AISTATS 2023 Papers

The lasso is the most famous sparse regression and feature selection method. One reason for its popularity is the speed at which the underlying optimization problem can be solved. Sorted L-One Penalized Estimation (SLOPE) is a generalization of the lasso with appealing statistical properties. In spite of this, the method has not yet reached widespread interest. A major reason for this is that current software packages that fit SLOPE rely on algorithms that perform poorly in high dimensions. To tackle this issue, we propose a new fast algorithm to solve the SLOPE optimization problem, which combines proximal gradient descent and proximal coordinate descent steps. We provide new results on the directional derivative of the SLOPE penalty and its related SLOPE thresholding operator, as well as provide convergence guarantees for our proposed solver. In extensive benchmarks on simulated and real data, we demonstrate our method’s performance against a long list of competing algorithms.

We revisit the domain of off-policy policy optimization in RL from the perspective of coordinate ascent. One commonly-used approach is to leverage the off-policy policy gradient to optimize a surrogate objective – the total discounted in expectation return of the target policy with respect to the state distribution of the behavior policy. However, this approach has been shown to suffer from the distribution mismatch issue, and therefore significant efforts are needed for correcting this mismatch either via state distribution correction or a counterfactual method. In this paper, we rethink off-policy learning via Coordinate Ascent Policy Optimization (CAPO), an off-policy actor-critic algorithm that decouples policy improvement from the state distribution of the behavior policy without using the policy gradient. This design obviates the need for distribution correction or importance sampling in the policy improvement step of off-policy policy gradient. We establish the global convergence of CAPO with general coordinate selection and then further quantify the convergence rates of several instances of CAPO with popular coordinate selection rules, including the cyclic and the randomized variants of CAPO. We then extend CAPO to neural policies for a more practical implementation. Through experiments, we demonstrate that CAPO provides a competitive approach to RL in practice.

Inverse decision theory (IDT) aims to learn a performance metric for classification by eliciting expert classifications on examples. However, elicitation in practical settings may require many classifications of potentially ambiguous examples. To improve the efficiency of elicitation, we propose the cooperative inverse decision theory (CIDT) framework as a formalization of the performance metric elicitation problem. In cooperative inverse decision theory, the expert and a machine play a game where both are rewarded according to the expert’s performance metric, but the machine does not initially know what this function is. We show that optimal policies in this framework produce active learning that leads to an exponential improvement in sample complexity over previous work. One of our key findings is that a broad class of sub-optimal experts can be represented as having uncertain preferences. We use this finding to show such experts naturally fit into our proposed framework extending inverse decision theory to efficiently deal with decision data that is sub-optimal due to noise, conflicting experts, or systematic error.

While deep neural networks are proven to be effective learning systems, their analysis is complex due to the high-dimensionality of their weight space. Persistent topological properties can be used as an additional descriptor, providing insights on how the network weights evolve during training. In this paper, we focus on convolutional neural networks, and define the topology of the space, populated by convolutional filters (i.e., kernels). We perform an extensive analysis of topological properties of the convolutional filters. Specifically, we define a metric based on persistent homology, namely, Convolutional Topology Representation, to determine an important factor in neural networks training: the generalizability of the model to the test set. We further analyse how various training methods affect the topology of convolutional layers.

The softmax function is a ubiquitous component at the output of neural networks and increasingly in intermediate layers as well. This paper provides convex lower bounds and concave upper bounds on the softmax function, which are compatible with convex optimization formulations for characterizing neural networks and other ML models. We derive bounds using both a natural exponential-reciprocal decomposition of the softmax as well as an alternative decomposition in terms of the log-sum-exp function. The new bounds are provably and/or numerically tighter than linear bounds obtained in previous work on robustness verification of transformers. As illustrations of the utility of the bounds, we apply them to verification of transformers as well as of the robustness of predictive uncertainty estimates of deep ensembles.

Stein Variational Gradient Descent (SVGD) is an important alternative to the Langevin-type algorithms for sampling from probability distributions of the form $\pi(x) \propto \exp(-V(x))$. In the existing theory of Langevin-type algorithms and SVGD, the potential function $V$ is often assumed to be $L$-smooth. However, this restrictive condition excludes a large class of potential functions such as polynomials of degree greater than $2$. Our paper studies the convergence of the SVGD algorithm for distributions with $(L_0,L_1)$-smooth potentials. This relaxed smoothness assumption was introduced by Zhang et al. [2019a] for the analysis of gradient clipping algorithms. With the help of trajectory-independent auxiliary conditions, we provide a descent lemma establishing that the algorithm decreases the KL divergence at each iteration and prove a complexity bound for SVGD in the population limit in terms of the Stein Fisher information.

Offline reinforcement learning (RL) is a promising approach for training intelligent medical agents to learn treatment policies and assist decision making in many healthcare applications, such as scheduling clinical visits and assigning dosages for patients with chronic conditions. In this paper, we investigate the potential usefulness of Decision Transformer (Chen et al., 2021)–a new offline RL paradign– in medical domains where decision making in continuous time is desired. As Decision Transformer only handles discrete-time (or turn-based) sequential decision making scenarios, we generalize it to Continuous-Time Decision Transformer that not only considers the past clinical measurements and treatments but also the timings of previous visits, and learns to suggest the timings of future visits as well as the treatment plan at each visit. Extensive experiments on synthetic datasets and simulators motivated by real-world medical applications demonstrate that Continuous-Time Decision Transformer is able to outperform competitors and has clinical utility in terms of improving patients’ health and prolonging their survival by learning high-performance policies from logged data generated using policies of different levels of quality.

We study contextual linear bandit problems under feature uncertainty; they are noisy with missing entries. To address the challenges of the noise, we analyze Bayesian oracles given observed noisy features. Our Bayesian analysis finds that the optimal hypothesis can be far from the underlying realizability function, depending on the noise characteristics, which are highly non-intuitive and do not occur for classical noiseless setups. This implies that classical approaches cannot guarantee a non-trivial regret bound. Therefore, we propose an algorithm that aims at the Bayesian oracle from observed information under this model, achieving $\tilde{O}(d\sqrt{T})$ regret bound when there is a large number of arms. We demonstrate the proposed algorithm using synthetic and real-world datasets.

Causal discovery algorithms aim at untangling complex causal relationships from data. Here, we study causal discovery and inference methods based on staged tree models, which can represent complex and asymmetric causal relationships between categorical variables. We provide a first graphical representation of the equivalence class of a staged tree, by looking only at a specific subset of its underlying independences. We further define a new pre-metric, inspired by the widely used structural intervention distance, to quantify the closeness between two staged trees in terms of their corresponding causal inference statements. A simulation study highlights the efficacy of staged trees in uncovering complexes, asymmetric causal relationships from data, and real-world data applications illustrate their use in practical causal analysis.

In contrast to ordinary supervised classification tasks that require massive data with high-quality labels, complementary-label learning (CLL) deals with the weakly-supervised learning scenario where each instance is equipped with a complementary label, which specifies a class the instance does not belong to. However, most of the existing statistically consistent CLL methods suffer from overfitting intrinsically, due to the negative empirical risk issue. In this paper, we aim to propose overfitting-resistant and theoretically grounded methods for CLL. Considering the unique property of the distribution of complementarily labeled samples, we provide a risk estimator via order-preserving losses, which are naturally non-negative and thus can avoid overfitting caused by negative terms in risk estimators. Moreover, we provide classifier-consistency analysis and statistical guarantee for this estimator. Furthermore, we provide a weighed version of the proposed risk estimator to further enhance its generalization ability and prove its statistical consistency. Experiments on benchmark datasets demonstrate the effectiveness of our proposed methods.

Causal discovery methods typically extract causal relations between multiple nodes (variables) based on univariate observations of each node. However, one frequently encounters situations where each node is multivariate, i.e. has multiple observational modalities. Furthermore, the observed modalities may be generated through an unknown mixing process, so that some original latent variables are entangled inside the nodes. In such a multimodal case, the existing frameworks cannot be applied. To analyze such data, we propose a new causal representation learning framework called connectivity-contrastive learning (CCL). CCL disentangles the observational mixing and extracts a set of mutually independent latent components, each having a separate causal structure between the nodes. The actual learning proceeds by a novel self-supervised learning method in which the pretext task is to predict the label of a pair of nodes from the observations of the node pairs. We present theorems which show that CCL can indeed identify both the latent components and the multimodal causal structure under weak technical assumptions, up to some indeterminacy. Finally, we experimentally show its superior causal discovery performance compared to state-of-the-art baselines, in particular demonstrating robustness against latent confounders.

One of the training strategies of generative models is to minimize the Jensen–Shannon divergence between the model distribution and the data distribution. Since data distribution is unknown, generative adversarial networks (GANs) formulate this problem as a game between two models, a generator and a discriminator. The training can be formulated in the context of game theory and the local Nash equilibrium (LNE). It does not seem feasible to derive guarantees of stability or optimality for the existing methods. This optimization problem is far more challenging than the single objective setting. Here, we use the conjugate gradient method to reliably and efficiently solve the LNE problem in GANs. We give a proof and convergence analysis under mild assumptions showing that the proposed method converges to a LNE with three different learning rate update rules, including a constant learning rate. Finally, we demonstrate that the proposed method outperforms stochastic gradient descent (SGD) and momentum SGD in terms of best Frechet inception distance (FID) score and outperforms Adam on average.

We develop a predictive inference procedure that combines conformal prediction (CP) with unconditional quantile regression (QR)-a commonly used tool in econometrics that involves regressing the recentered influence function (RIF) of the quantile functional over input covariates. Unlike the more widely known conditional QR, unconditional QR explicitly captures the impact of changes in covariate distribution on the quantiles of the marginal distribution of outcomes. Leveraging this property, our procedure issues adaptive predictive intervals with localized frequentist coverage guarantees. It operates by fitting a machine learning model for the RIFs using training data, and then applying the CP procedure for any test covariate with respect to a “hypothetical” covariate distribution localized around the new instance. Experiments show that our procedure is adaptive to heteroscedasticity, provides transparent coverage guarantees that are relevant to the test instance at hand, and performs competitively with existing methods in terms of efficiency.

Off-policy evaluation is critical in a number of applications where new policies need to be evaluated offline before online deployment. Most existing methods focus on the expected return, define the target parameter through averaging and provide a point estimator only. In this paper, we develop a novel procedure to produce reliable interval estimators for a target policy’s return starting from any initial state. Our proposal accounts for the variability of the return around its expectation, focuses on the individual effect and offers valid uncertainty quantification. Our main idea lies in designing a pseudo policy that generates subsamples as if they were sampled from the target policy so that existing conformal prediction algorithms are applicable to prediction interval construction. Our methods are justified by theories, synthetic data and real data from short-video platforms.

Boosted trees is a dominant ML model, exhibiting high accuracy. However, boosted trees are hardly intelligible, and this is a problem whenever they are used in safety-critical applications. Indeed, in such a context, provably sound explanations for the predictions made are expected. Recent work have shown how subset-minimal abductive explanations can be derived for boosted trees, using automated reasoning techniques. However, the generation of such well-founded explanations is intractable in the general case. To improve the scalability of their generation, we introduce the notion of tree-specific explanation for a boosted tree. We show that tree-specific explanations are provably sound abductive explanations that can be computed in polynomial time. We also explain how to derive a subset-minimal abductive explanation from a tree-specific explanation. Experiments on various datasets show the computational benefits of leveraging tree-specific explanations for deriving subset-minimal abductive explanations.

Kernel two-sample testing provides a powerful framework for distinguishing any pair of distributions based on n sample points. However, existing kernel tests either run in $n^2$ time or sacrifice undue power to improve runtime. To address these shortcomings, we introduce Compress Then Test (CTT), a new framework for high-powered kernel testing based on sample compression. CTT cheaply approximates an expensive test by compressing each n point sample into a small but provably high-fidelity coreset. For standard kernels and subexponential distributions, CTT inherits the statistical behavior of a quadratic-time test—recovering the same optimal detection boundary—while running in near-linear time. We couple these advances with cheaper permutation testing, justified by new power analyses; improved time-vs.-quality guarantees for low-rank approximation; and a fast aggregation procedure for identifying especially discriminating kernels. In our experiments with real and simulated data, CTT and its extensions provide 20–200x speed-ups over state-of-the-art approximate MMD tests with no loss of power.

Probabilistic circuits (PCs) are a class of tractable probabilistic models, which admit efficient inference routines depending on their structural properties. In this paper, we introduce md-vtrees, a novel structural formulation of (marginal) determinism in structured decomposable PCs, which generalizes previously proposed classes such as probabilistic sentential decision diagrams. Crucially, we show how md-vtrees can be used to derive tractability conditions and efficient algorithms for advanced inference queries expressed as arbitrary compositions of basic probabilistic operations, such as marginalization, multiplication and reciprocals, in a sound and generalizable manner. In particular, we derive the first polytime algorithms for causal inference queries such as backdoor adjustment on PCs. As a practical instantiation of the framework, we propose MDNets, a novel PC architecture using md-vtrees, and empirically demonstrate their application to causal inference.

Randomized sketches of a tensor product of $p$ vectors follow a tradeoff between statistical efficiency and computational acceleration. Commonly used approaches avoid computing the high-dimensional tensor product explicitly, resulting in a suboptimal dependence of $O(3^p)$ in the embedding dimension. We propose a simple Complex-to-Real (CtR) modification of well-known sketches that replaces real random projections by complex ones, incurring a lower $O(2^p)$ factor in the embedding dimension. The output of our sketches is real-valued, which renders their downstream use straightforward. In particular, we apply our sketches to $p$-fold self-tensored inputs corresponding to the feature maps of the polynomial kernel. We show that our method achieves state-of-the-art performance in terms of accuracy and speed compared to other randomized approximations from the literature.

For many of the classic online learning settings, it is known that having a “hint” about the loss function before making a prediction yields significantly better regret guarantees. In this work we study the question, do hints allow us to go beyond the standard notion of regret (which competes against the best fixed strategy) and compete against adaptive or dynamic strategies? After all, if hints were perfect, we can clearly compete against a fully dynamic strategy. For some common online learning settings, we provide upper and lower bounds for the switching regret, i.e., the difference between the loss incurred by the algorithm and the optimal strategy in hindsight that switches state at most $L$ times, where $L$ is some parameter. We show positive results for online linear optimization and the classic experts problem. Interestingly, such results turn out to be impossible for the classic bandit setting.