ICML 2020 Papers

We study Nesterov’s accelerated gradient method with constant step-size and momentum parameters in the stochastic approximation setting (unbiased gradients with bounded variance) and the finite-sum setting (where randomness is due to sampling mini-batches). To build better insight into the behavior of Nesterov’s method in stochastic settings, we focus throughout on objectives that are smooth, strongly-convex, and twice continuously differentiable. In the stochastic approximation setting, Nesterov’s method converges to a neighborhood of the optimal point at the same accelerated rate as in the deterministic setting. Perhaps surprisingly, in the finite-sum setting, we prove that Nesterov’s method may diverge with the usual choice of step-size and momentum, unless additional conditions on the problem related to conditioning and data coherence are satisfied. Our results shed light as to why Nesterov’s method may fail to converge or achieve acceleration in the finite-sum setting.

We consider the product of determinantal point processes (DPPs), a point process whose probability mass is proportional to the product of principal minors of multiple matrices as a natural, promising generalization of DPPs. We study the computational complexity of computing its normalizing constant, which is among the most essential probabilistic inference tasks. Our complexity-theoretic results (almost) rule out the existence of efficient algorithms for this task, unless input matrices are forced to have favorable structures. In particular, we prove the following: (1) Computing $\sum_{S} \det(\mathbf{A}_{S,S})^p$ exactly for every (fixed) positive even integer $p$ is $\textsf{UP}$-hard and $\textsf{Mod}_3\textsf{P}$-hard, which gives a negative answer to an open question posed by Kulesza and Taskar (2012). (2) $\sum_{S} \det(\mathbf{A}_{S,S}) \det(\mathbf{B}_{S,S}) \det(\mathbf{C}_{S,S})$ is $\textsf{NP}$-hard to approximate within a factor of $ 2^{\mathcal{O}(|\mathcal{I}|^{1-\epsilon})} $ for any $\epsilon > 0$, where $|\mathcal{I}|$ is the input size. This result is stronger than $\sharp\textsf{P}$-hardness for the case of two matrices by Gillenwater (2014). (3) There exists a $ k^{\mathcal{O}(k)} |\mathcal{I}|^{\mathcal{O}(1)} $-time algorithm for computing $\sum_{S} \det(\mathbf{A}_{S,S}) \det(\mathbf{B}_{S,S})$, where $k$ is “the maximum rank of $\mathbf{A}$ and $\mathbf{B}$” or “the treewidth of the graph formed by nonzero entries of $\mathbf{A}$ and $\mathbf{B}$.” Such parameterized algorithms are said to be fixed-parameter tractable.

Tuning hyperparameters for unsupervised learning problems is difficult in general due to the lack of ground truth for validation. However, the success of most clustering methods depends heavily on the correct choice of the involved hyperparameters. Take for example the Lagrange multipliers of penalty terms in semidefinite programming (SDP) relaxations of community detection in networks, or the bandwidth parameter needed in the Gaussian kernel used to construct similarity matrices for spectral clustering. Despite the popularity of these clustering algorithms, there are not many provable methods for tuning these hyperparameters. In this paper, we provide an overarching framework with provable guarantees for tuning hyperparameters in the above class of problems under two different models. Our framework can be augmented with a cross validation procedure to do model selection as well. In a variety of simulation and real data experiments, we show that our framework outperforms other widely used tuning procedures in a broad range of parameter settings.

In this work, we propose a robust test for the multivariate two-sample problem through projective ensemble, which is a generalization of the Cramer-von Mises statistic. The proposed test statistic has a simple closed-form expression without any tuning parameters involved, it is easy to implement can be computed in quadratic time. Moreover, our test is insensitive to the dimension and consistent against all fixed alternatives, it does not require the moment assumption and is robust to the presence of outliers. We study the asymptotic behaviors of the test statistic under the null and two kinds of alternative hypotheses. We also suggest a permutation procedure to approximate critical values and employ its consistency. We demonstrate the effectiveness of our test through extensive simulation studies and a real data application.

The variational autoencoder (VAE) can learn the manifold of natural images on certain datasets, as evidenced by meaningful interpolating or extrapolating in the continuous latent space. However, on discrete data such as text, it is unclear if unsupervised learning can discover similar latent space that allows controllable manipulation. In this work, we find that sequence VAEs trained on text fail to properly decode when the latent codes are manipulated, because the modified codes often land in holes or vacant regions in the aggregated posterior latent space, where the decoding network fails to generalize. Both as a validation of the explanation and as a fix to the problem, we propose to constrain the posterior mean to a learned probability simplex, and performs manipulation within this simplex. Our proposed method mitigates the latent vacancy problem and achieves the first success in unsupervised learning of controllable representations for text. Empirically, our method outperforms unsupervised baselines and strong supervised approaches on text style transfer, and is capable of performing more flexible fine-grained control over text generation than existing methods.

In the current era of personalized recommendation, one major interest is to develop an optimal individualized decision rule that assigns individuals with the best treatment option according to their covariates. Estimation of optimal decision rules (ODR) has been extensively investigated recently, however, at present, no testing procedure is proposed to verify whether these ODRs are significantly better than the naive decision rule that always assigning individuals to a fixed treatment option. In this paper, we propose a testing procedure for detecting the existence of an ODR that is better than the naive decision rule under the randomized trials. We construct the proposed test based on the difference of estimated value functions using the augmented inverse probability weighted method. The asymptotic distributions of the proposed test statistic under the null and local alternative hypotheses are established. Based on the established asymptotic distributions, we further develop a sample size calculation formula for testing the existence of an ODR in designing A/B tests. Extensive simulations and a real data application to a schizophrenia clinical trial data are conducted to demonstrate the empirical validity of the proposed methods.

We provide a computational complexity analysis for the Sinkhorn algorithm that solves the entropic regularized Unbalanced Optimal Transport (UOT) problem between two measures of possibly different masses with at most $n$ components. We show that the complexity of the Sinkhorn algorithm for finding an $\varepsilon$-approximate solution to the UOT problem is of order $\widetilde{\mathcal{O}}(n^2/ \varepsilon)$. To the best of our knowledge, this complexity is better than the best known complexity upper bound of the Sinkhorn algorithm for solving the Optimal Transport (OT) problem, which is of order $\widetilde{\mathcal{O}}(n^2/\varepsilon^2)$. Our proof technique is based on the geometric convergence rate of the Sinkhorn updates to the optimal dual solution of the entropic regularized UOT problem and scaling properties of the primal solution. It is also different from the proof technique used to establish the complexity of the Sinkhorn algorithm for approximating the OT problem since the UOT solution does not need to meet the marginal constraints of the measures.

There has been a growing realization of the potential of Bayesian machine learning as a platform that can provide both flexible modeling, accurate predictions as well as coherent uncertainty statements. In particular, Bayesian Additive Regression Trees (BART) have emerged as one of today’s most effective general approaches to predictive modeling under minimal assumptions. Statistical theoretical developments for machine learning have been mostly concerned with approximability or rates of estimation when recovering infinite dimensional objects (curves or densities). Despite the impressive array of available theoretical results, the literature has been largely silent about uncertainty quantification. In this work, we continue the theoretical investigation of BART initiated recently by Rockova and van der Pas (2017). We focus on statistical inference questions. In particular, we study the Bernstein-von Mises (BvM) phenomenon (i.e. asymptotic normality) for smooth linear functionals of the regression surface within the framework of non-parametric regression with fixed covariates. Our semi-parametric BvM results show that, beyond rate-optimal estimation, BART can be also used for valid statistical inference.

With the rapid adoption of machine learning systems in sensitive applications, there is an increasing need to make black-box models explainable. Often we want to identify an influential group of training samples in a particular test prediction for a given machine learning model. Existing influence functions tackle this problem by using first-order approximations of the effect of removing a sample from the training set on model parameters. To compute the influence of a group of training samples (rather than an individual point) in model predictions, the change in optimal model parameters after removing that group from the training set can be large. Thus, in such cases, the first-order approximation can be loose. In this paper, we address this issue and propose second-order influence functions for identifying influential groups in test-time predictions. For linear models, across different sizes and types of groups, we show that using the proposed second-order influence function improves the correlation between the computed influence values and the ground truth ones. We also show that second-order influence functions could be used with optimization techniques to improve the selection of the most influential group for a test-sample.

We take a more rigorous look at Relativistic Generative Adversarial Networks (RGANs) and prove that the objective function of the discriminator is a statistical divergence for any concave function f with minimal properties. We devise additional variants of relativistic f-divergences. We show that the Wasserstein distance is weaker than f-divergences which are weaker than relativistic f-divergences. Given the good performance of RGANs, this suggests that Wasserstein GAN does not performs well primarily because of the weak metric, but rather because of regularization and the use of a relativistic discriminator. We introduce the minimum-variance unbiased estimator (MVUE) for Relativistic paired GANs (RpGANs; originally called RGANs which could bring confusion) and show that it does not perform better. We show that the estimator of Relativistic average GANs (RaGANs) is asymptotically unbiased and that the finite-sample bias is small; removing this bias does not improve performance.

Recent papers have demonstrated that ensemble stumps and trees could be vulnerable to small input perturbations, so robustness verification and defense for those models have become an important research problem. However, due to the structure of decision trees, where each node makes decision purely based on one feature value, all the previous works only consider the $\ell_\infty$ norm perturbation. To study robustness with respect to a general $\ell_p$ norm perturbation, one has to consider the correlation between perturbations on different features, which has not been handled by previous algorithms. In this paper, we study the problem of robustness verification and certified defense with respect to general $\ell_p$ norm perturbations for ensemble decision stumps and trees. For robustness verification of ensemble stumps, we prove that complete verification is NP-complete for $p\in(0, \infty)$ while polynomial time algorithms exist for $p=0$ or $\infty$. For $p\in(0, \infty)$ we develop an efficient dynamic programming based algorithm for sound verification of ensemble stumps. For ensemble trees, we generalize the previous multi-level robustness verification algorithm to $\ell_p$ norm. We demonstrate the first certified defense method for training ensemble stumps and trees with respect to $\ell_p$ norm perturbations, and verify its effectiveness empirically on real datasets.

Recent work has shown generative adversarial networks (GANs) can generate highly realistic images, that are often indistinguishable (by humans) from real images. Most images so generated are not contained in the training dataset, suggesting potential for augmenting training sets with GAN-generated data. While this scenario is of particular relevance when there are limited data available, there is still the issue of training the GAN itself based on that limited data. To facilitate this, we leverage existing GAN models pretrained on large-scale datasets (like ImageNet) to introduce additional knowledge (which may not exist within the limited data), following the concept of transfer learning. Demonstrated by natural-image generation, we reveal that low-level filters (those close to observations) of both the generator and discriminator of pretrained GANs can be transferred to facilitate generation in a perceptually-distinct target domain with limited training data. To further adapt the transferred filters to the target domain, we propose adaptive filter modulation (AdaFM). An extensive set of experiments is presented to demonstrate the effectiveness of the proposed techniques on generation with limited data.

Learning from unordered sets is a fundamental learning setup, recently attracting increasing attention. Research in this area has focused on the case where elements of the set are represented by feature vectors, and far less emphasis has been given to the common case where set elements themselves adhere to their own symmetries. That case is relevant to numerous applications, from deblurring image bursts to multi-view 3D shape recognition and reconstruction. In this paper, we present a principled approach to learning sets of general symmetric elements. We first characterize the space of linear layers that are equivariant both to element reordering and to the inherent symmetries of elements, like translation in the case of images. We further show that networks that are composed of these layers, called Deep Sets for Symmetric Elements layers (DSS), are universal approximators of both invariant and equivariant functions. DSS layers are also straightforward to implement. Finally, we show that they improve over existing set-learning architectures in a series of experiments with images, graphs, and point-clouds.

The goal of universal machine translation is to learn to translate between any pair of languages. Despite impressive empirical results and an increasing interest in massively multilingual models, theoretical analysis on translation errors made by such universal machine translation models is only nascent. In this paper, we formally prove certain impossibilities of this endeavour in general, as well as prove positive results in the presence of additional (but natural) structure of data. For the former, we derive a lower bound on the translation error in the many-to-many translation setting, which shows that any algorithm aiming to learn shared sentence representations among multiple language pairs has to make a large translation error on at least one of the translation tasks, if no assumption on the structure of the languages is made. For the latter, we show that if the paired documents in the corpus follow a natural \emph{encoder-decoder} generative process, we can expect a natural notion of “generalization”: a linear number of language pairs, rather than quadratic, suffices to learn a good representation. Our theory also explains what kinds of connection graphs between pairs of languages are better suited: ones with longer paths result in worse sample complexity. We believe our theoretical insights and implications contribute to the future algorithmic design of universal machine translation.

The Transformer is widely used in natural language processing tasks. To train a Transformer however, one usually needs a carefully designed learning rate warm-up stage, which is shown to be crucial to the final performance but will slow down the optimization and bring more hyper-parameter tunings. In this paper, we first study theoretically why the learning rate warm-up stage is essential and show that the location of layer normalization matters. Specifically, we prove with mean field theory that at initialization, for the original-designed Post-LN Transformer, which places the layer normalization between the residual blocks, the expected gradients of the parameters near the output layer are large. Therefore, using a large learning rate on those gradients makes the training unstable. The warm-up stage is practically helpful for avoiding this problem. On the other hand, our theory also shows that if the layer normalization is put inside the residual blocks (recently proposed as Pre-LN Transformer), the gradients are well-behaved at initialization. This motivates us to remove the warm-up stage for the training of Pre-LN Transformers. We show in our experiments that Pre-LN Transformers without the warm-up stage can reach comparable results with baselines while requiring significantly less training time and hyper-parameter tuning on a wide range of applications.

While the impact of variational inference (VI) on posterior inference in a fixed generative model is well-characterized, its role in regularizing a learned generative model when used in variational autoencoders (VAEs) is poorly understood. We study the regularizing effects of variational distributions on learning in generative models from two perspectives. First, we analyze the role that the choice of variational family plays in imparting uniqueness to the learned model by restricting the set of optimal generative models. Second, we study the regularization effect of the variational family on the local geometry of the decoding model. This analysis uncovers the regularizer implicit in the $\beta$-VAE objective, and leads to an approximation consisting of a deterministic autoencoding objective plus analytic regularizers that depend on the Hessian or Jacobian of the decoding model, unifying VAEs with recent heuristics proposed for training regularized autoencoders. We empirically verify these findings, observing that the proposed deterministic objective exhibits similar behavior to the $\beta$-VAE in terms of objective value and sample quality.

We consider nonconvex-concave minimax problems, $\min_{\mathbf{x}} \max_{\mathbf{y} \in \mathcal{Y}} f(\mathbf{x}, \mathbf{y})$, where $f$ is nonconvex in $\mathbf{x}$ but concave in $\mathbf{y}$ and $\mathcal{Y}$ is a convex and bounded set. One of the most popular algorithms for solving this problem is the celebrated gradient descent ascent (GDA) algorithm, which has been widely used in machine learning, control theory and economics. Despite the extensive convergence results for the convex-concave setting, GDA with equal stepsize can converge to limit cycles or even diverge in a general setting. In this paper, we present the complexity results on two-time-scale GDA for solving nonconvex-concave minimax problems, showing that the algorithm can find a stationary point of the function $\Phi(\cdot) := \max_{\mathbf{y} \in \mathcal{Y}} f(\cdot, \mathbf{y})$ efficiently. To the best our knowledge, this is the first nonasymptotic analysis for two-time-scale GDA in this setting, shedding light on its superior practical performance in training generative adversarial networks (GANs) and other real applications.

A classic problem in unsupervised learning and data analysis is to find simpler and easy-to-visualize representations of the data that preserve its essential properties. A widely-used method to preserve the underlying hierarchical structure of the data while reducing its complexity is to find an embedding of the data into a tree or an ultrametric, but computing such an embedding on a data set of $n$ points in $\Omega(\log n)$ dimensions incurs a quite prohibitive running time of $\Theta(n^2)$. In this paper, we provide a new algorithm which takes as input a set of points $P$ in $\R^d$, and for every $c\ge 1$, runs in time $n^{1+\frac{\rho}{c^2}}$ (for some universal constant $\rho>1$) to output an ultrametric $\Delta$ such that for any two points $u,v$ in $P$, we have $\Delta(u,v)$ is within a multiplicative factor of $5c$ to the distance between $u$ and $v$ in the best ultrametric representation of $P$. Here, the best ultrametric is the ultrametric $\tilde\Delta$ that minimizes the maximum distance distortion with respect to the $\ell_2$ distance, namely that minimizes $\underset{u,v \in P}{\max} \nicefrac{\tilde\Delta(u,v)}{\|u-v\|_2}$. We complement the above result by showing that under popular complexity theoretic assumptions, for every constant $\varepsilon>0$, no algorithm with running time $n^{2-\varepsilon}$ can distinguish between inputs in $\ell_\infty$-metric that admit isometric embedding and those that incur a distortion of $\nicefrac{3}{2}$. Finally, we present empirical evaluation on classic machine learning datasets and show that the output of our algorithm is comparable to the output of the linkage algorithms while achieving a much faster running time.

Efficient construction of checkpoints/snapshots is a critical tool for training and diagnosing deep learning models. In this paper, we propose a lossy compression scheme for checkpoint constructions (called LC-Checkpoint). LC-Checkpoint simultaneously maximizes the compression rate and optimizes the recovery speed, under the assumption that SGD is used to train the model. LC-Checkpoint uses quantization and priority promotion to store the most crucial information for SGD to recover, and then uses a Huffman coding to leverage the non-uniform distribution of the gradient scales. Our extensive experiments show that LC-Checkpoint achieves a compression rate up to 28{\texttimes} and recovery speedup up to 5.77{\texttimes} over a state-of-the-art algorithm (SCAR).