First high probability convergence proof for gradients to zero in Delayed AdaGrad with momentum.
Adaptive quantization improves SGD accuracy in data-parallel settings.
problem Fixed gradient quantization schemes lead to suboptimal performance in deep learning.
method Developed adaptive quantization schemes ALQ and AMQ that update compression schemes based on gradient statistics.
result Improved validation accuracy on CIFAR-10 and ImageNet datasets by 2% and 1% respectively.
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.
Local AdaAlter reduces communication in SGD with adaptive learning rates.
problem Communication overhead in distributed training.
method Novel SGD variant with adaptive learning rates and reduced communication.
result Empirically reduces communication overhead by up to 30%.
Private adaptive methods improve on traditional SGD for convex optimization.
problem Differential privacy constraints in gradient optimization.
method Differentially private variants of SGD and AdaGrad with adaptive stepsizes and non-isotropic clipping.
result Private AdaGrad outperforms private SGD in high-dimensional problems.
New adaptive SGD algorithms for federated learning over physical channels.
problem Reducing communication cost in federated learning over physical channels.
method Proposed adaptive federated SGD algorithms considering channel noise and hardware constraints.
result Demonstrated convergence rates adaptive to stochastic gradient noise level.
AvaGrad optimizes vision tasks by decoupling learning rate and adaptability.
problem Improving optimization methods for vision tasks.
method Derives AvaGrad, a new optimizer that decouples learning rate and adaptability.
result AvaGrad outperforms SGD on vision tasks when adaptability is properly tuned.
Adaptive SGD learns optimal batch size for strong convex functions.
problem Finding optimal batch size for SGD in practice.
method Adaptive SGD method that learns optimal batch size.
result Adaptive SGD exhibits nearly optimal performance in experiments.
IntSGD compresses SGD gradients without floats, converging as SGD.
problem Efficiently compressing stochastic gradients in distributed SGD.
method Adaptive integer compression of gradients, estimating scaling adaptively.
result IntSGD matches SGD's iteration complexity for convex and non-convex functions.
Adapts SGD to noise and problem specifics for faster convergence.
problem Minimizing smooth, strongly-convex functions with varying noise and problem constants.
method Adaptive SGD with exponentially decreasing step-sizes, Nesterov acceleration, and stochastic line-search.
result Achieves near-optimal convergence rates without knowing noise or problem specifics.
New method trains deep networks robustly without adaptive methods.
problem Training deep networks with robustness and efficiency.
method Scale invariant architecture + SGD + weight decay + gradient clipping.
result SGD can achieve similar performance to adaptive methods like Adam.
Proposes a new adaptive learning rate for SGD.
problem Finding an efficient learning rate for SGD.
method Stochastic Polyak step-size (SPS).
result SGD with SPS converges faster for over-parameterized models.
AdaSGD combines SGD and Adam benefits, eliminating the need for transition.
problem Understanding when to transition from Adam to SGD for optimal performance.
method Adapting a single global learning rate for SGD (AdaSGD).
result AdaSGD combines the benefits of both SGD and Adam, improving convergence and generalization.
ARFF reduces spectral bias in SGD-trained neural networks.
problem Spectral bias in two-layer neural networks.
method Comparison of SGD and ARFF on spectral bias and robustness.
result ARFF yields a closer to zero spectral bias compared to SGD.
Adaptive optimization methods, which perform local optimization with a metric constructed from the history of iterates, are becoming increasingly popular for training deep neural networks. Examples include AdaGrad, RMSProp, and Adam. We show that for simple overparameterized problems, adaptive methods often find drasti…
Improved SGD methods converge faster for nonconvex optimization.
problem Nonconvex optimization challenges in machine learning.
method Adaptive SGD with line-search and Polyak stepsizes.
result Unified convergence rates for various nonconvex functions.
Develops a parameter-free SGD algorithm with optimal convergence rate.
problem Optimizing parameters in stochastic convex optimization.
method A novel parameter-free algorithm for SGD with high-probability guarantees and adaptive properties.
result Achieves optimal convergence rate with only a double-logarithmic factor increase compared to known-parameter settings.
Improves distributed SGD convergence speed with reduced computation load.
problem Mitigating stragglers in distributed SGD to speed up convergence.
method Modeling communication and computation times, adapting number of workers and computation load dynamically.
result Significantly reduces computation load while improving convergence speed.
End-to-end analysis of SGD for STL with adaptive sub-sampling.
problem Designing SGD for STL with statistical guarantees without prior knowledge of source quality.
method Mixed-sample SGD procedure that alternates between source and target data, maintaining transfer guarantees.
result Mixed-sample SGD converges to a target-adaptive solution with 1 / T 1/\sqrt{T} 1/ T rate. Adaptive gradient methods, which adopt historical gradient information to automatically adjust the learning rate, despite the nice property of fast convergence, have been observed to generalize worse than stochastic gradient descent (SGD) with momentum in training deep neural networks. This leaves how to close the gene…
Adaptive-SGD method optimizes machine learning training with dynamic batch and step sizes.
problem Optimizing machine learning training with adaptive batch and step sizes.
method Adaptive-SGD method that dynamically adjusts batch size and step size based on local curvature and probability of descent directions.
result Adaptive-SGD achieves global linear convergence on self-concordant functions and compares favorably to fine-tuned methods.
Stochastic Gradient Descent (SGD) is one of the most widely used techniques for online optimization in machine learning. In this work, we accelerate SGD by adaptively learning how to sample the most useful training examples at each time step. First, we show that SGD can be used to learn the best possible sampling distr…
Unified analysis for decentralized SGD across various topologies and updates.
problem Analysis of decentralized SGD methods with changing topologies and local updates.
method Unified convergence analysis covering local SGD updates and adaptive network topology.
result Universal convergence rates for smooth problems, interpolating between heterogeneous and iid-data settings.
Classical stochastic gradient methods for optimization rely on noisy gradient approximations that become progressively less accurate as iterates approach a solution. The large noise and small signal in the resulting gradients makes it difficult to use them for adaptive stepsize selection and automatic stopping. We prop…
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.
This paper analyzes SGD with biased gradients for deep learning models.
problem Analyzing SGD with biased gradients for deep learning models.
method Non-asymptotic analysis of SGD with adaptive steps for non-convex smooth functions.
result Adagrad, RMSProp, and AMSGRAD converge to critical points at a similar rate to unbiased case.
Privacy preserving machine learning algorithms are crucial for learning models over user data to protect sensitive information. Motivated by this, differentially private stochastic gradient descent (SGD) algorithms for training machine learning models have been proposed. At each step, these algorithms modify the gradie…
AdaScale SGD adapts learning rates for large-batch training efficiently.
problem Adapting learning rates for large-batch training to balance speed-ups and model quality.
method Adaptive learning rate adaptation based on gradient variance.
result AdaScale achieves reliable speed-ups for a wide range of batch sizes without degrading model quality.
Improved analysis shows momentum in SGD reduces batch size needs for non-convex objectives.
problem Reducing batch size requirements for SGD in non-convex optimization.
method Normalized SGD with momentum, adaptive method for small gradient variance.
result Normalized SGD with momentum achieves ε ε ε -critical points in O ( 1 / ε 3.5 ) O(1/ε^{3.5}) O ( 1/ ε 3.5 ) iterations. AdaS adapts SGD learning rate based on knowledge gain metrics.
problem Empirical step-size selection in SGD optimization lacks consistency and insight.
method Introduces AdaS algorithm that adapts SGD learning rate based on knowledge gain metrics.
result AdaS outperforms existing adaptive learning methods in convergence and generalization.
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…
Improved loss scaling for stochastic momentum algorithms in high dimensions.
problem Improving loss scaling for stochastic momentum algorithms in high dimensions.
method Dimension-adapted Nesterov acceleration (DANA) scales momentum hyperparameters based on model size and data complexity.
result DANA improves loss scaling exponents across various data and target complexities.
CSER improves SGD efficiency by resetting errors and partial synchronization.
problem Limited scalability of Distributed Stochastic Gradient Descent (SGD) due to communication bottlenecks.
method Introduces 'error reset' technique and partial synchronization for gradients and models.
result Proves convergence for smooth non-convex problems and accelerates distributed training significantly.
AdaGrad outperforms SGD in non-convex optimization problems by a factor of d.
problem Finding near-stationary points in stochastic non-convex optimization.
method Refined assumptions on smoothness and gradient noise variance, l 1 l_1 l 1 -norm stationarity measure. result AdaGrad achieves a convergence rate favorable over SGD in certain non-convex settings.
AdaBelief optimizes deep learning models with faster convergence and better stability.
problem Combining fast convergence and stability in deep learning models.
method Adapts stepsize based on the belief in observed gradients using exponential moving average (EMA) of noisy gradients.
result AdaBelief outperforms other methods in image classification and GAN training, achieving comparable accuracy to SGD on ImageNet.
GALA adapts learning rates online by aligning gradients, improving deep learning model performance.
problem Fine-tuning learning rates for deep learning models requires extensive grid search.
method GALA dynamically adjusts learning rates by tracking gradient alignment and local curvature.
result GALA produces a flexible, adaptive learning rate schedule that increases when gradients align.
Although stochastic gradient descent (SGD) method and its variants (e.g., stochastic momentum methods, AdaGrad) are the choice of algorithms for solving non-convex problems (especially deep learning), there still remain big gaps between the theory and the practice with many questions unresolved. For example, there is s…
Optimization algorithms affect fairness in deep learning models, especially with adaptive methods like RMSProp.
problem The impact of optimization algorithms on fairness in deep learning models, particularly under imbalance.
method Stochastic differential equation analysis of optimization dynamics in an analytically tractable setup.
result RMSProp, an adaptive optimizer, converges to fairer minima than SGD under severe imbalance.
The paper analyzes Adam and SGD in nonstationary optimization, revealing tradeoffs between noise and drift.
problem Analyzing Adam and SGD in nonstationary optimization problems.
method Theoretical analysis of Adam and SGD under non-stationary stochastic objectives, separating two regimes.
result Characterizes the tradeoff between noise and drift in Adam and SGD, revealing when adaptive step-sizing is beneficial or harmful.