More Like this

1. Stability of SGD: Tightness analysis and improved bounds

   Zhang, Yikai ; Zhang, Wenjia ; Bald, Sammy ; Pingali, Vamsi ; Chen, Chao ; Goswami, Mayank ( August 2022 , Uncertainty in artificial intelligence)

   Stochastic Gradient Descent (SGD) based methods have been widely used for training large-scale machine learning models that also generalize well in practice. Several explanations have been offered for this generalization performance, a prominent one being algorithmic stability [18]. However, there are no known examples of smooth loss functions for which the analysis can be shown to be tight. Furthermore, apart from the properties of the loss function, data distribution has also been shown to be an important factor in generalization performance. This raises the question: is the stability analysis of [18] tight for smooth functions, and if not, for what kind of loss functions and data distributions can the stability analysis be improved? In this paper we first settle open questions regarding tightness of bounds in the data-independent setting: we show that for general datasets, the existing analysis for convex and strongly-convex loss functions is tight, but it can be improved for non-convex loss functions. Next, we give a novel and improved data-dependent bounds: we show stability upper bounds for a large class of convex regularized loss functions, with negligible regularization parameters, and improve existing data-dependent bounds in the non-convex setting. We hope that our results will initiate further efforts to better understand the data-dependent setting under non-convex loss functions, leading to an improved understanding of the generalization abilities of deep networks. 

   more » « less
2. Batched Predecessor and Sorting with Size-Priced Information in External Memory

   Bender, Michael ; Goswami, Mayank ; Medjedovic, Dzejla ; Montes, Pablo ; Tsichlas, Kostas ( January 2021 , Latin American Theoretical Informatics Symposium)

   In the unit-cost comparison model, a black box takes an input two items and outputs the result of the comparison. Problems like sorting and searching have been studied in this model, and it has been general- ized to include the concept of priced information, where different pairs of items (say database records) have different comparison costs. These comparison costs can be arbitrary (in which case no algorithm can be close to optimal (Charikar et al. STOC 2000)), structured (for exam- ple, the comparison cost may depend on the length of the databases (Gupta et al. FOCS 2001)), or stochastic (Angelov et al. LATIN 2008). Motivated by the database setting where the cost depends on the sizes of the items, we consider the problems of sorting and batched predecessor where two non-uniform sets of items A and B are given as input. (1) In the RAM setting, we consider the scenario where both sets have n keys each. The cost to compare two items in A is a, to compare an item of A to an item of B is b, and to compare two items in B is c. We give upper and lower bounds for the case a ≤ b ≤ c, the case that serves as a warmup for the generalization to the external-memory model. Notice that the case b = 1,a = c = ∞ is the famous “nuts and bolts” problem. ) In the Disk-Access Model (DAM), where transferring elements between disk and internal memory is the main bottleneck, we con- sider the scenario where elements in B are larger than elements in A. The larger items take more I/Os to be brought into memory, consume more space in internal memory, and are required in their entirety for comparisons. A key observation is that the complexity of sorting depends heavily on the interleaving of the small and large items in the final sorted order. If all large elements come after all small elements in the final sorted order, sorting each type separately and concatenating is optimal. However, if the set of predecessors of B in A has size k ≪ n, one must solve an associated batched predecessor problem in order to achieve optimality. We first give output-sensitive lower and upper bounds on the batched predecessor problem, and use these to derive bounds on the complexity of sorting in the two models. Our bounds are tight in most cases, and require novel generalizations of the classical lower bound techniques in external memory to accommodate the non-uniformity of keys. 

   more » « less
3. Batched Predecessor and Sorting with Size-Priced Information in External Memory

   https://doi.org/10.1007/978-3-030-61792-9_13  

   Bender, Michael A ; Goswami, Mayank ; Medjedovic, Dzejla ; Montes, Pablo ; Tsichlas, Kostas ( January 2021 , Latin American Symposium on Theoretical Informatics)

   The fundamental problems of sorting and searching, traditionally studied in the unit-cost comparison model, have been generalized to include priced information, where different pairs of items have different comparison costs. These costs can be arbitrary (Charikar et al. STOC 2000), structured (Gupta et al. FOCS 2001), or stochastic (Angelov et al. LATIN 2008). Motivated by the database setting where the comparison cost depends on the sizes of the records, we consider the problems of sorting and batched predecessor where two non-uniform sets of items A and B are given as input. In the RAM model, pairwise comparisons (A-A, A-B and B-B) have respective comparison costs a, b and c. We give upper and lower bounds for the case a<= b <= c, which serves as a warmup for the generalization to the external-memory model. In the Disk-Access Model (DAM), where transferring elements between disk and RAM is the main bottleneck, we consider the scenario where elements in B are larger than elements in A. All items are required in their entirety for comparisons in RAM. A key observation is that the complexity of sorting depends on the interleaving of the small and large items in the final sorted order, and with a high degree of interleaving, the lower bound is dominated by an associated batched predecessor problem. We give output-sensitive bounds on the batched predecessor and sorting; our bounds are tight in most cases. Our lower bounds require novel generalizations of lower bound techniques in external memory to accommodate non-uniform keys. 

   more » « less
4. High Probability Convergence of Stochastic Gradient Methods

   Liu, Zijian ; Nguyen, Ta Duy ; Nguyen, Thien Hang ; Ene, Alina ; Nguyen, Huy ( July 2023 , International Conference on Machine Learning)

   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
5. Computationally and Statistically Efficient Truncated Regression

   Daskalakis, Constantinos ; Gouleakis, Themis ; Tzamos, Christos ; Zampetakis, Manolis ( January 2019 , Proceedings of Machine Learning Research)

   We provide a computationally and statistically efficient estimator for the classical problem of trun-cated linear regression, where the dependent variabley=wTx+εand its corresponding vector ofcovariatesx∈Rkare only revealed if the dependent variable falls in some subsetS⊆R; otherwisethe existence of the pair(x,y)is hidden. This problem has remained a challenge since the earlyworks of Tobin (1958); Amemiya (1973); Hausman and Wise (1977); Breen et al. (1996), its appli-cations are abundant, and its history dates back even further to the work of Galton, Pearson, Lee,and Fisher Galton (1897); Pearson and Lee (1908); Lee (1914); Fisher (1931). While consistent es-timators of the regression coefficients have been identified, the error rates are not well-understood,especially in high-dimensional settings.Under a “thickness assumption” about the covariance matrix of the covariates in the revealed sample, we provide a computationally efficient estimator for the coefficient vectorwfromnre-vealed samples that attains`2errorO(√k/n), recovering the guarantees of least squares in thestandard (untruncated) linear regression setting. Our estimator uses Projected Stochastic Gradi-ent Descent (PSGD) on the negative log-likelihood of the truncated sample, and only needs ora-cle access to the setS, which may otherwise be arbitrary, and in particular may be non-convex.PSGD must be restricted to an appropriately defined convex cone to guarantee that the negativelog-likelihood is strongly convex, which in turn is established using concentration of matrices onvariables with sub-exponential tails. We perform experiments on simulated data to illustrate the accuracy of our estimator.As a corollary of our work, we show that SGD provably learns the parameters of single-layerneural networks with noisy Relu activation functions Nair and Hinton (2010); Bengio et al. (2013);Gulcehre et al. (2016), given linearly many, in the number of network parameters, input-outputpairs in the realizable setting. 

   more » « less