UAI 2022 Papers

Unsupervised domain adaptation (UDA) aims to transfer knowledge from a model trained on a labeled source domain to an unlabeled target domain. To this end, we propose in this paper a novel cycle class-consistent model based on optimal transport (OT) and knowledge distillation. The model consists of two agents, a teacher and a student cooperatively working in a cycle process under the guidance of the distributional optimal transport and distillation manner. The OT distance is designed to bridge the gap between the distribution of the target data and a distribution over the source class-conditional distributions. The optimal probability matrix then provides pseudo labels to learn a teacher that achieves a good classification performance on the target domain. Knowledge distillation is performed in the next step in which the teacher distills and transfers its knowledge to the student. And finally, the student produces its prediction for the optimal transport step. This process forms a closed cycle in which the teacher and student networks are simultaneously trained to conduct transfer learning from the source to the target domain. Extensive experiments show that our proposed method outperforms existing methods, especially the class-aware and OT-based ones on benchmark datasets including Office-31, Office-Home, and ImageCLEF-DA.

Despite the impressive success achieved in various domains, deep reinforcement learning (DRL) is still faced with the sample inefficiency problem. Transfer learning (TL), which leverages prior knowledge from different but related tasks to accelerate the target task learning, has emerged as a promising direction to improve RL efficiency. The majority of prior work considers TL across tasks with the same state-action spaces, while transferring across domains with different state-action spaces is relatively unexplored. Furthermore, such existing cross-domain transfer approaches only enable transfer from a single source policy, leaving open the important question of how to best transfer from multiple source policies. This paper proposes a novel framework called Cross-domain Adaptive Transfer (CAT) to accelerate DRL. CAT learns the state-action correspondence from each source task to the target task and adaptively transfers knowledge from multiple source task policies to the target policy. CAT can be easily combined with existing DRL algorithms and experimental results show that CAT significantly accelerates learning and outperforms other cross-domain transfer methods on multiple continuous action control tasks.

Automated decision support systems that are able to infer second opinions from experts can potentially facilitate a more efficient allocation of resources—they can help decide when and from whom to seek a second opinion. In this paper, we look at the design of this type of support systems from the perspective of counterfactual inference. We focus on a multiclass classification setting and first show that, if experts make predictions on their own, the underlying causal mechanism generating their predictions needs to satisfy a desirable set invariant property. Further, we show that, for any causal mechanism satisfying this property, there exists an equivalent mechanism where the predictions by each expert are generated by independent sub-mechanisms governed by a common noise. This motivates the design of a set invariant Gumbel-Max structural causal model where the structure of the noise governing the sub-mechanisms underpinning the model depends on an intuitive notion of similarity between experts which can be estimated from data. Experiments on both synthetic and real data show that our model can be used to infer second opinions more accurately than its non-causal counterpart.

Model interpretability, fairness, and recourse for end users have increased as machine learning models have become increasingly popular in areas including criminal justice, finance, healthcare, and job marketplaces. This work presents a novel method of addressing these issues by producing meaningful counterfactuals that are aimed at providing recourse to users and highlighting potential model biases. A meaningful counterfactual is a reasonable alternative scenario that illustrates how input data perturbations can influence the model’s output. The CounteRGAN method generates meaningful counterfactuals for a target classifier by utilizing a novel Residual Generative Adversarial Network (RGAN). We compare our method against leading state-of-the-art approaches on image and tabular datasets over a variety of performance metrics. The results indicate a significant improvement over existing techniques in combined metric performance, with a latency reduction of 2 to 7 orders of magnitude which enables providing real-time recourse to users. The code for reproducibility can be found here: https://github.com/gan-counterfactuals/countergan.

Linear coupling is recently proposed to accelerate first-order algorithms by linking gradient descent and mirror descent together, which is able to achieve the accelerated convergence rate for first-order algorithms. This work focuses on the convergence analysis of linear coupling for convex composite minimization when the proximal operator cannot be exactly computed. It is of particular interest to study the convergence of linear coupling because it not only achieves the accelerated convergence rate for first-order algorithm but also works for generic norms. We present convergence analysis of linear coupling by allowing the proximal operator to be computed up to a certain precision. Our analysis illustrates that the accelerated convergence rate of linear coupling with inexact proximal operator can be preserved if the error sequence of inexact proximal operator decreases in a sufficiently fast rate. More importantly, our analysis leads to better bounds than existing works on inexact proximal operator. Experiment results on several real-world datasets verify our theoretical results.

Deep latent variable models such as variational autoencoders and energy-based models are widely used for neural text generation. Most of them focus on matching the prior distribution with the posterior distribution of the latent variable for text reconstruction. In addition to instance-level reconstruction, this paper aims to integrate contrastive learning in the latent space, forcing the latent variables to learn high-level semantics by exploring inter-instance relationships. Experiments on various text generation benchmarks show the effectiveness of our proposed method. We also empirically show that our method can mitigate the posterior collapse issue for latent variable based text generation models.

Denoising diffusion models have recently emerged as a powerful class of generative models. They provide state-of-the-art results, not only for unconditional simulation, but also when used to solve conditional simulation problems arising in a wide range of inverse problems. A limitation of these models is that they are computationally intensive at generation time as they require simulating a diffusion process over a long time horizon. When performing unconditional simulation, a Schr{ö}dinger bridge formulation of generative modeling leads to a theoretically grounded algorithm shortening generation time which is complementary to other proposed acceleration techniques. We extend the Schrödinger bridge framework to conditional simulation. We demonstrate this novel methodology on various applications including image super-resolution, optimal filtering for state-space models and the refinement of pre-trained networks. Our code can be found at https://github.com/vdeborto/cdsb.

Bayesian optimization is a popular method for solving the problem of global optimization of an expensive-to-evaluate black-box function. It relies on a probabilistic surrogate model of the objective function, upon which an acquisition function is built to determine where next to evaluate the objective function. In general, Bayesian optimization with Gaussian process regression operates on a continuous space. When input variables are categorical or discrete, an extra care is needed. A common approach is to use one-hot encoded or Boolean representation for categorical variables which might yield a combinatorial explosion problem. In this paper we present a method for Bayesian optimization in a combinatorial space, which can operate well in a large combinatorial space. The main idea is to use a random mapping which embeds the combinatorial space into a convex polytope in a continuous space, on which all essential process is performed to determine a solution to the black-box optimization in the combinatorial space. We describe our combinatorial Bayesian optimization algorithm and present its regret analysis. Numerical experiments demonstrate that our method shows satisfactory performance compared to existing methods.

Notable progress has been made in numerous fields of machine learning based on neural network-driven mutual information (MI) bounds. However, utilizing the conventional MI-based losses is often challenging due to their practical and mathematical limitations. In this work, we first identify the symptoms behind their instability: (1) the neural network not converging even after the loss seemed to converge, and (2) saturating neural network outputs causing the loss to diverge. We mitigate both issues by adding a novel regularization term to the existing losses. We theoretically and experimentally demonstrate that added regularization stabilizes training. Finally, we present a novel benchmark that evaluates MI-based losses on both the MI estimation power and its capability on the downstream tasks, closely following the pre-existing supervised and contrastive learning settings. We evaluate six different MI-based losses and their regularized counterparts on multiple benchmarks to show that our approach is simple yet effective.

Existing BERT-based models for Chinese spelling correction (CSC) have three issues. 1) Bert tends to rectify a correct low-frequency collocation into a high-frequency one and leads to over-correcting. 2) It fails to completely detect phonic or morphological errors by the current learned similarity knowledge between Chinese characters, and the recall rate still has room to improve. 3) Two-dimensional glyph information of Chinese characters is overlooked and some morphological misused characters may be difficult to detect. This paper proposes a hybrid approach, CoSPA, to address these issues. 1) This paper proposes an alterable copy mechanism to alleviate over-correcting by jointly learning to copy a correct character from input sentence, or generate a character from BERT. No method has used copy mechanism in BERT for CSC. 2) The attention mechanism is further applied on the phonic and shape representation of each character at the output layer. 3) Shape representation is enhanced by mining character glyph with ResNet, and fused with stroke representation via an adaptive gating unit. The experimental results show that CoSPA outperforms the previous state-of-the-art methods on SIGHAN2015 datasets.

This paper focuses on the Matrix Factorization based Clustering (MFC) method which is one of the few closed-form algorithms for the subspace clustering algorithm. Despite being simple, closed-form, and computation-efficient, MFC can outperform the other sophisticated subspace clustering methods in many challenging scenarios. We reveal the connection between MFC and the Innovation Pursuit (iPursuit) algorithm which was shown to be able to outperform the other spectral clustering based methods with a notable margin especially when the span of clusters are close. A novel theoretical study is presented which sheds light on the key performance factors of both algorithms (MFC/iPursuit) and it is shown that both algorithms can be robust to notable intersections between the span of clusters. Importantly, in contrast to the theoretical guarantees of other algorithms which emphasized on the distance between the subspaces as the key performance factor and without making the innovation assumption, it is shown that the performance of MFC/iPursuit mainly depends on the distance between the innovative components of the clusters.

We propose a kernel-based nonparametric estimator for the causal effect when the cause is corrupted by error. We do so by generalizing estimation in the instrumental variable setting. Despite significant work on regression with measurement error, additionally handling unobserved confounding in the continuous setting is non-trivial: we have seen little prior work. As a by-product of our investigation, we clarify a connection between mean embeddings and characteristic functions, and how learning one simultaneously allows one to learn the other. This opens the way for kernel method research to leverage existing results in characteristic function estimation. Finally, we empirically show that our proposed method, MEKIV, improves over baselines and is robust under changes in the strength of measurement error and to the type of error distributions.

Despite the increasing relevance of forecasting methods, causal implications of these algorithms remain largely unexplored. This is concerning considering that, even under simplifying assumptions such as causal sufficiency, the statistical risk of a model can differ significantly from its causal risk. Here, we study the problem of causal generalization—generalizing from the observational to interventional distributions—in forecasting. Our goal is to find answers to the question: How does the efficacy of an autoregressive (VAR) model in predicting statistical associations compare with its ability to predict under interventions? To this end, we introduce the framework of causal learning theory for forecasting. Using this framework, we obtain a characterization of the difference between statistical and causal risks, which helps identify sources of divergence between them. Under causal sufficiency, the problem of causal generalization amounts to learning under covariate shifts albeit with additional structure (restriction to interventional distributions under the VAR model). This structure allows us to obtain uniform convergence bounds on causal generalizability for the class of VAR models. To the best of our knowledge, this is the first work that provides theoretical guarantees for causal generalization in the time-series setting.

We consider the problem of causal discovery (structure learning) from heterogeneous observational data. Most existing methods assume homogeneous sampling scheme and causal mechanism, which may lead to misleading conclusions when violated. We propose a novel approach that exploits data heterogeneity to infer possibly cyclic causal structures from causally insufficient systems. The core idea is to model the direct causal effects as functions of exogenous covariates that help explain sampling and causal heterogeneity. We investigate the structure identifiability properties of the proposed model. Structure learning is carried out in a fully Bayesian fashion, which provides natural uncertainty quantification. We demonstrate its utility through extensive simulations and two real-world applications.

Inferring causal relationships from observational data is rarely straightforward, but the problem is especially difficult in high dimensions. For these applications, causal discovery algorithms typically require parametric restrictions or extreme sparsity constraints. We relax these assumptions and focus on an important but more specialized problem, namely recovering the causal order among a subgraph of variables known to descend from some (possibly large) set of confounding covariates, i.e. a $\textit{confounder blanket}$. This is useful in many settings, for example when studying a dynamic biomolecular subsystem with genetic data providing background information. Under a structural assumption called the $\textit{confounder blanket principle}$, which we argue is essential for tractable causal discovery in high dimensions, our method accommodates graphs of low or high sparsity while maintaining polynomial time complexity. We present a structure learning algorithm that is provably sound and complete with respect to a so-called $\textit{lazy oracle}$. We design inference procedures with finite sample error control for linear and nonlinear systems, and demonstrate our approach on a range of simulated and real-world datasets. An accompanying $\texttt{R}$ package, $\texttt{cbl}$, is available from $\texttt{CRAN}$.

Importance sampling (IS) is often used to perform off-policy evaluation but it is prone to several issues—especially when the behavior policy is unknown and must be estimated from data. Significant differences between target and behavior policies can result in uncertain value estimates due to, for example, high variance. Standard practices such as inspecting IS weights may be insufficient to diagnose such problems and determine for which type of inputs the policies differ in suggested actions and resulting values. To address this, we propose estimating the behavior policy for IS using prototype learning. The learned prototypes provide a condensed summary of the input-action space, which allows for describing differences between policies and assessing the support for evaluating a certain target policy. In addition, we can describe a value estimate in terms of prototypes to understand which parts of the target policy have the most impact on the estimate. We find that this provides new insights in the examination of a learned policy for sepsis management. Moreover, we study the bias resulting from restricting models to use prototypes, how bias propagates to IS weights and estimated values and how this varies with history length.

End-to-end learning of dynamical systems with black-box models, such as neural ordinary differential equations (ODEs), provides a flexible framework for learning dynamics from data without prescribing a mathematical model for the dynamics. Unfortunately, this flexibility comes at the cost of understanding the dynamical system, for which ODEs are used ubiquitously. Further, experimental data are collected under various conditions (inputs), such as treatments, or grouped in some way, such as part of sub-populations. Understanding the effects of these system inputs on system outputs is crucial to have any meaningful model of a dynamical system. To that end, we propose a structured latent ODE model that explicitly captures system input variations within its latent representation. Building on a static latent variable specification, our model learns (independent) stochastic factors of variation for each input to the system, thus separating the effects of the system inputs in the latent space. This approach provides actionable modeling through the controlled generation of time-series data for novel input combinations (or perturbations). Additionally, we propose a flexible approach for quantifying uncertainties, leveraging a quantile regression formulation. Results on challenging biological datasets show consistent improvements over competitive baselines in the controlled generation of observational data and inference of biologically meaningful system inputs.

Mean-Field Control (MFC) is a powerful tool to solve Multi-Agent Reinforcement Learning (MARL) problems. Recent studies have shown that MFC can well-approximate MARL when the population size is large and the agents are exchangeable. Unfortunately, the presumption of exchangeability implies that all agents uniformly interact with one another which is not true in many practical scenarios. In this article, we relax the assumption of exchangeability and model the interaction between agents via an arbitrary doubly stochastic matrix. As a result, in our framework, the mean-field ‘seen’ by different agents are different. We prove that, if the reward of each agent is an affine function of the mean-field seen by that agent, then one can approximate such a non-uniform MARL problem via its associated MFC problem within an error of $e=\mathcal{O}(\frac{1}{\sqrt{N}}[\sqrt{|\mathcal{X}|} + \sqrt{|\mathcal{U}|}])$ where $N$ is the population size and $|\mathcal{X}|$, $|\mathcal{U}|$ are the sizes of state and action spaces respectively. Finally, we develop a Natural Policy Gradient (NPG) algorithm that can provide a solution to the non-uniform MARL with an error $\mathcal{O}(\max\{e,\epsilon\})$ and a sample complexity of $\mathcal{O}(\epsilon^{-3})$ for any $\epsilon >0$.

We often see undesirable tradeoffs in robust machine learning where out-of-distribution (OOD) accuracy is at odds with in-distribution (ID) accuracy. A robust classifier obtained via specialized techniques such as removing spurious features often has better OOD but worse ID accuracy compared to a standard classifier trained via vanilla ERM. In this paper, we find that a simple approach of ensembling the standard and robust models, after calibrating on only ID data, outperforms prior state-of-the-art both ID and OOD. On ten natural distribution shift datasets, ID-calibrated ensembles get the best of both worlds: strong ID accuracy of the standard model and OOD accuracy of the robust model. We analyze this method in stylized settings, and identify two important conditions for ensembles to perform well on both ID and OOD: (1) standard and robust models should be calibrated (on ID data, because OOD data is unavailable), (2) OOD has no anticorrelated spurious features.