UAI 2023 Papers

Accurate outlier detection is not only a necessary preprocessing step, but can itself give important insights into the data. However, especially, for non-linear regression the detection of outliers is non-trivial, and actually ambiguous. We propose a new method that identifies outliers by finding a subset of data points T such that the marginal likelihood of all remaining data points S is maximized. Though the idea is more general, it is particular appealing for Gaussian processes regression, where the marginal likelihood has an analytic solution. While maximizing the marginal likelihood for hyper-parameter optimization is a well established non-convex optimization problem, optimizing the set of data points S is not. Indeed, even a greedy approximation is computationally challenging due to the high cost of evaluating the marginal likelihood. As a remedy, we propose an efficient projected gradient descent method with provable convergence guarantees. Moreover, we also establish the breakdown point when jointly optimizing hyper-parameters and S. For various datasets and types of outliers, our experiments demonstrate that the proposed method can improve outlier detection and robustness when compared with several popular alternatives like the student-t likelihood.

We introduce the notion of risk-limiting financial audits (RLFA): procedures that manually evaluate a subset of $N$ financial transactions to check the validity of a claimed assertion $\mathcal{A}$ about the transactions. More specifically, RLFA satisfy two properties: (i) if $\mathcal{A}$ is false, they correctly disprove it with probability at least $1-\delta$, and (ii) they validate the correctness of $\mathcal{A}$ with probability $1$, if it is true. We propose a general RLFA strategy, by constructing new confidence sequences (CSs) for the weighted average of $N$ unknown values, based on samples drawn without replacement from a (randomized) weighted sampling scheme. Next, we develop methods to improve the quality of CSs by incorporating side information about the unknown values. We show that when the side information is sufficiently accurate, it can directly drive the sampling. For the case where the accuracy is unknown

Automated curriculum generation for reinforcement learning (RL) aims to speed up learning by designing a sequence of tasks of increasing difficulty. Such tasks are usually drawn from probability distributions with exponentially bounded tails, such as uniform or Gaussian distributions. However, existing approaches overlook heavy-tailed distributions. Under such distributions, current methods may fail to learn optimal policies in rare and risky tasks, which fall under the tails and yield the lowest returns, respectively. We address this challenge by proposing a risk-aware curriculum generation algorithm that simultaneously creates two curricula: 1) a primary curriculum that aims to maximize the expected discounted return with respect to a distribution over target tasks, and an auxiliary curriculum that identifies and over-samples rare and risky tasks observed in the primary curriculum. Our empirical results evidence that the proposed algorithm achieves significantly higher returns in frequent as well as rare tasks compared to the state-of-the-art methods.

Self-paced reinforcement learning (RL) aims to improve the data efficiency of learning by automatically creating sequences, namely curricula, of probability distributions over contexts. However, existing techniques for self-paced RL fail in long-horizon planning tasks that involve temporally extended behaviors. We hypothesize that taking advantage of prior knowledge about the underlying task structure can improve the effectiveness of self-paced RL. We develop a self-paced RL algorithm guided by reward machines, i.e., a type of finite-state machine that encodes the underlying task structure. The algorithm integrates reward machines in 1) the update of the policy and value functions obtained by any RL algorithm of choice, and 2) the update of the automated curriculum that generates context distributions. Our empirical results evidence that the proposed algorithm achieves optimal behavior reliably even in cases in which existing baselines cannot make any meaningful progress. It also decreases the curriculum length and reduces the variance in the curriculum generation process by up to one-fourth and four orders of magnitude, respectively.

The problem of structure learning in Bayesian networks asks for a directed acyclic graph (DAG) that maximizes a given scoring function. Since the problem is NP-hard, research effort has been put into discovering restricted classes of DAGs for which the search problem can be solved in polynomial time. Here, we initiate investigation of questions that have received less attention thus far: Are the known polynomial algorithms close to the best possible, or is there room for significant improvements? If the interest is in Bayesian learning, that is, in sampling or weighted counting of DAGs, can we obtain similar complexity results? Focusing on DAGs with bounded vertex cover number—a class studied in Korhonen and Parviainen’s seminal work (NIPS 2015)—we answer the questions in the affirmative. We also give, apparently the first, proof that the counting problem is $#$P-hard in general. In addition, we show that under the vertex-cover constraint counting is $#$W[1]-hard.

Neural networks are universal approximators and are studied for their use in solving differential equations. However, a major criticism is the lack of error bounds for obtained solutions. This paper proposes a technique to rigorously evaluate the error bound of Physics-Informed Neural Networks (PINNs) on most linear ordinary differential equations (ODEs), certain nonlinear ODEs, and first-order linear partial differential equations (PDEs). The error bound is based purely on equation structure and residual information and does not depend on assumptions of how well the networks are trained. We propose algorithms that bound the error efficiently. Some proposed algorithms provide tighter bounds than others at the cost of longer run time.

The utilization of online optimization techniques is prevalent in many fields of artificial intelligence, enabling systems to continuously learn and adjust to their surroundings. This paper outlines a regularized online optimization problem, where the regularizer is defined on the average of the actions taken. The objective is to maximize the sum of rewards and the regularizer value while adhering to resource constraints, where the reward function is assumed to be DR-submodular. Both concave and DR-submodular regularizers are analyzed. Concave functions are useful in describing the impartiality of decisions, while DR-submodular functions can be employed to represent the overall effect of decisions on all relevant parties. We have developed two algorithms for each of the concave and DR-submodular regularizers. These algorithms are easy to implement, efficient, and produce sublinear regret in both cases. The performance of the proposed algorithms and regularizers has been verified through numerical experiments in the context of internet advertising.

Virtually all state-of-the-art methods for training supervised machine learning models are variants of Stochastic Gradient Descent (SGD), enhanced with a number of additional tricks, such as minibatching, momentum, and adaptive stepsizes. However, one of the most basic questions in the design of successful SGD methods, one that is orthogonal to the aforementioned tricks, is the choice of the next training data point to be learning from. Standard variants of SGD employ a sampling with replacement strategy, which means that the next training data point is sampled from the entire data set, often independently of all previous samples. While standard SGD is well understood theoretically, virtually all widely used machine learning software is based on sampling without replacement as this is often empirically superior. That is, the training data is randomly shuffled/permuted, either only once at the beginning, strategy known as random shuffling (RS), or before every epoch, strategy known as random reshuffling (RR), and training proceeds in the data order dictated by the shuffling. RS and RR strategies have for a long time remained beyond the reach of theoretical analysis that would satisfactorily explain their success. However, very recently, Mishchenko et al. [2020] provided tight sublinear convergence rates through a novel analysis, and showed that these strategies can improve upon standard SGD in certain regimes. Inspired by these results, we seek to further improve the rates of shuffling-based methods. In particular, we show that it is possible to enhance them with a variance reduction mechanism, obtaining linear convergence rates. To the best of our knowledge, our linear convergence rates are the best for any method based on sampling without replacement.

Dataset condensation aims to condense the original training dataset into a small synthetic dataset for data-efficient learning. The recently proposed dataset condensation techniques allow the model trainers with limited resources to learn acceptable deep learning models on a small amount of synthetic data. However, in an adversarial environment, given the original dataset as a poisoned dataset, dataset condensation may encode the poisoning information into the condensed synthetic dataset. To explore the vulnerability of dataset condensation to data poisoning, we revisit the state-of-the-art targeted data poisoning method and customize a targeted data poisoning algorithm for dataset condensation. By executing the two poisoning methods, we demonstrate that, when the synthetic dataset is condensed from a poisoned dataset, the models trained on the synthetic dataset may predict the targeted sample as the attack-targeted label. To defend against data poisoning, we introduce the concept of poisoned deviation to quantify the poisoning effect. We further propose a poisoning-resilient dataset condensation algorithm with a calibration method to reduce poisoned deviation. Extensive evaluations demonstrate that our proposed algorithm can protect the synthetic dataset from data poisoning with minor performance drop.

Bayesian methods are a popular choice for statistical inference in small-data regimes due to the regularization effect induced by the prior. %, which serves to counteract overfitting. In the context of density estimation, the standard nonparametric Bayesian approach is to target the posterior predictive of the Dirichlet process mixture model. In general, direct estimation of the posterior predictive is intractable and so methods typically resort to approximating the posterior distribution as an intermediate step. The recent development of quasi-Bayesian predictive copula updates, however, has made it possible to perform tractable predictive density estimation without the need for posterior approximation. Although these estimators are computationally appealing, they struggle on non-smooth data distributions. This is due to the comparatively restrictive form of the likelihood models from which the proposed copula updates were derived. To address this shortcoming, we consider a Bayesian nonparametric model with an autoregressive likelihood decomposition and a Gaussian process prior. While the predictive update of such a model is typically intractable, we derive a quasi-Bayesian update that achieves state-of-the-art results in small-data regimes.

Pruning deep neural networks is a widely used strategy to alleviate the computational burden in machine learning. Overwhelming empirical evidence suggests that pruned models retain very high accuracy even with a tiny fraction of parameters. However, relatively little work has gone into characterising the small pruned networks obtained, beyond a measure of their accuracy. In this paper, we use the sparse double descent approach to identify univocally and characterise pruned models associated with classification tasks. We observe empirically that, for a given task, iterative magnitude pruning (IMP) tends to converge to networks of comparable sizes even when starting from full networks with sizes ranging over orders of magnitude. We analyse the best pruned models in a controlled experimental setup and show that their number of parameters reflects task difficulty and that they are much better than full networks at capturing the true conditional probability distribution of the labels. On real data, we similarly observe that pruned models are less prone to overconfident predictions. Our results suggest that pruned models obtained via IMP not only have advantageous computational properties but also provide a better representation of uncertainty in learning.

The quantification of aleatoric and epistemic uncertainty in terms of conditional entropy and mutual information, respectively, has recently become quite common in machine learning. While the properties of these measures, which are rooted in information theory, seem appealing at first glance, we identify various incoherencies that call their appropriateness into question. In addition to the measures themselves, we critically discuss the idea of an additive decomposition of total uncertainty into its aleatoric and epistemic constituents. Experiments across different computer vision tasks support our theoretical findings and raise concerns about current practice in uncertainty quantification.

The success of deep reinforcement learning (DRL) lies in its ability to learn a representation that is well-suited for the exploration and exploitation task. To understand how the choice of representation can improve the efficiency of reinforcement learning (RL), we study representation selection for a class of low-rank Markov Decision Processes (MDPs) where the transition kernel can be represented in a bilinear form. We propose an efficient algorithm, called ReLEX, for representation learning in both online and offline RL. Specifically, we show that the online version of ReLEX, called ReLEX-UCB, always performs no worse than the state-of-the-art algorithm without representation selection, and achieves a strictly better constant regret if the representation function class has a "coverage" property over the entire state-action space. For the offline counterpart, ReLEX-LCB, we show that the algorithm can find the optimal policy if the representation class can cover the state-action space and achieves gap-dependent sample complexity. This is the first result with constant sample complexity for representation learning in offline RL.

Imitation learning (IL) has proven to be an effective method for learning good policies from expert demonstrations. Adversarial imitation learning (AIL), a subset of IL methods, is particularly promising, but its theoretical foundation in the presence of unknown transitions has yet to be fully developed. This paper explores the theoretical underpinnings of AIL in this context, where the stochastic and uncertain nature of environment transitions presents a challenge. We examine the expert sample complexity and interaction complexity required to recover good policies. To this end, we establish a framework connecting reward-free exploration and AIL, and propose an algorithm, MB-TAIL, that achieves the minimax optimal expert sample complexity of $\widetilde{\mathcal{O}} (H^{3/2} |\mathcal{S}|/\varepsilon)$ and interaction complexity of $\widetilde{\mathcal{O}} (H^{3} |\mathcal{S}|^2 |\mathcal{A}|/\varepsilon^2)$. Here, $H$ represents the planning horizon, $|\mathcal{S}|$ is the state space size, $|\mathcal{A}|$ is the action space size, and $\varepsilon$ is the desired imitation gap. MB-TAIL is the first algorithm to achieve this level of expert sample complexity in the unknown transition setting and improves upon the interaction complexity of the best-known algorithm, OAL, by $\mathcal{O} (H)$. Additionally, we demonstrate the generalization ability of MB-TAIL by extending it to the function approximation setting and proving that it can achieve expert sample and interaction complexity independent of $|\mathcal{S}|$.

Conformal prediction (CP) is a framework to quantify uncertainty of machine learning classifiers including deep neural networks. Given a testing example and a trained classifier, CP produces a prediction set of candidate labels with a user-specified coverage (i.e., true class label is contained with high probability). Almost all the existing work on CP assumes clean testing data and there is not much known about the robustness of CP algorithms w.r.t natural/adversarial perturbations to testing examples. This paper studies the problem of probabilistically robust conformal prediction (PRCP) which ensures robustness to most perturbations around clean input examples. PRCP generalizes the standard CP (cannot handle perturbations) and adversarially robust CP (ensures robustness w.r.t worst-case perturbations) to achieve better trade-offs between nominal performance and robustness. We propose a novel adaptive PRCP (aPRCP) algorithm to achieve probabilistically robust coverage. The key idea behind aPRCP is to determine two parallel thresholds, one for data samples and another one for the perturbations on data (aka "quantile-of-quantile” design). We provide theoretical analysis to show that aPRCP algorithm achieves robust coverage. Our experiments on CIFAR-10, CIFAR-100, and ImageNet datasets using deep neural networks demonstrate that aPRCP achieves better trade-offs than state-of-the-art CP and adversarially robust CP algorithms.

Probabilistic circuits (PCs) are models that allow exact and tractable probabilistic inference. In contrast to neural networks, they are often assumed to be well-calibrated and robust to out-of-distribution (OOD) data. In this paper, we show that PCs are in fact not robust to OOD data, i.e., they don’t know what they don’t know. We then show how this challenge can be overcome by model uncertainty quantification. To this end, we propose tractable dropout inference (TDI), an inference procedure to estimate uncertainty by deriving an analytical solution to Monte Carlo dropout (MCD) through variance propagation. Unlike MCD in neural networks, which comes at the cost of multiple network evaluations, TDI provides tractable sampling-free uncertainty estimates in a single forward pass. TDI improves the robustness of PCs to distribution shift and OOD data, demonstrated through a series of experiments evaluating the classification confidence and uncertainty estimates on real-world data.

Multi-dimensional classification (MDC) can be employed in a range of applications where one needs to predict multiple class variables for each given instance. Many existing MDC methods suffer from at least one of inaccuracy, scalability, limited use to certain types of data, hardness of interpretation or lack of probabilistic (uncertainty) estimations. This paper is an attempt to address all these disadvantages simultaneously. We propose a formal framework for probabilistic MDC in which learning an optimal multi-dimensional classifier can be decomposed, without loss of generality, into learning a set of (smaller) single-variable multi-class probabilistic classifiers and a directed acyclic graph. Current and future developments of both probabilistic classification and graphical model learning can directly enhance our framework, which is flexible and provably optimal. A collection of experiments is conducted to highlight the usefulness of this MDC framework.

We consider the problem of increasing the expressivity of probabilistic circuits by augmenting them with the successful generative models of normalizing flows. To this effect, we theoretically establish the requirement of decomposability for such combinations to retain tractability of the learned models. Our model, called Probabilistic Flow Circuits, essentially extends circuits by allowing for normalizing flows at the leaves. Our empirical evaluation clearly establishes the expressivity and tractability of this new class of probabilistic circuits.

Most existing approaches of differentially private (DP) machine learning focus on