AISTATS 2023 论文列表

We propose a meta-learning method to improve the anomaly detection performance on unseen target tasks that have only unlabeled data. Existing meta-learning methods for anomaly detection have shown remarkable performance but require labeled data in target tasks. Although they can treat unlabeled data as normal assuming anomalies in the unlabeled data are negligible, this assumption is often violated in practice. As a result, the methods have low performance. Our method meta-learns with related tasks that have labeled and unlabeled data such that the expected test anomaly detection performance is directly improved when the anomaly detector is adapted to given unlabeled data. Our method is based on autoencoders (AEs), which are widely used neural network-based anomaly detectors. We model anomalous attributes for each unlabeled instance in the reconstruction loss of the AE, which are used to prevent the anomalies from being reconstructed; they can remove the effect of the anomalies. We formulate adaptation to the unlabeled data as a learning problem of the last layer of the AE and the anomalous attributes. This formulation enables the optimum solution to be obtained with a closed-form alternate update formula, which is preferable to efficiently maximize the expected test anomaly detection performance. The effectiveness of our method is experimentally shown with four real-world datasets.

Bayesian model comparison (BMC) offers a principled probabilistic approach to study and rank competing models. In standard BMC, we construct a discrete probability distribution over the set of possible models, conditional on the observed data of interest. These posterior model probabilities (PMPs) are measures of uncertainty, but—when derived from a finite number of observations—are also uncertain themselves. In this paper, we conceptualize distinct levels of uncertainty which arise in BMC. We explore a fully probabilistic framework for quantifying meta-uncertainty, resulting in an applied method to enhance any BMC workflow. Drawing on both Bayesian and frequentist techniques, we represent the uncertainty over the uncertain PMPs via meta-models which combine simulated and observed data into a predictive distribution for PMPs on new data. We demonstrate the utility of the proposed method in the context of conjugate Bayesian regression, likelihood-based inference with Markov chain Monte Carlo, and simulation-based inference with neural networks.

Model Agnostic Meta-Learning (MAML) is widely used to find a good initialization for a family of tasks. Despite its success, a critical challenge in MAML is to calculate the gradient w.r.t. the initialization of a long training trajectory for the sampled tasks, because the computation graph can rapidly explode and the computational cost is very expensive. To address this problem, we propose Adjoint MAML (A-MAML). We view gradient descent in the inner optimization as the evolution of an Ordinary Differential Equation (ODE). To efficiently compute the gradient of the validation loss w.r.t. the initialization, we use the adjoint method to construct a companion, backward ODE. To obtain the gradient w.r.t. the initialization, we only need to run the standard ODE solver twice — one is forward in time that evolves a long trajectory of gradient flow for the sampled task; the other is backward and solves the adjoint ODE. We need not create or expand any intermediate computational graphs, adopt aggressive approximations, or impose proximal regularizers in the training loss. Our approach is cheap, accurate, and adaptable to different trajectory lengths. We demonstrate the advantage of our approach in both synthetic and real-world meta-learning tasks. The code is available at https://github.com/shib0li/Adjoint-MAML.

Data is the foundation of most science. Unfortunately, sharing data can be obstructed by the risk of violating data privacy, impeding research in fields like healthcare. Synthetic data is a potential solution. It aims to generate data that has the same distribution as the original data, but that does not disclose information about individuals. Membership Inference Attacks (MIAs) are a common privacy attack, in which the attacker attempts to determine whether a particular real sample was used for training of the model. Previous works that propose MIAs against generative models either display low performance—giving the false impression that data is highly private—or need to assume access to internal generative model parameters—a relatively low-risk scenario, as the data publisher often only releases synthetic data, not the model. In this work we argue for a realistic MIA setting that assumes the attacker has some knowledge of the underlying data distribution. We propose DOMIAS, a density-based MIA model that aims to infer membership by targeting local overfitting of the generative model. Experimentally we show that DOMIAS is significantly more successful at MIA than previous work, especially at attacking uncommon samples. The latter is disconcerting since these samples may correspond to underrepresented groups. We also demonstrate how DOMIAS’ MIA performance score provides an interpretable metric for privacy, giving data publishers a new tool for achieving the desired privacy-utility trade-off in their synthetic data.

In mediated uncoupled learning (MU-learning), the goal is to predict an output variable $Y$ given an input variable $X$ as in ordinary supervised learning while the training dataset has no joint samples of $(X, Y)$ but only independent samples of $(X, U)$ and $(U, Y)$ each observed with a mediating variable $U$. The existing MU-learning methods can only handle the squared loss, which prohibited the use of other popular loss functions such as the cross-entropy loss. We propose a general MU-learning framework that allows for the problems with Bregman divergences, which cover a wide range of loss functions useful for various types of tasks, in a unified manner. This loss family has maximal generality among those whose minimizers characterize the conditional expectation. We prove that the proposed objective function is a tighter approximation to the oracle loss that one would minimize if ordinary supervised samples of $(X, Y)$ were available. We also propose an estimator of an interval containing the expected test loss of predictions of a trained model only using $(X, U)$- and $(U, Y)$-data. We provide a theoretical analysis on the excess risk for the proposed method and confirm its practical usefulness with regression experiments with synthetic data and low-quality image classification experiments with benchmark datasets.

We study the fair regression problem under the notion of Mean Parity (MP) fairness, which requires the conditional mean of the learned function output to be constant with respect to the sensitive attributes. We address this problem by leveraging reproducing kernel Hilbert space (RKHS) to construct the functional space whose members are guaranteed to satisfy the fairness constraints. The proposed functional space suggests a closed-form solution for the fair regression problem that is naturally compatible with multiple sensitive attributes. Furthermore, by formulating the fairness-accuracy tradeoff as a relaxed fair regression problem, we derive a corresponding regression function that can be implemented efficiently and provides interpretable tradeoffs. More importantly, under some mild assumptions, the proposed method can be applied to regression problems with a covariance-based notion of fairness. Experimental results on benchmark datasets show the proposed methods achieve competitive and even superior performance compared with several state-of-the-art methods.

We consider the problem of finding the matching map between two sets of $d$-dimensional noisy feature-vectors. The distinctive feature of our setting is that we do not assume that all the vectors of the first set have their corresponding vector in the second set. If $n$ and $m$ are the sizes of these two sets, we assume that the matching map that should be recovered is defined on a subset of unknown cardinality $k^*\le \min(n,m)$. We show that, in the high-dimensional setting, if the signal-to-noise ratio is larger than $5(d\log(4nm/\alpha))^{1/4}$, then the true matching map can be recovered with probability $1-\alpha$. Interestingly, this threshold does not depend on $k^*$ and is the same as the one obtained in prior work in the case of $k = \min(n,m)$. The procedure for which the aforementioned property is proved is obtained by a data-driven selection among candidate mappings $\{\hat\pi_k:k\in[\min(n,m)]\}$. Each $\hat\pi_k$ minimizes the sum of squares of distances between two sets of size $k$. The resulting optimization problem can be formulated as a minimum-cost flow problem, and thus solved efficiently. Finally, we report the results of numerical experiments on both synthetic and real-world data that illustrate our theoretical results and provide further insight into the properties of the algorithms studied in this work.

Shapley values are model-agnostic methods for explaining model predictions. Many commonly used methods of computing Shapley values, known as off-manifold methods, rely on model evaluations on out-of-distribution input samples. Consequently, explanations obtained are sensitive to model behaviour outside the data distribution, which may be irrelevant for all practical purposes. While on-manifold methods have been proposed which do not suffer from this problem, we show that such methods are overly dependent on the input data distribution, and therefore result in unintuitive and misleading explanations. To circumvent these problems, we propose ManifoldShap, which respects the model’s domain of validity by restricting model evaluations to the data manifold. We show, theoretically and empirically, that ManifoldShap is robust to off-manifold perturbations of the model and leads to more accurate and intuitive explanations than existing state-of-the-art Shapley methods.

We introduce a method, MMD-B-Fair, to learn fair representations of data via kernel two-sample testing. We find neural features of our data where a maximum mean discrepancy (MMD) test cannot distinguish between different values of sensitive attributes, while preserving information about the target. Minimizing the power of an MMD test is more difficult than maximizing it (as done in previous work), because the test threshold’s complex behavior cannot be simply ignored. Our method exploits the simple asymptotics of block testing schemes to efficiently find fair representations without requiring the complex adversarial optimization or generative modelling schemes widely used by existing work on fair representation learning. We evaluate our approach on various datasets, showing its ability to hide information about sensitive attributes, and its effectiveness in downstream transfer tasks.

Tensor decomposition methods have proven effective in various applications, including compression and acceleration of neural networks. At the same time, the problem of determining optimal decomposition ranks, which present the crucial parameter controlling the compressionaccuracy trade-off, is still acute. In this paper, we introduce MARS - a new efficient method for the automatic selection of ranks in general tensor decompositions. During training, the procedure learns binary masks over decomposition cores that “select” the optimal tensor structure. The learning is performed via relaxed maximum a posteriori (MAP) estimation in a specific Bayesian model and can be naturally embedded into the standard neural network training routine. Diverse experiments demonstrate that MARS achieves better results compared to previous works in various tasks.

Training neural networks on a large dataset requires substantial computational costs. Dataset reduction selects or synthesizes data instances based on the large dataset, while minimizing the degradation in generalization performance from the full dataset. Existing methods utilize the neural network during the dataset reduction procedure, so the model parameter becomes important factor in preserving the performance after reduction. By depending upon the importance of parameters, this paper introduces a new reduction objective, coined LCMat, which Matches the Loss Curvatures of the original dataset and reduced dataset over the model parameter space, more than the parameter point. This new objective induces a better adaptation of the reduced dataset on the perturbed parameter region than the exact point matching. Particularly, we identify the worst case of the loss curvature gap from the local parameter region, and we derive the implementable upper bound of such worst-case with theoretical analyses. Our experiments on both coreset selection and condensation benchmarks illustrate that LCMat shows better generalization performances than existing baselines.

Recent theoretical analyses of the convergence of gradient descent (GD) to a global minimum for over-parametrized neural networks make strong assumptions on the step size (infinitesimal), the hidden-layer width (infinite), or the initialization (spectral, balanced). In this work, we relax these assumptions and derive a linear convergence rate for two-layer linear networks trained using GD on the squared loss in the case of finite step size, finite width and general initialization. Despite the generality of our analysis, our rate estimates are significantly tighter than those of prior work. Moreover, we provide a time-varying step size rule that monotonically improves the convergence rate as the loss function decreases to zero. Numerical experiments validate our findings.

We propose the NeRF-LEBM, a likelihoodbased top-down 3D-aware 2D image generative model that incorporates 3D representation via Neural Radiance Fields (NeRF) and 2D imaging process via differentiable volume rendering. The model represents an image as a rendering process from 3D object to 2D image and is conditioned on some latent variables that account for object characteristics and are assumed to follow informative trainable energy-based prior models. We propose two likelihood-based learning frameworks to train the NeRF-LEBM: (i) maximum likelihood estimation with Markov chain Monte Carlo-based inference and (ii) variational inference with the reparameterization trick. We study our models in the scenarios with both known and unknown camera poses. Experiments on several benchmark datasets demonstrate that the NeRF-LEBM can infer 3D object structures from 2D images, generate 2D images with novel views and objects, learn from incomplete 2D images, and learn from 2D images with known or unknown camera poses.

Programmatic Weak Supervision (PWS) has emerged as a widespread paradigm to synthesize training labels efficiently. The core component of PWS is the label model, which infers true labels by aggregating the outputs of multiple noisy supervision sources abstracted as labeling functions (LFs). Existing statistical label models typically rely only on the outputs of LF, ignoring the instance features when modeling the underlying generative process. In this paper, we attempt to incorporate the instance features into a statistical label model via the proposed FABLE. In particular, it is built on a mixture of Bayesian label models, each corresponding to a global pattern of correlation, and the coefficients of the mixture components are predicted by a Gaussian Process classifier based on instance features. We adopt an auxiliary variable-based variational inference algorithm to tackle the non-conjugate issue between the Gaussian Process and Bayesian label models. Extensive empirical comparison on eleven benchmark datasets sees FABLE achieving the highest averaged performance across nine baselines. Our implementation of FABLE can be found in https://github.com/JieyuZ2/wrench/blob/main/wrench/labelmodel/fable.py.

It has been known by neuroscience studies that partial and transient forgetting of memory often plays an important role in the brain to improve performance for certain intellectual activities. In machine learning, associative memory models such as classical and modern Hopfield networks have been proposed to express memories as attractors in the feature space of a closed recurrent network. In this work, we propose learning with partial forgetting (LwPF), where a partial forgetting functionality is designed by element-wise non-bijective projections, for memory neurons in modern Hopfield networks to improve model performance. We incorporate LwPF into the attention mechanism also, whose process has been shown to be identical to the update rule of a certain modern Hopfield network, by modifying the corresponding Lagrangian. We evaluated the effectiveness of LwPF on three diverse tasks such as bit-pattern classification, immune repertoire classification for computational biology, and image classification for computer vision, and confirmed that LwPF consistently improves the performance of existing neural networks including DeepRC and vision transformers.

In real-world decision-making, uncertainty is important yet difficult to handle. Stochastic dominance provides a theoretically sound approach to comparing uncertain quantities, but optimization with stochastic dominance constraints is often computationally expensive, which limits practical applicability. In this paper, we develop a simple yet efficient approach for the problem, Light Stochastic Dominance Solver (light-SD), by leveraging properties of the Lagrangian. We recast the inner optimization in the Lagrangian as a learning problem for surrogate approximation, which bypasses the intractability and leads to tractable updates or even closed-form solutions for gradient calculations. We prove convergence of the algorithm and test it empirically. The proposed light-SD demonstrates superior performance on several representative problems ranging from finance to supply chain management.

Learning to optimize (L2O) has gained increasing popularity, which automates the design of optimizers by data-driven approaches. However, current L2O methods often suffer from poor generalization performance in at least two folds: (i) applying the L2O-learned optimizer to unseen optimizees, in terms of lowering their loss function values (optimizer generalization, or “generalizable learning of optimizers”); and (ii) the test performance of an optimizee (itself as a machine learning model), trained by the optimizer, in terms of the accuracy over unseen data (optimizee generalization, or “learning to generalize”). While the optimizer generalization has been recently studied, the optimizee generalization (or learning to generalize) has not been rigorously studied in the L2O context, which is the aim of this paper. We first theoretically establish an implicit connection between the local entropy and the Hessian, and hence unify their roles in the handcrafted design of generalizable optimizers as equivalent metrics of the landscape flatness of loss functions. We then propose to incorporate these two metrics as flatness-aware regularizers into the L2O framework in order to meta-train optimizers to learn to generalize, and theoretically show that such generalization ability can be learned during the L2O meta-training process and then transformed to the optimizee loss function. Extensive experiments consistently validate the effectiveness of our proposals with substantially improved generalization on multiple sophisticated L2O models and diverse optimizees.

We study the statistical properties of learning to defer (L2D) to multiple experts. In particular, we address the open problems of deriving a consistent surrogate loss, confidence calibration, and principled ensembling of experts. Firstly, we derive two consistent surrogates—one based on a softmax parameterization, the other on a one-vs-all (OvA) parameterization—that are analogous to the single expert losses proposed by Mozannar and Sontag (2020) and Verma and Nalisnick (2022), respectively. We then study the frameworks’ ability to estimate $P( m_j = y | x )$, the probability that the $j$th expert will correctly predict the label for $x$. Theory shows the softmax-based loss causes mis-calibration to propagate between the estimates while the OvA-based loss does not (though in practice, we find there are trade offs). Lastly, we propose a conformal inference technique that chooses a subset of experts to query when the system defers. We perform empirical validation on tasks for galaxy, skin lesion, and hate speech classification.

Motivated by the limited qubit capacity of current quantum systems, we study the quantum sample complexity of k-qubit quantum operators, i.e., operations applicable on only k out of d qubits. The problem is studied according to the quantum probably approximately correct (QPAC) model abiding by quantum mechanical laws such as no-cloning, state collapse, and measurement incompatibility. With the delicacy of quantum samples and the richness of quantum operations, one expects a significantly larger quantum sample complexity. This paper proves the contrary. We show that the quantum sample complexity of k-qubit quantum operations is comparable to the classical sample complexity of their counterparts (juntas), at least when $\frac{k}{d}\ll 1$. This is surprising, especially since sample duplication is prohibited, and measurement incompatibility would lead to an exponentially larger sample complexity with standard methods. Our approach is based on the Pauli decomposition of quantum operators and a technique called Quantum Shadow Sampling (QSS) to reduce the sample complexity exponentially. The results are proved by developing (i) a connection between the learning loss and the Pauli decomposition; (ii) a scalable QSS circuit for estimating the Pauli coefficients; and (iii) a quantum algorithm for learning $k$-qubit operators with sample complexity $O(\frac{k4^k}{\epsilon^2}\log d)$.