UAI 2022 论文列表

Learning physics models in the form of Partial Differential Equations (PDEs) is carried out through back-propagation to match the simulations of the physics model with experimental observations. Nevertheless, such matching involves computation over billions of elements, presenting a significant computational overhead. We notice many PDEs in real world problems are sparse and decomposable, where the temporal updates and the spatial features are sparsely concentrated on small interface regions. We propose RAPID-PDE, an algorithm to expedite the learning of sparse and decomposable PDEs. Our RAPID-PDE first uses random projection to compress the high dimensional sparse updates and features into low dimensional representations and then use these compressed signals during learning. Crucially, such a conversion is only carried out once prior to learning and the entire learning process is conducted in the compressed space. Theoretically, we derive a constant factor approximation between the projected loss function and the original one with logarithmic number of projected dimensions. Empirically, we demonstrate RAPID-PDE with data compressed to 0.05% of its original size learns similar models compared with uncompressed algorithms in learning a set of phase-field models which govern the spatial-temporal dynamics of nano-scale structures in metallic materials.

Dynamic topic modeling is a well established tool for capturing the temporal dynamics of the topics of a corpus. In this work, we develop a scalable dynamic topic model by utilizing the correlation among the words in the vocabulary. By correlating previously independent temporal processes for words, our new model allows us to reliably estimate the topic representations containing less frequent words. We develop an amortised variational inference method with self-normalised importance sampling approximation to the word distribution that dramatically reduces the computational complexity and the number of variational parameters in order to handle large vocabularies. With extensive experiments on text datasets, we show that our method significantly outperforms the previous works by modeling word correlations, and it is able to handle real world data with a large vocabulary which could not be processed by previous continuous dynamic topic models.

An established way to improve the transferability of black-box evasion attacks is to craft the adversarial examples on an ensemble-based surrogate to increase diversity. We argue that transferability is fundamentally related to uncertainty. Based on a state-of-the-art Bayesian Deep Learning technique, we propose a new method to efficiently build a surrogate by sampling approximately from the posterior distribution of neural network weights, which represents the belief about the value of each parameter. Our extensive experiments on ImageNet, CIFAR-10 and MNIST show that our approach improves the success rates of four state-of-the-art attacks significantly (up to 83.2 percentage points), in both intra-architecture and inter-architecture transferability. On ImageNet, our approach can reach 94% of success rate while reducing training computations from 11.6 to 2.4 exaflops, compared to an ensemble of independently trained DNNs. Our vanilla surrogate achieves 87.5% of the time higher transferability than three test-time techniques designed for this purpose. Our work demonstrates that the way to train a surrogate has been overlooked, although it is an important element of transfer-based attacks. We are, therefore, the first to review the effectiveness of several training methods in increasing transferability. We provide new directions to better understand the transferability phenomenon and offer a simple but strong baseline for future work.

The intersection of learning to rank and choice modeling is an active area of research with applications in e-commerce, information retrieval and the social sciences. In some applications such as recommendation systems, the statistician is primarily interested in recovering the set of the top ranked items from a large pool of items as efficiently as possible using passively collected discrete choice data, i.e., the user picks one item from a set of multiple items. Motivated by this practical consideration, we propose the choice-based Borda count algorithm as a fast and accurate ranking algorithm for top $K$-recovery i.e., correctly identifying all of the top $K$ items. We show that the choice-based Borda count algorithm has optimal sample complexity for top-$K$ recovery under a broad class of random utility models. We prove that in the limit, the choice-based Borda count algorithm produces the same top-$K$ estimate as the commonly used Maximum Likelihood Estimate method but the former’s speed and simplicity brings considerable advantages in practice. Experiments on both synthetic and real datasets show that the counting algorithm is competitive with commonly used ranking algorithms in terms of accuracy while being several orders of magnitude faster.

To address spatial imbalances in the supply and demand of drivers, ridesharing platforms can make use of policies to direct driver relocation. We study a simple model of this problem, which allows us to give a constructive characterization of the unique fixpoint of system dynamics. Using this construction, we design a dynamic policy that provides stronger, than previous work, guarantees about its rate of convergence to the fixpoint. Simulations demonstrate the benefits of our approach.

The problem of detecting the Out-of-Distribution (OoD) inputs is of paramount importance for Deep Neural Networks. It has been previously shown that even Deep Generative Models that allow estimating the density of the inputs may not be reliable and often tend to make over-confident predictions for OoDs, assigning to them a higher density than to the in-distribution data. This over-confidence in a single model can be potentially mitigated with Bayesian inference over the model parameters that take into account epistemic uncertainty. This paper investigates three approaches to Bayesian inference: stochastic gradient Markov chain Monte Carlo, Bayes by Backpropagation, and Stochastic Weight Averaging-Gaussian. The inference is implemented over the weights of the deep neural networks that parameterize the likelihood of the Variational Autoencoder. We empirically evaluate the approaches against several benchmarks that are often used for OoD detection: estimation of the marginal likelihood utilizing sampled model ensemble, typicality test, disagreement score, and Watanabe-Akaike Information Criterion. Finally, we introduce two simple scores that demonstrate the state-of-the-art performance.

Current deep neural networks (DNNs) are vulnerable to adversarial attacks, where adversarial perturbations to the inputs can change or manipulate classification. To defend against such attacks, an effective and popular approach, known as adversarial training (AT), has been shown to mitigate the negative impact of adversarial attacks by virtue of a min-max robust training method. While effective, it remains unclear whether it can successfully be adapted to the distributed learning context. The power of distributed optimization over multiple machines enables us to scale up robust training over large models and datasets. Spurred by that, we propose distributed adversarial training (DAT), a large-batch adversarial training framework implemented over multiple machines. We show that DAT is general, which supports training over labeled and unlabeled data, multiple types of attack generation methods, and gradient compression operations favored for distributed optimization. Theoretically, we provide, under standard conditions in the optimization theory, the convergence rate of DAT to the first-order stationary points in general non-convex settings. Empirically, we demonstrate that DAT either matches or outperforms state-of-the-art robust accuracies and achieves a graceful training speedup (e.g., on ResNet-50 under ImageNet). Codes are available at https://github.com/dat-2022/dat.

This study addresses the problem of learning an extended summary causal graph from time series. The algorithms we propose fit within the well-known constraint-based framework for causal discovery and make use of information-theoretic measures to determine (in)dependencies between time series. We first introduce generalizations of the causation entropy measure to any lagged or instantaneous relations, prior to using this measure to construct extended summary causal graphs by adapting two well-known algorithms, namely PC and FCI. The behaviour of our method is illustrated through several experiments.

Dimension reduction for high-dimensional count data with a large proportion of zeros is an important task in various applications. As a large number of dimension reduction methods rely on the proximity measure, we develop a dissimilarity measure that is well-suited for small counts based on the Kullback-Leibler divergence. We compare the proposed measure with other widely used dissimilarity measures and show that the proposed one has superior discriminative ability when applied to high-dimensional count data having an excess of zeros. Extensive empirical results, on both simulated and publicly-available real-world datasets that contain many zeros, demonstrate that the proposed dissimilarity measure can improve a wide range of dimension reduction methods.

Differentially Private (DP) data release is a promising technique to disseminate data without compromising the privacy of data subjects. However the majority of prior work has focused on scenarios where a single party owns all the data. In this paper we focus on the multi-party setting, where different stakeholders own disjoint sets of attributes belonging to the same group of data subjects. Within the context of linear regression that allow all parties to train models on the complete data without the ability to infer private attributes or identities of individuals, we start with directly applying Gaussian mechanism and show it has the small eigenvalue problem. We further propose our novel method and prove it asymptotically converges to the optimal (non-private) solutions with increasing dataset size. We substantiate the theoretical results through experiments on both artificial and real-world datasets.

Stochastic gradient descent ascent (SGDA) and its variants have been the workhorse for solving minimax problems. However, in contrast to the well-studied stochastic gradient descent (SGD) with differential privacy (DP) constraints, there is little work on understanding the generalization (utility) of SGDA with DP constraints. In this paper, we use the algorithmic stability approach to establish the generalization (utility) of DP-SGDA in different settings. In particular, for the convex-concave setting, we prove that the DP-SGDA can achieve an optimal utility rate in terms of the weak primal-dual population risk in both smooth and non-smooth cases. To our best knowledge, this is the first-ever-known result for DP-SGDA in the non-smooth case. We further provide its utility analysis in the nonconvex-strongly-concave setting which is the first-ever-known result in terms of the primal population risk. The convergence and generalization results for this nonconvex setting are new even in the non-private setting. Finally, numerical experiments are conducted to demonstrate the effectiveness of DP-SGDA for both convex and nonconvex cases.

The deterministic policy gradient (DPG) method proposed in Silver et al. [2014] has been demonstrated to exhibit superior performance particularly for applications with multi-dimensional and continuous action spaces. However, it remains unclear whether DPG converges, and if so, how fast it converges and whether it converges as efficiently as other PG methods. In this paper, we provide a theoretical analysis of DPG to answer those questions. We study the single timescale DPG (often the case in practice) in both on-policy and off-policy settings, and show that both algorithms attain an $\epsilon$-accurate stationary policy with a sample complexity of $\mathcal{O}(\epsilon^{-2})$. Moreover, we establish the convergence rate for DPG under Gaussian noise exploration, which is widely adopted in practice to improve the performance of DPG. To our best knowledge, this is the first non-asymptotic convergence characterization for DPG methods.

A line of work has shown that natural text processing models are vulnerable to adversarial examples. Correspondingly, various defense methods are proposed to mitigate the threat of textual adversarial examples, \textit{e.g.} adversarial training, input transformations, detection, \textit{etc}. In this work, we treat the optimization process for synonym substitution based textual adversarial attacks as a specific sequence of word replacement, in which each word mutually influences other words. We identify that we could destroy such mutual interaction and eliminate the adversarial perturbation by randomly substituting a word with its synonyms. Based on this observation, we propose a novel textual adversarial example detection method, termed \textit{Randomized Substitution and Vote} (RS&V), which votes the prediction label by accumulating the logits of $k$ samples generated by randomly substituting the words in the input text with synonyms. The proposed RS&V is generally applicable to any existing neural networks without modification on the architecture or extra training, and it is orthogonal to prior work on making the classification network itself more robust. Empirical evaluations on three benchmark datasets demonstrate that our RS&V could detect the textual adversarial examples more successfully than the existing detection methods while maintaining the high classification accuracy on benign samples.

In this paper we propose the deep Dirichlet process mixture (DDPM) model, which is an unsupervised method that simultaneously performs clustering and feature learning. The traditional Dirichlet process mixture model can infer the number of mixture components, but its flexibility is restricted since the clustering is performed in the raw feature space. Our method alleviates this limitation by using the flow-based deep neural network to learn more expressive features. DDPM unifies Dirichlet processes and the flow-based model with Monte Carlo expectation-maximization, and uses Gibbs sampling to sample from the posterior. This combination allows our method to exploit the mutually beneficial relation between clustering and feature learning. The effectiveness of DDPM is demonstrated by thorough experiments in various synthetic and real-world datasets.

In open multiagent systems, the set of agents operating in the environment changes over time and in ways that are nontrivial to predict. For example, if collaborative robots were tasked with fighting wildfires, they may run out of suppressants and be temporarily unavailable to assist their peers. Because an agent’s optimal action depends on the actions of others, each agent must not only predict the actions of its peers, but, before that, reason whether they are even present to perform an action. Addressing openness thus requires agents to model each other’s presence, which can be enhanced through agents communicating about their presence in the environment. At the same time, communicative acts can also incur costs (e.g., consuming limited bandwidth), and thus an agent must tradeoff the benefits of enhanced coordination with the costs of communication. We present a new principled, decision-theoretic method in the context provided by the recent communicative interactive POMDP framework for planning in open agent settings that balances this tradeoff. Simulations of multiagent wildfire suppression problems demonstrate how communication can improve planning in open agent environments, as well as how agents tradeoff the benefits and costs of communication under different scenarios.

Conventional machine learning applications typically assume that data samples are independently and identically distributed (i.i.d.). However, practical scenarios often involve a data-generating process that produces highly dependent data samples, which are known to heavily bias the stochastic optimization process and slow down the convergence of learning. In this paper, we conduct a fundamental study on how different stochastic data sampling schemes affect the sample complexity of online stochastic gradient descent (SGD) over highly dependent data. Specifically, with a $\phi$-mixing process of data, we show that online SGD with proper periodic data-subsampling achieves an improved sample complexity over the standard online SGD in the full spectrum of the data dependence level. Interestingly, even subsampling a subset of data samples can accelerate the convergence of online SGD over highly dependent data. Moreover, we show that online SGD with mini-batch sampling can further substantially improve the sample complexity over online SGD with periodic data-subsampling over highly dependent data. Numerical experiments validate our theoretical results.

Off-policy Evaluation (OPE) methods are a crucial tool for evaluating policies in high-stakes domains such as healthcare, where exploration is often infeasible, unethical, or expensive. However, the extent to which such methods can be trusted under adversarial threats to data quality is largely unexplored. In this work, we make the first attempt at investigating the sensitivity of OPE methods to marginal adversarial perturbations to the data. We design a generic data poisoning attack framework leveraging influence functions from robust statistics to carefully construct perturbations that maximize error in the policy value estimates. We carry out extensive experimentation with multiple healthcare and control datasets. Our results demonstrate that many existing OPE methods are highly prone to generating value estimates with large errors when subject to data poisoning attacks, even for small adversarial perturbations. These findings question the reliability of policy values derived using OPE methods and motivate the need for developing OPE methods that are statistically robust to train-time data poisoning attacks.

Randomized smoothing is a recent technique that achieves state-of-art performance in training certifiably robust deep neural networks. While the smoothing family of distributions is often connected to the choice of the norm used for certification, the parameters of these distributions are always set as global hyper parameters independent from the input data on which a network is certified. In this work, we revisit Gaussian randomized smoothing and show that the variance of the Gaussian distribution can be optimized at each input so as to maximize the certification radius for the construction of the smooth classifier. Since the data dependent classifier does not directly enjoy sound certification with existing approaches, we propose a memory-enhanced data dependent smooth classifier that is certifiable by construction. This new approach is generic, parameter-free, and easy to implement. In fact, we show that our data dependent framework can be seamlessly incorporated into 3 randomized smoothing approaches, leading to consistent improved certified accuracy. When this framework is used in the training routine of these approaches followed by a data dependent certification, we achieve 9% and 6% improvement over the certified accuracy of the strongest baseline for a radius of 0.5 on CIFAR10 and ImageNet.

Bayesian neural networks that incorporate data augmentation implicitly use a “randomly perturbed log-likelihood [which] does not have a clean interpretation as a valid likelihood function” (Izmailov et al. 2021). Here, we provide several approaches to developing principled Bayesian neural networks incorporating data augmentation. We introduce a “finite orbit” setting which allows valid likelihoods to be computed exactly, and for the more usual “full orbit” setting we derive multi-sample bounds tighter than those used previously. These models cast light on the origin of the cold posterior effect. In particular, we find that the cold posterior effect persists even in these principled models incorporating data augmentation. This suggests that the cold posterior effect cannot be dismissed as an artifact of data augmentation using incorrect likelihoods.