AISTATS 2024 论文列表

Domain adversarial adaptation in a continual setting poses significant challenges due to the limitations of accessing previous source domain data. Despite extensive research in continual learning, adversarial adaptation cannot be effectively accomplished using only a small number of stored source domain data, a standard setting in memory replay approaches. This limitation arises from the erroneous empirical estimation of $\mathcal{H}$-divergence with few source domain samples. To tackle this problem, we propose a double-head discriminator algorithm by introducing an addition source-only domain discriminator trained solely on the source learning phase. We prove that by introducing a pre-trained source-only domain discriminator, the empirical estimation error of $\mathcal{H}$-divergence related adversarial loss is reduced from the source domain side. Further experiments on existing domain adaptation benchmarks show that our proposed algorithm achieves more than 2$%$ improvement on all categories of target domain adaptation tasks while significantly mitigating the forgetting of the source domain.

Estimating the structure of directed acyclic graphs (DAGs) from observational data remains a significant challenge in machine learning. Most research in this area concentrates on learning a single DAG for the entire population. This paper considers an alternative setting where the graph structure varies across individuals based on available "contextual" features. We tackle this contextual DAG problem via a neural network that maps the contextual features to a DAG, represented as a weighted adjacency matrix. The neural network is equipped with a novel projection layer that ensures the output matrices are sparse and satisfy a recently developed characterization of acyclicity. We devise a scalable computational framework for learning contextual DAGs and provide a convergence guarantee and an analytical gradient for backpropagating through the projection layer. Our experiments suggest that the new approach can recover the true context-specific graph where existing approaches fail.

Contextual bandit algorithms are commonly used in digital health to recommend personalized treatments. However, to ensure the effectiveness of the treatments, patients are often requested to take actions that have no immediate benefit to them, which we refer to as pro-treatment actions. In practice, clinicians have a limited budget to encourage patients to take these actions and collect additional information. We introduce a novel optimization and learning algorithm to address this problem. This algorithm effectively combines the strengths of two algorithmic approaches in a seamless manner, including 1) an online primal-dual algorithm for deciding the optimal timing to reach out to patients, and 2) a contextual bandit learning algorithm to deliver personalized treatment to the patient. We prove that this algorithm admits a sub-linear regret bound. We illustrate the usefulness of this algorithm on both synthetic and real-world data.

Multiplayer bandits have recently garnered significant attention due to their relevance in cognitive radio networks. While the existing body of literature predominantly focuses on synchronous players, real-world radio networks, such as those in IoT applications, often feature asynchronous (i.e., randomly activated) devices. This highlights the need for addressing the more challenging asynchronous multiplayer bandits problem. Our first result shows that a natural extension of UCB achieves a minimax regret of $\mathcal{O}(\sqrt{T\log(T)})$ in the centralized setting. More significantly, we introduce Cautious Greedy, which uses $\mathcal{O}(\log(T))$ communications and whose instance-dependent regret is constant if the optimal policy assigns at least one player to each arm (a situation proven to occur when arm means are sufficiently close). Otherwise, the regret is, as usual, $\log(T)$ times the sum of some inverse sub-optimality gaps. We substantiate the optimality of Cautious Greedy through lower-bound analysis based on data-dependent terms. Therefore, we establish a strong baseline for asynchronous multiplayer bandits, at least with $\mathcal{O}(\log(T))$ communications.

Proper scoring rules evaluate the quality of probabilistic predictions, playing an essential role in the pursuit of accurate and well-calibrated models. Every proper score decomposes into two fundamental components – proper calibration error and refinement – utilizing a Bregman divergence. While uncertainty calibration has gained significant attention, current literature lacks a general estimator for these quantities with known statistical properties. To address this gap, we propose a method that allows consistent, and asymptotically unbiased estimation of all proper calibration errors and refinement terms. In particular, we introduce Kullback-Leibler calibration error, induced by the commonly used cross-entropy loss. As part of our results, we prove the relation between refinement and f-divergences, which implies information monotonicity in neural networks, regardless of which proper scoring rule is optimized. Our experiments validate empirically the claimed properties of the proposed estimator and suggest that the selection of a post-hoc calibration method should be determined by the particular calibration error of interest.

Given samples from two joint distributions, we consider the problem of Optimal Transportation (OT) between them when conditioned on a common variable. We focus on the general setting where the conditioned variable may be continuous, and the marginals of this variable in the two joint distributions may not be the same. In such settings, standard OT variants cannot be employed, and novel estimation techniques are necessary. Since the main challenge is that the conditional distributions are not explicitly available, the key idea in our OT formulation is to employ kernelized-least-squares terms computed over the joint samples, which implicitly match the transport plan’s marginals with the empirical conditionals. Under mild conditions, we prove that our estimated transport plans, as a function of the conditioned variable, are asymptotically optimal. For finite samples, we show that the deviation in terms of our regularized objective is bounded by $O(m^{-1/4})$, where $m$ is the number of samples. We also discuss how the conditional transport plan could be modelled using explicit probabilistic models as well as using implicit generative ones. We empirically verify the consistency of our estimator on synthetic datasets, where the optimal plan is analytically known. When employed in applications like prompt learning for few-shot classification and conditional-generation in the context of predicting cell responses to treatment, our methodology improves upon state-of-the-art methods.

In multi-class hierarchical classification, a natural evaluation metric is the tree distance loss that takes the value of two labels’ distance on the pre-defined tree hierarchy. This metric is motivated by that its Bayes optimal solution is the deepest label on the tree whose induced superclass (subtree rooted at it) includes the true label with probability at least $\frac{1}{2}$. However, it can hardly handle the risk sensitivity of different tasks since its accuracy requirement for induced superclasses is fixed at $\frac{1}{2}$. In this paper, we first introduce a new evaluation metric that generalizes the tree distance loss, whose solution’s accuracy constraint $\frac{1+c}{2}$ can be controlled by a penalty value $c$ tailored for different tasks: a higher c indicates the emphasis on prediction’s accuracy and a lower one indicates that on specificity. Then, we propose a novel class of consistent surrogate losses based on an intuitive presentation of our generalized metric and its regret, which can be compatible with various binary losses. Finally, we theoretically derive the regret transfer bounds for our proposed surrogates and empirically validate their usefulness on benchmark datasets.

We analyze a paradigm for interpretable Manifold Learning for scientific data analysis, whereby one parametrizes a manifold with d smooth functions from a scientist-provided dictionary of meaningful, domain-related functions. When such a parametrization exists, we provide an algorithm for finding it based on sparse regression in the manifold tangent bundle, bypassing more standard, agnostic manifold learning algorithms. We prove conditions for the existence of such parameterizations in function space and the first end to end recovery results from finite samples. The method is demonstrated on both synthetic problems and with data from a real scientific domain.

The Random Forests classifier, a widely utilized off-the-shelf classification tool, assumes training and test samples come from the same distribution as other standard classifiers. However, in safety-critical scenarios like medical diagnosis and network attack detection, discrepancies between the training and test sets, including the potential presence of novel outlier samples not appearing during training, can pose significant challenges. To address this problem, we introduce the Conformalized Semi-Supervised Random Forest (CSForest), which couples the conformalization technique Jackknife+aB with semi-supervised tree ensembles to construct a set-valued prediction $C(x)$. Instead of optimizing over the training distribution, CSForest employs unlabeled test samples to enhance accuracy and flag unseen outliers by generating an empty set. Theoretically, we establish CSForest to cover true labels for previously observed inlier classes under arbitrarily label-shift in the test data. We compare CSForest with state-of-the-art methods using synthetic examples and various real-world datasets, under different types of distribution changes in the test domain. Our results highlight CSForest’s effective prediction of inliers and its ability to detect outlier samples unique to the test data. In addition, CSForest shows persistently good performance as the sizes of the training and test sets vary. Codes of CSForest are available at https://github.com/yujinhan98/CSForest.

Uncertainty estimation is critical in high-stakes machine learning applications. One effective way to estimate uncertainty is conformal prediction, which can provide predictive inference with statistical coverage guarantees. We present a new conformal regression method, Spline Prediction Intervals via Conformal Estimation (SPICE), that estimates the conditional density using neural- network-parameterized splines. We prove universal approximation and optimality results for SPICE, which are empirically reflected by our experiments. SPICE is compatible with two different efficient-to- compute conformal scores, one designed for size-efficient marginal coverage (SPICE-ND) and the other for size-efficient conditional coverage (SPICE-HPD). Results on benchmark datasets demonstrate SPICE-ND models achieve the smallest average prediction set sizes, including average size reductions of nearly 50% for some datasets compared to the next best baseline. SPICE-HPD models achieve the best conditional coverage compared to baselines. The SPICE implementation is made available.

Data-driven approaches to predict-then-optimize decision-making problems seek to mitigate the risk of uncertainty region misspecification in safety-critical settings. Current approaches, however, suffer from considering overly conservative uncertainty regions, often resulting in suboptimal decision-making. To this end, we propose Conformal-Predict-Then-Optimize (CPO), a framework for leveraging highly informative, nonconvex conformal prediction regions over high-dimensional spaces based on conditional generative models, which have the desired distribution-free coverage guarantees. Despite guaranteeing robustness, such black-box optimization procedures alone inspire little confidence owing to the lack of explanation of why a particular decision was found to be optimal. We, therefore, augment CPO to additionally provide semantically meaningful visual summaries of the uncertainty regions to give qualitative intuition for the optimal decision. We highlight the CPO framework by demonstrating results on a suite of simulation-based inference benchmark tasks and a vehicle routing task based on probabilistic weather prediction.

Machine learning models are widely applied in various fields. Stakeholders often use post-hoc feature importance methods to better understand the input features’ contribution to the models’ predictions. The interpretation of the importance values provided by these methods is frequently based on the relative order of the features (their ranking) rather than the importance values themselves. Since the order may be unstable, we present a framework for quantifying the uncertainty in global importance values. We propose a novel method for the post-hoc interpretation of feature importance values that is based on the framework and pairwise comparisons of the feature importance values. This method produces simultaneous confidence intervals for the features’ ranks, which include the “true” (infinite sample) ranks with high probability, and enables the selection of the set of top-k important features.

Learning from Preferential Feedback (LfPF) plays an essential role in training Large Language Models, as well as certain types of interactive learning agents. However, a substantial gap exists between the theory and application of LfPF algorithms. Current results guaranteeing the existence of optimal policies in LfPF problems assume that both the preferences and transition dynamics are determined by a Markov Decision Process. We introduce the Direct Preference Process, a new framework for analyzing LfPF problems in partially-observable, non-Markovian environments. Within this framework, we establish conditions that guarantee the existence of optimal policies by considering the ordinal structure of the preferences. We show that a decision-making problem can have optimal policies – that are characterized by recursive optimality equations – even when no reward function can express the learning goal. These findings underline the need to explore preference-based learning strategies which do not assume that preferences are generated by reward.

We consider the problem of identifying a conditional causal effect through covariate adjustment. We focus on the setting where the causal graph is known up to one of two types of graphs: a maximally oriented partially directed acyclic graph (MPDAG) or a partial ancestral graph (PAG). Both MPDAGs and PAGs represent equivalence classes of possible underlying causal models. After defining adjustment sets in this setting, we provide a necessary and sufficient graphical criterion – the conditional adjustment criterion – for finding these sets under conditioning on variables unaffected by treatment. We further provide explicit sets from the graph that satisfy the conditional adjustment criterion, and therefore, can be used as adjustment sets for conditional causal effect identification.

Metrics such as the outbreak size in an epidemic process on a network are fundamental quantities used in public health analyses. The datasets used in such models used in practice, e.g., the contact network and disease states, are sensitive in many settings. We study the complexity of computing epidemic outbreak size within a given time horizon, under edge differential privacy. These quantities have high sensitivity, and we show that giving algorithms with good utility guarantees is impossible for general graphs. To address these hardness results, we consider a smaller class of graphs with similar properties as social networks (called expander graphs) and give a polynomial-time algorithm with strong utility guarantees. Our results are the first to give any non-trivial guarantees for differentially private infection size estimation.

Compression schemes have been extensively used in Federated Learning (FL) to reduce the communication cost of distributed learning. While most approaches rely on a bounded variance assumption of the noise produced by the compressor, this paper investigates the use of compression and aggregation schemes that produce a specific error distribution, e.g., Gaussian or Laplace, on the aggregated data. We present and analyze different aggregation schemes based on layered quantizers achieving exact error distribution. We provide different methods to leverage the proposed compression schemes to obtain compression-for-free in differential privacy applications. Our general compression methods can recover and improve standard FL schemes with Gaussian perturbations such as Langevin dynamics and randomized smoothing.

We analyze the sample complexity of single-loop quadratic penalty and augmented Lagrangian algorithms for solving nonconvex optimization problems with functional equality constraints. We consider three cases, in all of which the objective is stochastic, that is, an expectation over an unknown distribution that is accessed by sampling. The nature of the equality constraints differs among the three cases: deterministic and linear in the first case, deterministic and nonlinear in the second case, and stochastic and nonlinear in the third case. Variance reduction techniques are used to improve the complexity. To find a point that satisfies $\varepsilon$-approximate first-order conditions, we require $\widetilde{O}(\varepsilon^{-3})$ complexity in the first case, $\widetilde{O}(\varepsilon^{-4})$ in the second case, and $\widetilde{O}(\varepsilon^{-5})$ in the third case. For the first and third cases, they are the first algorithms of “single loop” type that also use $O(1)$ samples at each iteration and still achieve the best-known complexity guarantees.

We derive generic information-theoretic and PAC-Bayesian generalization bounds involving an arbitrary convex comparator function, which measures the discrepancy between the training loss and the population loss. The bounds hold under the assumption that the cumulant-generating function (CGF) of the comparator is upper-bounded by the corresponding CGF within a family of bounding distributions. We show that the tightest possible bound is obtained with the comparator being the convex conjugate of the CGF of the bounding distribution, also known as the Cramér function. This conclusion applies more broadly to generalization bounds with a similar structure. This confirms the near-optimality of known bounds for bounded and sub-Gaussian losses and leads to novel bounds under other bounding distributions.

Federated Learning (FL) enables large-scale distributed training of machine learning models, while still allowing individual nodes to maintain data locally. However, executing FL at scale comes with inherent practical challenges: 1) heterogeneity of the local node data distributions, 2) heterogeneity of node computational speeds (asynchrony), but also 3) constraints in the amount of communication between the clients and the server. In this work, we present the first variant of the classic federated averaging (FedAvg) algorithm which, at the same time, supports data heterogeneity, partial client asynchrony, and communication compression. Our algorithm comes with a novel, rigorous analysis showing that, in spite of these system relaxations, it can provide similar convergence to FedAvg in interesting parameter regimes. Experimental results in the rigorous LEAF benchmark on setups of up to $300$ nodes show that our algorithm ensures fast convergence for standard federated tasks, improving upon prior quantized and asynchronous approaches.