Adaptive optimization algorithms such as Adam are widely used in deep learning. The stability of such algorithms is often improved with a warmup schedule for the learning rate. Motivated by the difficulty of choosing and tuning warmup schedules, recent work proposes automatic variance rectification of Adam's adaptive l…
On-device research index
arXiv research
A locally-built, LLM-digested index of recent arXiv papers in quant finance, geometry/topology, and statistical ML — keyword search served straight from SQLite on this machine.
168,657 papers · 148 categories
Trend · papers per month
38 results for “warmup”
The paper shows how warming up the learning rate improves deep learning performance.
problem Improving deep learning performance through better handling of larger learning rates.
method Systematic experiments with SGD and Adam showing the benefits of warmup and different regimes of operation.
result Properly choosing can eliminate the need for warmup and improve performance.
Paper proposes faster certified robust training methods with short warmup.
problem Certified robust training methods require long warmup schedules, making training costly.
method Proposes three improvements: new weight initialization, BN, and regularization.
result Achieves 65.03% verified error on CIFAR-10 and 82.36% on TinyImageNet with short warmup.
The learning rate warmup heuristic achieves remarkable success in stabilizing training, accelerating convergence and improving generalization for adaptive stochastic optimization algorithms like RMSprop and Adam. Here, we study its mechanism in details. Pursuing the theory behind warmup, we identify a problem of the ad…
Training deep neural networks using a large batch size has shown promising results and benefits many real-world applications. However, the optimizer converges slowly at early epochs and there is a gap between large-batch deep learning optimization heuristics and theoretical underpinnings. In this paper, we propose a no…
The convergence rate and final performance of common deep learning models have significantly benefited from heuristics such as learning rate schedules, knowledge distillation, skip connections, and normalization layers. In the absence of theoretical underpinnings, controlled experiments aimed at explaining these strate…
Optimal Learning-Rate Schedules under Functional Scaling Laws: Power Decay and Warmup-Stable-Decaystat.ML
Optimal learning rates decay to zero in easy tasks and maintain a warmup phase in hard tasks.
problem Optimizing learning rates under functional scaling laws for model training.
method Deriving optimal learning-rate schedules based on exponents and .
result Sharp phase transition between easy and hard tasks, with different decay behaviors.
GradPower speeds up language model training with minimal code changes.
problem Slower training of large language models.
method Elementwise sign-power transformation applied to gradients.
result Consistently lower terminal loss across various models and datasets.
SignSGD outperforms SGD in linear regression with optimal scaling laws under PLRF model.
problem Improving linear regression performance with signSGD under power-law random features.
method Analysis of signSGD risk under PLRF model, comparison with SGD, identification of unique effects.
result SignSGD can have a steeper compute-optimal slope than SGD in noisy regimes, especially with WSD schedule.
WSqD extends learning rate schedules for large model training without fixed horizons.
problem Fixed learning rate schedules limit training horizon extension.
method WSqD replaces constant stable phase with a shifted inverse-square-root base, retaining linear cooldown.
result WSqD achieves minimax-optimal convergence rate and horizon-independence.
This work improves regret minimization for logistic bandits by reducing dependence on a large constant.
problem Minimizing regret in logistic bandits with reduced dependence on a large constant.
method Experimental design procedure and warmup sampling algorithm.
result Achieves a minimax regret of \(O(\sqrt{d \dotμT\log(|\mathcal{X}|)})\) in the fixed arm setting.
New method prevents entropy collapse in Transformer training, leading to more stable and robust models.
problem Training instability in Transformers, especially in attention layers.
method Spectral normalization with a learned scalar to prevent entropy collapse.
result Prevents entropy collapse, leading to more stable training.
Adaptive batch size schedules improve language model training efficiency and generalization.
problem Dilemma of choosing batch sizes in large-scale model training.
method General-purpose adaptive batch size schedules compatible with data and model parallelism.
result Adaptive batch size schedules outperform constant batch sizes and heuristic warmup schedules.
nGPT learns to transfer learning rates across model dimensions and token horizons.
problem nGPT does not transfer learning rates across model size and token horizon.
method Combining numerical experiments with alignment exponents, a novel nGPT parameterization νGPT is developed.
result νGPT exhibits learning rate transfer across width, depth, and token horizon.
WSD schedule improves model training efficiency by adapting learning rates dynamically.
problem Fixed compute budgets limit training efficiency of language models.
method Introduces a WSD schedule that uses a constant learning rate followed by a rapid decay phase.
result WSD schedule generates a non-traditional loss curve with stable and decay phases.
We evaluate three simple, normalization-centric changes to improve Transformer training. First, we show that pre-norm residual connections (PreNorm) and smaller initializations enable warmup-free, validation-based training with large learning rates. Second, we propose normalization with a single scale paramete…
ScheduleFree+ improves large language model training without schedules or learning rates.
problem Scaling up Schedule-Free Learning to large language models.
method Learning-rate-free and schedule-free method for training large language models.
result ScheduleFree+ outperforms SOTA schedules by 31% at 1000 tokens per parameter.
The paper explores how symmetries and noise in SGD influence parameter dynamics.
problem Understanding the dynamics of parameter updates in SGD with symmetries.
method Proved the existence of noise equilibria and showed their role in balancing gradient noise.
result Gradient noise creates a systematic motion of parameters to a unique fixed point, called noise equilibria.
Our work presents extensive empirical evidence that layer rotation, i.e. the evolution across training of the cosine distance between each layer's weight vector and its initialization, constitutes an impressively consistent indicator of generalization performance. In particular, larger cosine distances between final an…
Random Function Descentmath.OC
Random Function Descent improves optimization in high dimensions.
problem Lack of effective optimization methods in high-dimensional spaces.
method Introducing a 'random function' framework to optimize classical optimization problems.
result Random Function Descent (RFD) is a scalable optimization method that bridges Bayesian and classical optimization.
Robust Transformer-Based One-Step Stock Index Forecasting via Shifted Data Augmentation
problem Robust stock index forecasting
method Modified Transformer architecture with Shifted Data Augmentation
result Best performance on benchmark datasets
Residual networks (ResNet) and weight normalization play an important role in various deep learning applications. However, parameter initialization strategies have not been studied previously for weight normalized networks and, in practice, initialization methods designed for un-normalized networks are used as a proxy.…
Large-batch training approaches have enabled researchers to utilize large-scale distributed processing and greatly accelerate deep-neural net (DNN) training. For example, by scaling the batch size from 256 to 32K, researchers have been able to reduce the training time of ResNet50 on ImageNet from 29 hours to 2.2 minute…
This paper explains why distributional reinforcement learning is better than vanilla RL using small-loss bounds.
problem Understanding when and why distributional reinforcement learning (DistRL) is superior to vanilla reinforcement learning (RL).
method The paper uses small-loss bounds to explain the benefits of DistRL, proposing algorithms and proving bounds for different RL settings.
result Distributional reinforcement learning (DistRL) outperforms vanilla RL when optimal costs are small, as shown by small-loss bounds.
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.
ATLAS adapts HMC step size and trajectory length for complex geometries.
problem Sampling complex geometries with constant step size HMC/NUTS.
method Adapts step size and trajectory length using local Hessian and no U-turn condition.
result ATLAS accurately samples complex geometries, outperforming NUTS.
Through the River: Understanding the Benefit of Schedule-Free Methods for Language Model Trainingcs.LG
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.
The large asymptotics (perturbation series) for integrals of the form , where is a smooth top form and is a smooth function on a manifold , both of which are invariant under the action of a symmetry group , may be computed using the stationary phase approximation…
SMC analysis reveals key transient effects in macroeconomic ABM.
problem Analysis of complex ABMs is challenging and often relies on ad hoc methods.
method Statistical model checking (SMC) implemented through MultiVeStA.
result Clear contrast across parameter families in macro-financial and structural sweeps.
The paper presents a multi-power law for predicting loss curves across different learning rate schedules.
problem Understanding and optimizing the relationship between model performance and hyperparameters, especially learning rates.
method Proposes a multi-power law that combines power laws based on the sum of learning rates and additional laws for loss reduction due to decay.
result The multi-power law accurately predicts loss curves for unseen learning rate schedules and finds a schedule that outperforms cosine learning rate.
We discover scaling laws for kernel regression loss under various learning rate schedules.
problem Understanding loss dynamics and learning rate schedules in kernel regression.
method Theoretical analysis of stochastic gradient descent on a power-law kernel regression model.
result Established a Functional Scaling Law (FSL) capturing the full loss trajectory under arbitrary learning rate schedules.
Transformers use a unique Hessian structure that differs from classical networks, affecting optimization.
problem Understanding the unique optimization landscape of Transformers.
method Theoretical Hessian analysis of a single self-attention layer in Transformers.
result Transformers have a highly non-linear Hessian structure, distinguishing them from classical networks.
A new method reduces complexity of normalizing flows for MCMC preconditioning.
problem Improving sampling efficiency in MCMC algorithms for complex target distributions.
method Factorized preconditioning architecture combining a linear component and a conditional NF.
result Significantly better tail samples and higher effective sample sizes on various distributions.
PolarGrad optimizes deep learning models by considering matrix structure, outperforming Adam and Muon.
problem Efficient optimization of large-scale neural networks and language models.
method A unifying framework for analyzing matrix-aware preconditioned methods, including PolarGrad.
result PolarGrad outperforms Adam and Muon in various tasks.
Injective and surjective neural operators for function spaces.
problem Tackles injective and surjective neural operators in function spaces.
method Combines prior work in ReLU and operator learning, uses Fredholm theory and Leray-Schauder degree theory.
result Injective and surjective neural operators are universal approximators and maintain their properties in finite-rank implementations.
This paper simplifies fine-tuning for small LLMs, reducing barriers for developers.
problem Limited resources for fine-tuning large language models (LLMs) by individual developers and small organizations.
method Instruction-tuning datasets, small-sized LLMs (3B to 7B parameters), various training configurations and strategies.
result Improved model performance on benchmarks with specific training configurations, and insights into early termination and hyperparameter simplifications.
Theory of Optimal Learning Rate Schedules and Scaling Laws for a Random Feature Modelcond-mat.dis-nn
We find optimal learning rate schedules for a random feature model.
problem Choosing optimal learning rates for deep learning models.
method We analyze a powerlaw random feature model trained with SGD, considering optimal schedules as numerical and analytical problems.
result We discover two regimes: easy and hard phases, with different optimal learning rate schedules.
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.