Is AdamW effective under heavy-tailed noise?
problem Stochastic gradient noise in LLM pretraining is typically heavy-tailed.
method Formulate as an open problem, prove a positive weighted-metric benchmark, and give a corridor lower-bound mechanism.
result No rigorous convergence theory for AdamW established in heavy-tailed regime.
AdamW optimizes a constrained loss with ℓ ∞ \ell_\infty ℓ ∞ norm constraint.
problem Understanding the optimization behavior of AdamW with ℓ ∞ \ell_\infty 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 ∞ \ell_\infty ℓ ∞ norm constraint. MARS optimizes large model training by reducing variance, outperforming AdamW.
problem Training large models efficiently and scalably.
method Unified optimization framework MARS combining preconditioned gradient updates and variance reduction.
result MARS outperforms AdamW in training GPT-2 models.
Uniform scaling limits in AdamW-trained transformers converge to ODEs.
problem Understanding the dynamics of large-depth transformers trained with AdamW.
method Modeling transformer dynamics as an interacting particle system coupled through attention, proving convergence to ODEs.
result The joint dynamics of hidden states and backpropagated variables converge uniformly to an ODE system.
Logarithmic-time schedules boost large-scale language model training efficiency.
problem Improving performance and efficiency in large-scale language model training.
method Designing time-varying hyperparameters ( β 1 , β 2 , λ ) (β_1, β_2, λ) ( β 1 , β 2 , λ ) for AdamW, specifically logarithmic-time scheduling with damping mechanisms. result ADANA optimizer achieves up to 40% compute efficiency compared to tuned AdamW, with gains persisting as model scale increases.
New optimizer SF-NorMuon matches tuned AdamW across various horizons.
problem Fixed learning-rate schedules in neural network training lead to strong path dependence and costly re-tuning.
method Schedule-Free Spectral Optimization (SF-NorMuon)
result SF-NorMuon outperforms tuned AdamW on 125M and 772M parameter models across different horizons.
Weibull weight-scale parameter λ λ λ evolves during AdamW training, with alignment, injection, and decay forces driving its growth and relaxation.
problem Understanding the evolution of the Weibull weight-scale parameter λ λ λ during AdamW training. method Deriving a leading-order three-force decomposition of the squared weight norm from AdamW updates.
result The alignment force dominates the rise phase, contributing 88-94% of the absolute force budget across four random seeds.
New algorithms improve generalization of Adam and AdamW optimizers.
problem Adam and AdamW optimizers generalize worse than SGD.
method Algorithmic stability analysis and clever momentum-based SGD integration.
result HomeAdam(W) algorithms achieve better generalization and faster convergence.
PACE optimizes training for averaged language models, improving performance.
problem How to optimize training for averaged language model iterates.
method Formulated as an optimal-control problem, solved for minimizing error of the average with a penalty on intervention size.
result PACE improves the limiting squared error of the iterate-average estimator by an arbitrarily large factor on some instances.
IVON improves LoRA finetuning with minimal overhead and significant accuracy gains.
problem Improving LoRA finetuning with Bayesian methods.
method IVON, a variational algorithm, with posterior pruning.
result Significant accuracy improvements over AdamW and other Bayesian methods.
Paper finds sharpness differences in transformer blocks accelerating LLM training.
problem Understanding and accelerating large language model pre-training.
method Uncovering sharpness disparity across transformer blocks and proposing Blockwise Learning Rate.
result Blockwise Learning Rate strategy accelerates LLM pre-training with lower loss and speedup.
Improved IVON boosts LoRA model accuracy and calibration.
problem Improving the accuracy and calibration of large language models.
method Replaced AdamW with IVON for finetuning Llama-2.
result IVON improves Llama-2 accuracy by 2.8% and calibration error by 4.6%.
Optimizer memory affects learning rate sensitivity in shuffle order, impacting fine-tuning noise.
problem Optimizer memory affects the learning rate sensitivity in shuffle order, leading to fine-tuning noise.
method Isolated the mechanism of fixed-clock optimizer memory affecting the learning rate sensitivity in shuffle order, deriving a fit-free way to size the noise.
result Fixed-clock optimizers like AdamW produce a larger first-order noise channel compared to memoryless optimizers, affecting fine-tuning comparisons.
Improving optimization for iterate-averaged language models
problem How to optimize the averaged model returned by Language Model pipelines
method Formulating optimizer design as an optimal-control problem
result Proven convergence rate and strict improvement in squared error
Muon optimizes training efficiency by improving data retention at large batch sizes.
problem Improving training efficiency and data retention at large batch sizes.
method Introducing Muon, a second-order optimizer, and combining it with muP for efficient hyperparameter transfer.
result Muon outperforms AdamW in retaining data efficiency at large batch sizes, enabling more economical training.
New method schedules learning rate without stopping time, outperforming existing methods.
problem Learning rate schedules that require stopping time are outperformed.
method Schedule-Free approach that avoids stopping time and introduces no additional hyper-parameters.
result Exhibits state-of-the-art performance across various problems.
We introduce a new weight-decay scaling rule to maintain sublayer gains across different widths in modern scale-invariant architectures.
problem In modern scale-invariant architectures, training quickly enters a steady state where normalization layers create backward scale sensitivity, degrading learning-rate transfer.
method We introduce a weight-decay scaling rule for AdamW that preserves sublayer gain across widths by equalizing the effective learning rate.
result Our empirical weight-decay scaling rule λ 2 ∝ d λ_2\propto \sqrt{d} λ 2 ∝ d approximately keeps sublayer gains width invariant, enabling zero-shot transfer of learning rate and weight decay. 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 L 1 L_1 L 1 -regularization. result Derives a unified mathematical framework for understanding and optimizing training acceleration.
Distributed Lion optimizes large model training by reducing communication costs.
problem Training large AI models efficiently with reduced communication costs.
method Adapted Lion optimizer for distributed training, using binary or lower-precision vectors for communication.
result Distributed Lion achieves comparable performance to standard optimizers but with significantly reduced communication bandwidth.
A technique identifies memoryless algorithms approximating memory-dependent optimization methods.
problem Understanding how memory in optimization algorithms affects loss and generalization.
method Introducing a general technique to replace past iterates with the current one and adding a correction term.
result Lion does not have the same implicit anti-regularization as AdamW, explaining its better generalization performance.
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.
A new optimizer DDC improves deep learning models by respecting symmetries.
problem Deep networks' loss is invariant to continuous symmetries, leading to optimization issues.
method DDC builds a Dead-Direction Conditioner that lifts a base optimizer into a G-equivariant one, preserving the quotient geometry.
result DDCAdam and DDCMuon outperform standard optimizers in various tasks, improving validation-train loss gaps and learning dynamics.
GPA improves LLM training speed by 8.71% for Llama-160M models.
problem Training Large Language Models (LLMs) with high memory overhead and slow convergence.
method Generalized Primal Averaging (GPA) extends Nesterov's method to eliminate memory-intensive two-loop structure.
result GPA achieves up to 10.13% speedup over AdamW in training Llama-1B model.
AlgoPerf competition evaluates neural network training speed-ups.
problem Improving neural network training speed using better algorithms.
method Compared 18 diverse submissions from 10 teams on multiple workloads.
result Schedule Free AdamW algorithm achieved best results in self-tuning ruleset.
New method SF-AdamW trains large models without decay phases or memory overhead.
problem Inadequate fixed compute budgets for large-scale training.
method Schedule-Free (SF) method revisited and refined.
result SF-AdamW effectively navigates loss landscape without decay phases or memory overhead.
We propose NovoGrad, an adaptive stochastic gradient descent method with layer-wise gradient normalization and decoupled weight decay. In our experiments on neural networks for image classification, speech recognition, machine translation, and language modeling, it performs on par or better than well tuned SGD with mom…
This paper quantifies hyperparameter transfer and finds embedding layer learning rate is key.
problem Quantifying optimal hyperparameters for large language models across scales.
method Developed three metrics to quantify hyperparameter transfer and investigated the importance of embedding layer learning rate.
result Maximal Update (μP) parameterization offers high-quality learning rate transfer compared to standard parameterization (SP).
Novel Adam-family method with decoupled weight decay for training neural networks.
problem Training nonsmooth neural networks with weight decay.
method Proposes a novel Adam-family method with decoupled weight decay, establishing convergence properties and demonstrating superior performance.
result Asymptotically approximates SGD and enhances generalization performance.
New study finds optimal hyperparameter tuning crucial for fair optimizer comparisons.
problem Inadequate hyperparameter tuning and misleading evaluation setups hinder fair comparisons of optimizers.
method Systematic study of ten optimizers across four model scales and data-to-model ratios.
result Optimal hyperparameters for one optimizer may be suboptimal for another, and many claimed speedups are lower than expected.
ADAHESSIAN optimizes machine learning models with adaptive second-order methods.
problem Efficiently optimizing machine learning models with second-order methods.
method Dynamic Hessian estimation via adaptive estimates, incorporating fast approximations and moving averages.
result ADAHESSIAN achieves state-of-the-art performance across various tasks.
New optimizers control network width scaling, improving stability and transfer across different model sizes.
problem Designing stable optimizers for networks of varying widths.
method Interpreting optimizers as steepest descent under mean-normalized operator norms, enabling layerwise composability and width-independent bounds.
result New optimizers like row normalization and column normalization provide stable learning-rate transfer across different model widths.
We formulate the problem of neural network optimization as Bayesian filtering, where the observations are the backpropagated gradients. While neural network optimization has previously been studied using natural gradient methods which are closely related to Bayesian inference, they were unable to recover standard optim…
Adam optimization algorithm can have non-zero average regret under certain conditions.
problem Non-zero average regret in Adam optimization algorithm.
method Used a three-periodic sequence of linear functions on [-1,1] with slopes c, -1, -1, and analyzed Adam variants.
result Adam optimization algorithm can have non-zero average regret under certain conditions.
Cautious Weight Decay modifies weight decay for better optimization.
problem Improving optimization in deep learning models.
method Applies weight decay selectively based on parameter sign alignment.
result Consistently improves model performance across various tasks and scales.
SINGD improves KFAC for memory-efficiency and stability in low-precision training.
problem Memory inefficiency and numerical instability of KFAC in low-precision training.
method Formulated inverse-free KFAC update and imposed structures in Kronecker factors.
result SINGD is memory-efficient and numerically robust, often outperforming AdamW in half precision.
Enhances deep learning by boosting generalization and convergence.
problem Improving generalization and convergence in deep learning models.
method Implicit Regularization Enhancement (IRE) framework that decouples flat and sharp directions.
result IRE consistently improves generalization performance across various deep learning tasks and models.
Develops ECD framework for optimizing machine learning problems.
problem Optimizing machine learning models, especially non-convex ones.
method Energy Conserving Descent (ECD) framework, chaotic dynamical systems.
result ECDSep outperforms existing methods on various machine learning tasks.
RELTA-SGLD stabilizes nonconvex SGLD updates with a lighter taming scheme.
problem Stabilizing superlinear stochastic-gradient updates in nonconvex optimization.
method Threshold-based taming with relative-growth principle for stability.
result Polynomial moment stability and first-order stationary accuracy in nonconvex SGLD.
ProxSPS improves on SPS for regularization tasks, offering better stability and performance.
problem Handling regularization terms in adaptive step size schemes for stochastic gradient descent.
method Developed a proximal variant of the stochastic Polyak step size (SPS) scheme.
result ProxSPS is easier to tune and more stable with regularization, and performs well in image classification tasks.
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.
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
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.
AdamS uses momentum as a denominator to optimize LLMs efficiently.
problem Optimizing large language models (LLMs) with efficient and effective methods.
method AdamS introduces a novel denominator based on the root of the weighted sum of squares of momentum and current gradient.
result AdamS achieves superior optimization performance with minimal memory and compute requirements.
New algorithm avoids spurious sharpness minimization for NLP models.
problem SAM fails in NLP, leading to performance degradation.
method Developed Functional-SAM, which modifies logit statistics instead of function geometry.
result Functional-SAM and combined methods outperform AdamW and SAM in NLP tasks.
Extends hyperparameter transfer across model sizes and modules, improving training speed.
problem Training stability and performance of large-scale models with optimal hyperparameters.
method Complete ( d ) ^{(d)} ( d ) Parameterisation, per-module hyperparameter optimisation and transfer. result Hyperparameter transfer holds even in the per-module hyperparameter regime, improving training speed.
AdEMAMix optimizer improves model performance and convergence speed.
problem Suboptimal use of single EMA in momentum-based optimizers.
method Proposes AdEMAMix, a modified Adam optimizer with a mixture of two EMAs.
result Empirically shows gradients can remain relevant for tens of thousands of steps, leading to faster convergence and lower minima.
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.
Muon replaces matrix gradient with polar factor, optimizing flat spectrum updates
problem Optimization bias in matrix updates
method Using polar factor of gradient
result Muon update maximizes entropy among bounded updates