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.
Seesaw optimizes training by balancing learning rate and batch size, accelerating model pretraining.
problem Optimizing training efficiency for large language models with adaptive optimizers.
method Develops a principled framework for batch-size scheduling, introducing Seesaw which multiplies learning rate by 1/√2 and doubles batch size.
result Empirically, Seesaw reduces wall-clock time by approximately 36% compared to cosine decay, matching theoretical limits.
New framework optimizes deep learning training by deferring large batch sizes to late stages.
problem Optimizing batch size scheduling for deep learning training efficiency.
method Introduced the functional scaling law (FSL) framework to analyze and optimize batch size scheduling.
result Large batch sizes can be deferred to late training stages without sacrificing performance.
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.
New convergence results for NGVI with various step sizes and sample sizes.
problem Understanding convergence of stochastic NGVI for various schedules.
method Projected stochastic NGVI for exponential family variational distributions.
result Geometric convergence and $\mathcal{O}\left(\frac{1}{T^ρ}
ight)$ rates for different schedules.
We investigate the dynamical and convergent properties of stochastic gradient descent (SGD) applied to Deep Neural Networks (DNNs). Characterizing the relation between learning rate, batch size and the properties of the final minima, such as width or generalization, remains an open question. In order to tackle this pro…
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.
Mini-batch stochastic gradient descent and variants thereof have become standard for large-scale empirical risk minimization like the training of neural networks. These methods are usually used with a constant batch size chosen by simple empirical inspection. The batch size significantly influences the behavior of the …
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.
With an increasing demand for training powers for deep learning algorithms and the rapid growth of computation resources in data centers, it is desirable to dynamically schedule different distributed deep learning tasks to maximize resource utilization and reduce cost. In this process, different tasks may receive varyi…
We determine the critical batch size for large language models and find it scales with data size, not model size.
problem Determining the optimal batch size for large-scale model training.
method We propose a measure of critical batch size, pre-trained models, and systematic hyper-parameter sweeps.
result The critical batch size scales primarily with data size, not model size.
A major challenge in current optimization research for deep learning is to automatically find optimal step sizes for each update step. The optimal step size is closely related to the shape of the loss in the update step direction. However, this shape has not yet been examined in detail. This work shows empirically that…
Noise in SGD helps deep nets generalize better, even with smaller batch sizes.
problem The generalization benefit of using noise in SGD over large batch sizes.
method Carefully designed experiments and rigorous hyperparameter sweeps on various models.
result Small or moderately large batch sizes outperform very large batches on test sets.
It is common practice to decay the learning rate. Here we show one can usually obtain the same learning curve on both training and test sets by instead increasing the batch size during training. This procedure is successful for stochastic gradient descent (SGD), SGD with momentum, Nesterov momentum, and Adam. It reache…
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.
Bayesian method optimizes rescheduling for multipurpose batch processes with incomplete look-ahead information.
problem Optimizing rescheduling for multipurpose batch processes under incomplete look-ahead information.
method Proposes a Bayesian dynamic scheduling method that learns from disturbances and updates schedules online.
result Achieves statistically better long-term costs and system nervousness compared to existing periodic rescheduling strategies.
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.
New method shows how order of gradient updates impacts stability and convergence in deep learning.
problem Training deep learning models can be unstable and computationally expensive.
method Theoretical analysis and experiments with backward-SGD.
result The order of gradient updates affects stability and convergence, leading to improved performance.
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.
Recent work has shown how to train Convolutional Neural Networks (CNNs) rapidly on large image datasets, then transfer the knowledge gained from these models to a variety of tasks. Following [Radford 2017], in this work, we demonstrate similar scalability and transfer for Recurrent Neural Networks (RNNs) for Natural La…
The paper optimizes LLM inference systems through queueing theory.
problem Efficient LLM inference for AI agents under various routing topologies.
method Developed a fluid-limit framework for multi-class batched processing networks under K-FCFS scheduling.
result Proved that work-conserving scheduling algorithms maximize throughput for LLM inference.
Step sizes in neural network training are largely determined using predetermined rules such as fixed learning rates and learning rate schedules. These require user input or expensive global optimization strategies to determine their functional form and associated hyperparameters. Line searches are capable of adaptively…
Accelerates BERT pretraining from 3 days to 54 minutes.
problem Long training time of BERT due to large mini-batch sizes.
method LANS method and learning rate scheduler for large mini-batch training.
result Achieved fastest BERT training time of 54 minutes.
This paper introduces sample-averaged Q-learning for better RL performance.
problem Improving reinforcement learning algorithms by managing uncertainty.
method Integrates statistical inference into Q-learning through sample averaging and functional central limit theorem.
result Establishes a unified theoretical foundation for sample-averaged Q-learning.
Training deep neural networks with Stochastic Gradient Descent, or its variants, requires careful choice of both learning rate and batch size. While smaller batch sizes generally converge in fewer training epochs, larger batch sizes offer more parallelism and hence better computational efficiency. We have developed a n…
A new scaling law predicts optimal batch size for training models.
problem Finding the optimal batch size for training models efficiently.
method Proposed a three-term scaling law that considers model size, training data, training steps, and batch size.
result The three-term law accurately recovers the optimal batch size and can be robustly fit with fewer training runs.
Determining the appropriate batch size for mini-batch gradient descent is always time consuming as it often relies on grid search. This paper considers a resizable mini-batch gradient descent (RMGD) algorithm based on a multi-armed bandit for achieving best performance in grid search by selecting an appropriate batch s…
Study finds optimal learning rate schedules for sub-100M quantization-aware training across bit-widths.
problem Optimal learning rate schedules for quantization-aware training depend on bit-width.
method Factorial grid testing over bit-width, warmdown fraction, LR magnitude, model size, and seed.
result INT6 QAT requires a different schedule than higher-precision training, falsifying the primary hypothesis.
Large-batch SGD is important for scaling training of deep neural networks. However, without fine-tuning hyperparameter schedules, the generalization of the model may be hampered. We propose to use batch augmentation: replicating instances of samples within the same batch with different data augmentations. Batch augment…
Riemannian stochastic gradient descent converges faster with increasing batch size.
problem Improving convergence rate of Riemannian stochastic gradient descent.
method Theoretical analysis and numerical investigation of increasing batch size effects.
result Riemannian stochastic gradient descent converges faster with increasing batch size.
A new method SEBS optimizes SGD batch size for better performance.
problem Optimizing batch size for SGD to balance training speed and generalization.
method SEBS method uses a multi-stage geometric batch size enlargement scheme.
result SEBS reduces parameter updates without increasing generalization error.
Large batch sizes reduce gradient variance in DP-SGD, improving privacy.
problem Understanding why large batch sizes work in DP-SGD.
method Decomposed total gradient variance into subsampling and noise-induced variances, proving batch size independence in the limit.
result Large batch sizes reduce effective total gradient variance, improving privacy in DP-SGD.
SGD's performance improves with critical batch size, minimizing SFO complexity.
problem Optimizing SGD's performance with batch size and learning rate.
method Analysis of SGD using constant and decaying learning rates, focusing on batch size effects.
result SGD with critical batch size minimizes SFO complexity.
Proposes an automatic cyclical scheduling for gradient-based discrete sampling.
problem Gradient-based sampling in high-dimensional models can get stuck in local modes.
method Cyclical step size and balancing schedules with automatic hyperparameter tuning.
result Proves non-asymptotic convergence and inference guarantees for general discrete distributions.
CWGD measures gradient diversity weighted by curvature, improving SGD convergence.
problem Gradient noise in high-curvature directions is underestimated by standard methods.
method CWGD weights gradient diversity by the inverse square root of the Hessian.
result CWGD-Cosine reduces optimization error by up to 20% compared to standard cosine annealing.
Adaptive batch sizes improve active learning efficiency and flexibility.
problem Fixed batch sizes in active learning are inefficient due to dynamic cost-speed trade-offs.
method Probabilistic Numerics framework that adaptively changes batch sizes based on integration error and precision objectives.
result Significant enhancement in learning efficiency and flexibility across various applications.
When using active learning, smaller batch sizes are typically more efficient from a learning efficiency perspective. However, in practice due to speed and human annotator considerations, the use of larger batch sizes is necessary. While past work has shown that larger batch sizes decrease learning efficiency from a lea…
Adjusting the learning rate schedule in stochastic gradient methods is an important unresolved problem which requires tuning in practice. If certain parameters of the loss function such as smoothness or strong convexity constants are known, theoretical learning rate schedules can be applied. However, in practice, such …
We make policy optimization algorithms batch size-invariant by decoupling proximal and behavior policies.
problem Some policy optimization algorithms do not have batch size-invariance, leading to inefficiencies.
method We decouple the proximal policy from the behavior policy to achieve batch size-invariance.
result Our approach makes policy optimization algorithms more efficient and allows them to use stale data more effectively.
Current state-of-the-art NMT systems use large neural networks that are not only slow to train, but also often require many heuristics and optimization tricks, such as specialized learning rate schedules and large batch sizes. This is undesirable as it requires extensive hyperparameter tuning. In this paper, we propose…
AdAdaGrad optimizes batch sizes for deep learning models, reducing the generalization gap.
problem The generalization gap between large-batch and small-batch training in deep learning.
method AdAdaGrad introduces adaptive batch size strategies derived from adaptive sampling methods.
result AdAdaGradNorm converges to a first-order stationary point with a rate of O(1/K) in K iterations.
Empirical study on SGD hyperparameters and adversarial robustness.
problem Effect of SGD hyperparameters on adversarial robustness and generalization.
method Empirical observation of learning rate, batch size, and momentum effects on adversarial robustness and generalization.
result Constant learning rate to batch size ratio leads to good generalization and almost constant adversarial robustness.
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.
Batch normalisation doesn't affect variational inference but fails for larger batch sizes.
problem Failure of Monte Carlo Batch Normalisation (MCBN) for capturing epistemic uncertainty in larger batch sizes.
method Investigated MCBN as an approximate inference technique for Bayesian neural networks, showing its limitations and providing insights for improvement.
result For larger batch sizes, MCBN fails to capture epistemic uncertainty, requiring the batch size to be a variational parameter.
Adaptive batch sizes improve local gradient methods in distributed training.
problem Communication bottlenecks in distributed deep learning.
method Adaptive batch size strategies for local gradient methods.
result Adaptive batch sizes reduce minibatch gradient variance and improve training efficiency.
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.
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.
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.