An official website of the United States government
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
HARD: Hyperplane ARrangement Descent
The problem of clustering points on a union of subspaces finds numerous applications in machine learning and computer vision, and it has been extensively studied in the past two decades. When the subspaces are low-dimensional, the problem can be formulated as a convex sparse optimization problem, for which numerous accurate, efficient and robust methods exist. When the subspaces are of high relative dimension (e.g., hyperplanes), the problem is intrinsically non-convex, and existing methods either lack theory, are computationally costly, lack robustness to outliers, or learn hyperplanes one at a time. In this paper, we propose Hyperplane ARangentment Descent (HARD), a method that robustly learns all the hyperplanes simultaneously by solving a novel non-convex non-smooth ℓ1 minimization problem. We provide geometric conditions under which the ground-truth hyperplane arrangement is a coordinate-wise minimizer of our objective. Furthermore, we devise efficient algorithms, and give conditions under which they converge to coordinate-wise minimizes. We provide empirical evidence that HARD surpasses state-of-the-art methods and further show an interesting experiment in clustering deep features on CIFAR-10.
more »
« less
- Award ID(s):
- 1704458
- PAR ID:
- 10535366
- Publisher / Repository:
- CPAL
- Date Published:
- Format(s):
- Medium: X
- Location:
- https://proceedings.mlr.press/v234/ding24a
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
We consider the problem of simultaneously clustering and learning a linear representation of data lying close to a union of low-dimensional manifolds, a fundamental task in machine learning and computer vision. When the manifolds are assumed to be linear subspaces, this reduces to the classical problem of subspace clustering, which has been studied extensively over the past two decades. Unfortunately, many real-world datasets such as natural images can not be well approximated by linear subspaces. On the other hand, numerous works have attempted to learn an appropriate transformation of the data, such that data is mapped from a union of general non-linear manifolds to a union of linear subspaces (with points from the same manifold being mapped to the same subspace). However, many existing works have limitations such as assuming knowledge of the membership of samples to clusters, requiring high sampling density, or being shown theoretically to learn trivial representations. In this paper, we propose to optimize the Maximal Coding Rate Reduction metric with respect to both the data representation and a novel doubly stochastic cluster membership, inspired by state-of-the-art subspace clustering results. We give a parameterization of such a representation and membership, allowing efficient mini-batching and one-shot initialization. Experiments on CIFAR-10, -20, -100, and TinyImageNet-200 datasets show that the proposed method is much more accurate and scalable than state-of-the-art deep clustering methods, and further learns a latent linear representation of the data.more » « less
-
null (Ed.)Abstract Subspace clustering is the unsupervised grouping of points lying near a union of low-dimensional linear subspaces. Algorithms based directly on geometric properties of such data tend to either provide poor empirical performance, lack theoretical guarantees or depend heavily on their initialization. We present a novel geometric approach to the subspace clustering problem that leverages ensembles of the $$K$$-subspace (KSS) algorithm via the evidence accumulation clustering framework. Our algorithm, referred to as ensemble $$K$$-subspaces (EKSSs), forms a co-association matrix whose $(i,j)$th entry is the number of times points $$i$$ and $$j$$ are clustered together by several runs of KSS with random initializations. We prove general recovery guarantees for any algorithm that forms an affinity matrix with entries close to a monotonic transformation of pairwise absolute inner products. We then show that a specific instance of EKSS results in an affinity matrix with entries of this form, and hence our proposed algorithm can provably recover subspaces under similar conditions to state-of-the-art algorithms. The finding is, to the best of our knowledge, the first recovery guarantee for evidence accumulation clustering and for KSS variants. We show on synthetic data that our method performs well in the traditionally challenging settings of subspaces with large intersection, subspaces with small principal angles and noisy data. Finally, we evaluate our algorithm on six common benchmark datasets and show that unlike existing methods, EKSS achieves excellent empirical performance when there are both a small and large number of points per subspace.more » « less
-
We consider the semi-supervised dimension reduction problem: given a high dimensional dataset with a small number of labeled data and huge number of unlabeled data, the goal is to find the low-dimensional embedding that yields good classification results. Most of the previous algorithms for this task are linkage-based algorithms. They try to enforce the must-link and cannot-link constraints in dimension reduction, leading to a nearest neighbor classifier in low dimensional space. In this paper, we propose a new hyperplane-based semi-supervised dimension reduction method---the main objective is to learn the low-dimensional features that can both approximate the original data and form a good separating hyperplane. We formulate this as a non-convex optimization problem and propose an efficient algorithm to solve it. The algorithm can scale to problems with millions of features and can easily incorporate non-negative constraints in order to learn interpretable non-negative features. Experiments on real world datasets demonstrate that our hyperplane-based dimension reduction method outperforms state-of-art linkage-based methods when very few labels are available.more » « less
-
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
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Proceeding:
- The DOI is not currently available.
