An official website of the United States government
We introduce two complementary techniques for efficient optimization that reduce memory requirements while accelerating training of large-scale neural networks. The first technique, Subset-Norm step size, generalizes AdaGrad-Norm and AdaGrad(-Coordinate) through step-size sharing. Subset-Norm (SN) reduces AdaGrad’s memory footprint from O(d) to O(sqrt(d)), where d is the model size. For non-convex smooth objectives under coordinate-wise sub-gaussian noise, we show a noise-adapted high-probability convergence guarantee with improved dimensional dependence of SN over existing methods. Our second technique, Subspace-Momentum, reduces the momentum state’s memory footprint by restricting momentum to a low-dimensional subspace while performing SGD in the orthogonal complement. We prove a high-probability convergence result for Subspace-Momentum under standard assumptions. Empirical evaluation on pre-training and fine-tuning LLMs demonstrates the effectiveness of our methods. For instance, combining Subset-Norm with Subspace-Momentum achieves Adam’s validation perplexity for LLaMA 1B in approximately half the training tokens (6.8B vs 13.1B) while reducing Adam’s optimizer-states memory footprint by more than 80% with minimal additional hyperparameter tuning.
more »
« less
This content will become publicly available on June 6, 2026
Provable Complexity Improvement of AdaGrad over SGD: Upper and Lower Bounds in Stochastic Non-Convex Optimization
Adaptive gradient methods, such as AdaGrad, are among the most successful optimization algorithms for neural network training. While these methods are known to achieve better dimensional dependence than stochastic gradient descent (SGD) for stochastic convex optimization under favorable geometry, the theoretical justification for their success in stochastic non-convex optimization remains elusive. In fact, under standard assumptions of Lipschitz gradients and bounded noise variance, it is known that SGD is worst-case optimal in terms of finding a near-stationary point with respect to the l2-norm, making further improvements impossible. Motivated by this limitation, we introduce refined assumptions on the smoothness structure of the objective and the gradient noise variance, which better suit the coordinate-wise nature of adaptive gradient methods. Moreover, we adopt the l1-norm of the gradient as the stationarity measure, as opposed to the standard l2-norm, to align with the coordinate-wise analysis and obtain tighter convergence guarantees for AdaGrad. Under these new assumptions and the l1-norm stationarity measure, we establish an upper bound on the convergence rate of AdaGrad and a corresponding lower bound for SGD. In particular, we identify non-convex settings in which the iteration complexity of AdaGrad is favorable over SGD and show that, for certain configurations of problem parameters, it outperforms SGD by a factor of d, where d is the problem dimension. To the best of our knowledge, this is the first result to demonstrate a provable gain of adaptive gradient methods over SGD in a non-convex setting. We also present supporting lower bounds, including one specific to AdaGrad and one applicable to general deterministic first-order methods, showing that our upper bound for AdaGrad is tight and unimprovable up to a logarithmic factor under certain conditions.
more »
« less
- Award ID(s):
- 2505865
- PAR ID:
- 10631743
- Publisher / Repository:
- https://doi.org/10.48550/arXiv.2406.04592
- Date Published:
- ISSN:
- 2406.04592
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
In this work, we describe a generic approach to show convergence with high probability for both stochastic convex and non-convex optimization with sub-Gaussian noise. In previous works for convex optimization, either the convergence is only in expectation or the bound depends on the diameter of the domain. Instead, we show high probability convergence with bounds depending on the initial distance to the optimal solution. The algorithms use step sizes analogous to the standard settings and are universal to Lipschitz functions, smooth functions, and their linear combinations. The method can be applied to the non-convex case. We demonstrate an $$O((1+\sigma^{2}\log(1/\delta))/T+\sigma/\sqrt{T})$$ convergence rate when the number of iterations $$T$$ is known and an $$O((1+\sigma^{2}\log(T/\delta))/\sqrt{T})$$ convergence rate when $$T$$ is unknown for SGD, where $$1-\delta$$ is the desired success probability. These bounds improve over existing bounds in the literature. We also revisit AdaGrad-Norm (Ward et al., 2019) and show a new analysis to obtain a high probability bound that does not require the bounded gradient assumption made in previous works. The full version of our paper contains results for the standard per-coordinate AdaGrad.more » « less
-
In this work, we describe a generic approach to show convergence with high probability for both stochastic convex and non-convex optimization with sub-Gaussian noise. In previous works for convex optimization, either the convergence is only in expectation or the bound depends on the diameter of the domain. Instead, we show high probability convergence with bounds depending on the initial distance to the optimal solution. The algorithms use step sizes analogous to the standard settings and are universal to Lipschitz functions, smooth functions, and their linear combinations. The method can be applied to the non-convex case. We demonstrate an $$O((1+\sigma^{2}\log(1/\delta))/T+\sigma/\sqrt{T})$$ convergence rate when the number of iterations $$T$$ is known and an $$O((1+\sigma^{2}\log(T/\delta))/\sqrt{T})$$ convergence rate when $$T$$ is unknown for SGD, where $$1-\delta$$ is the desired success probability. These bounds improve over existing bounds in the literature. We also revisit AdaGrad-Norm \cite{ward2019adagrad} and show a new analysis to obtain a high probability bound that does not require the bounded gradient assumption made in previous works. The full version of our paper contains results for the standard per-coordinate AdaGrad.more » « less
-
null (Ed.)Stochastic Gradient Descent (SGD) and its variants are the most used algorithms in machine learning applications. In particular, SGD with adaptive learning rates and momentum is the industry standard to train deep networks. Despite the enormous success of these methods, our theoretical understanding of these variants in the non-convex setting is not complete, with most of the results only proving convergence in expectation and with strong assumptions on the stochastic gradients. In this paper, we present a high probability analysis for adaptive and momentum algorithms, under weak assumptions on the function, stochastic gradients, and learning rates. We use it to prove for the first time the convergence of the gradients to zero in high probability in the smooth nonconvex setting for Delayed AdaGrad with momentum.more » « less
-
Stochastic gradient descent (SGD) is the optimization algorithm of choice in many machine learning applications such as regularized empirical risk minimization and training deep neural networks. The classical analysis of convergence of SGD is carried out under the assumption that the norm of the stochastic gradient is uniformly bounded. While this might hold for some loss functions, it is always violated for cases where the objective function is strongly convex. In (Bottou et al.,2016) a new analysis of convergence of SGD is performed under the assumption that stochastic gradients are bounded with respect to the true gradient norm. Here we show that for stochastic problems arising in machine learning such bound always holds. Moreover, we propose an alternative convergence analysis of SGD with diminishing learning rate regime, which is results in more relaxed conditions that those in (Bottou et al.,2016). We then move on the asynchronous parallel setting, and prove convergence of the Hogwild! algorithm in the same regime, obtaining the first convergence results for this method in the case of diminished learning rate.more » « less
