UAI 2023 论文列表

For the stable optimization of deep neural networks, regularization methods such as dropout and batch normalization have been used in various tasks. Nevertheless, the correct position to apply dropout has rarely been discussed, and different positions have been employed depending on the practitioners. In this study, we investigate the correct position to apply dropout. We demonstrate that for a residual network with batch normalization, applying dropout at certain positions increases the performance, whereas applying dropout at other positions decreases the performance. Based on theoretical analysis, we provide the following guideline for the correct position to apply dropout: apply one dropout after the last batch normalization but before the last weight layer in the residual branch. We provide detailed theoretical explanations to support this claim and demonstrate them through module tests. In addition, we investigate the correct position of dropout in the head that produces the final prediction. Although the current consensus is to apply dropout after global average pooling, we prove that applying dropout before global average pooling leads to a more stable output. The proposed guidelines are validated through experiments using different datasets and models.

Visual object tracking has seen significant progress in recent years. However, the vast majority of this work focuses on tracking objects within the image plane of a single camera and ignores the uncertainty associated with predicted object locations. In this work, we focus on the geospatial object tracking problem using data from a distributed camera network. The goal is to predict an object’s track in geospatial coordinates along with uncertainty over the object’s location while respecting communication constraints that prohibit centralizing raw image data. We present a novel single-object geospatial tracking data set that includes high-accuracy ground truth object locations and video data from a network of four cameras. We present a modeling framework for addressing this task including a novel backbone model and explore how uncertainty calibration and fine-tuning through a differentiable tracker affect performance.

Linear bandit algorithms have been extensively studied and have shown successful in sequential decision tasks despite their simplicity. Many algorithms however work under the assumption that the reward is the sum of linear function of observed contexts and a sub-Gaussian error. In practical applications, errors can be heavy-tailed, especially in financial data. In such reward environments, algorithms designed for sub-Gaussian error may underexplore, resulting in suboptimal regret. In this paper, we relax the reward assumption and propose a novel linear bandit algorithm which works well under heavy-tailed errors as well. The proposed algorithm utilizes Huber regression. When contexts are stochastic with positive definite covariance matrix and the $(1+\delta)$-th moment of the error is bounded by a constant, we show that the high-probability upper bound of the regret is $O(\sqrt{d}T^{\frac{1}{1+\delta}}(\log dT)^{\frac{\delta}{1+\delta}})$, where $d$ is the dimension of context variables, $T$ is the time horizon, and $\delta\in (0,1]$. This bound improves on the state-of-the-art regret bound of the Median of Means and Truncation algorithm by a factor of $\sqrt{\log T}$ and $\sqrt{d}$ for the case where the time horizon $T$ is unknown. We also remark that when $\delta=1$, the order is the same as the regret bound of linear bandit algorithms designed for sub-Gaussian errors. We support our theoretical findings with synthetic experiments.

We study the problem of conservative off-policy evaluation (COPE) where given an offline dataset of environment interactions, collected by other agents, we seek to obtain a (tight) lower bound on a policy’s performance. This is crucial when deciding whether a given policy satisfies certain minimal performance/safety criteria before it can be deployed in the real world. To this end, we introduce HAMBO, which builds on an uncertainty-aware learned model of the transition dynamics. To form a conservative estimate of the policy’s performance, HAMBO hallucinates worst-case trajectories that the policy may take, within the margin of the models’ epistemic confidence regions. We prove that the resulting COPE estimates are valid lower bounds, and, under regularity conditions, show their convergence to the true expected return. Finally, we discuss scalable variants of our approach based on Bayesian Neural Networks and empirically demonstrate that they yield reliable and tight lower bounds in various continuous control environments.

Combining Gaussian processes with the expressive power of deep neural networks is commonly done nowadays through deep kernel learning (DKL). Unfortunately, due to the kernel optimization process, this often results in losing their Bayesian benefits. In this study, we present a novel approach for learning deep kernels by utilizing infinite-width neural networks. We propose to use the Neural Network Gaussian Process (NNGP) model as a guide to the DKL model in the optimization process. Our approach harnesses the reliable uncertainty estimation of the NNGPs to adapt the DKL target confidence when it encounters novel data points. As a result, we get the best of both worlds, we leverage the Bayesian behavior of the NNGP, namely its robustness to overfitting, and accurate uncertainty estimation, while maintaining the generalization abilities, scalability, and flexibility of deep kernels. Empirically, we show on multiple benchmark datasets of varying sizes and dimensionality, that our method is robust to overfitting, has good predictive performance, and provides reliable uncertainty estimations.

We consider the unlabeled linear regression reading as $\mathbf{Y} = \mathbf{\Pi}^{*}\mathbf{X}\mathbf{B}^* + \mathbf{W}$, where $\mathbf{\Pi}^{*}, \mathbf{B}^*$ and $\mathbf{W}$ represents missing (or incomplete) correspondence information, signals, and additive noise, respectively. Our goal is to perform data alignment between $\mathbf{Y}$ and $\mathbf{X}$, or equivalently, reconstruct the correspondence information encoded by $\mathbf{\Pi}^*$. Based on whether signal $\mathbf{B}^*$ is given a prior, we separately propose two greedy-selection-based estimators, which both reach the mini-max optimality. Compared with previous works, our work $(i)$ supports partial recovery of the correspondence information; and $(ii)$ applies to a general matrix family rather than the permutation matrices, to put more specifically, selection matrices, where multiple rows of $\mathbf{X}$ can correspond to the same row in $\mathbf{Y}$. Moreover, numerical experiments are provided to corroborate our claims.

Graph classification aims to categorise graphs based on their structure and node attributes. In this work, we propose to tackle this task using tools from graph signal processing by deriving spectral features, which we then use to design two variants of Gaussian process models for graph classification. The first variant uses spectral features based on the distribution of energy of a node feature signal over the spectrum of the graph. We show that even such a simple approach, having no learned parameters, can yield competitive performance compared to strong neural network and graph kernel baselines. A second, more sophisticated variant is designed to capture multi-scale and localised patterns in the graph by learning spectral graph wavelet filters, obtaining improved performance on synthetic and real-world data sets. Finally, we show that both models produce well calibrated uncertainty estimates, enabling reliable decision making based on the model predictions.

Self-supervised learning (SSL) for graphs is an essential problem since graph data are ubiquitous and labeling can be costly. We argue that existing SSL approaches for graphs have two limitations. First, they rely on corruption techniques such as node attribute perturbation and edge dropping to generate graph views for contrastive learning. These unnatural corruption techniques require extensive tuning efforts and provide marginal improvements. Second, the current approaches require the computation of multiple graph views, which is memory and computationally inefficient. These shortcomings of graph SSL call for a corruption-free single-view learning approach, but the strawman approach of using neighboring nodes as positive examples suffers two problems: it ignores the strength of connections between nodes implied by the graph structure on a macro level, and cannot deal with the high noise in real-world graphs. We propose Proximity Divergence Minimization (PDM), a corruption-free single-view graph SSL approach that overcomes these problems by leveraging node proximity to measure connection strength and denoise the graph structure. Through extensive experiments, we show that PDM achieves up to 4.55% absolute improvement in ROC-AUC on graph SSL tasks over state-of-the-art approaches while being more memory efficient. Moreover, PDM even outperforms supervised training on node classification tasks of ogbn-proteins dataset. Our code is publicly available.

Data for pretraining machine learning models often consists of collections of heterogeneous datasets. Although training on their union is reasonable in agnostic settings, it might be suboptimal when the target domain —where the model will ultimately be used— is known in advance. In that case, one would ideally pretrain only on the dataset(s) most similar to the target one. Instead of limiting this choice to those datasets already present in the pretraining collection, here we explore extending this search to all datasets that can be synthesized as ‘combinations’ of them. We define such combinations as multi-dataset interpolations, formalized through the notion of generalized geodesics from optimal transport (OT) theory. We compute these geodesics using a recent notion of distance between labeled datasets, and derive alternative interpolation schemes based on it: using either barycentric projections or optimal transport maps, the latter computed using recent neural OT methods. These methods are scalable, efficient, and —notably— can be used to interpolate even between datasets with distinct and unrelated label sets. Through various experiments in transfer learning in computer vision, we demonstrate this is a promising new approach for targeted on-demand dataset synthesis.

Not being able to understand and predict the behavior of deep learning systems makes it hard to decide what architecture and algorithm to use for a given problem. In science and engineering, modeling is a methodology used to understand complex systems whose internal processes are opaque. Modeling replaces a complex system with a simpler, more interpretable surrogate. Drawing inspiration from this, we construct a class of surrogate models for neural networks using Gaussian processes. Rather than deriving kernels for infinite neural networks, we learn kernels empirically from the naturalistic behavior of finite neural networks. We demonstrate our approach captures existing phenomena related to the spectral bias of neural networks, and then show that our surrogate models can be used to solve practical problems such as identifying which points most influence the behavior of specific neural networks and predicting which architectures and algorithms will generalize well for specific datasets.

We propose functional causal Bayesian optimization (fCBO), a method for finding interventions that optimize a target variable in a known causal graph. fCBO extends the CBO family of methods to enable functional interventions, which set a variable to be a deterministic function of other variables in the graph. fCBO models the unknown objectives with Gaussian processes whose inputs are defined in a reproducing kernel Hilbert space, thus allowing to compute distances among vector-valued functions. In turn, this enables to sequentially select functions to explore by maximizing an expected improvement acquisition functional while keeping the typical computational tractability of standard BO settings. We introduce graphical criteria that establish when considering functional interventions allows attaining better target effects, and conditions under which selected interventions are also optimal for conditional target effects. We demonstrate the benefits of the method in a synthetic and in a real-world causal graph.

We study the problem of best-arm identification (BAI) in the fixed-budget setting with heterogeneous reward variances. We propose two variance-adaptive BAI algorithms for this setting: SHVar for known reward variances and SHAdaVar for unknown reward variances. Our algorithms rely on non-uniform budget allocations among the arms where the arms with higher reward variances are pulled more often than those with lower variances. The main algorithmic novelty is in the design of SHAdaVar, which allocates budget greedily based on overestimating the unknown reward variances. We bound probabilities of misidentifying the best arms in both SHVar and SHAdaVar. Our analyses rely on novel lower bounds on the number of pulls of an arm that do not require closed-form solutions to the budget allocation problem. Since one of our budget allocation problems is analogous to the optimal experiment design with unknown variances, we believe that our results are of a broad interest. Our experiments validate our theory, and show that SHVar and SHAdaVar outperform algorithms from prior works with analytical guarantees.

Nash Q-learning may be considered one of the first and most known algorithms in multi-agent reinforcement learning (MARL) for learning policies that constitute a Nash equilibrium of an underlying general-sum Markov game. Its original proof provided asymptotic guarantees and was for the tabular case. Recently, finite-sample guarantees have been provided using more modern RL techniques for the tabular case. Our work analyzes Nash Q-learning using linear function approximation – a representation regime introduced when the state space is large or continuous – and provides finite-sample guarantees that indicate its sample efficiency. We find that the obtained performance nearly matches an existing efficient result for single-agent RL under the same representation and has a polynomial gap when compared to the best-known result for the tabular case.

One fundamental problem in machine learning is out-of-distribution generalization. A method named the surgery estimator incorporates the causal structure in the form of a directed acyclic graph (DAG) to find predictors that are invariant across target domains using distributional invariances via Pearl’s do-calculus. However, finding a surgery estimator can take exponential time as the current methods need to search through all possible predictors. In this work, we first provide a graphical characterization of the identifiability of conditional causal queries. Next, we leverage this characterization together with a greedy search step to develop a polynomial-time algorithm for finding invariant predictors using the causal graph. Given the correct causal graph, our method is guaranteed to find at least one invariant predictor, if it exists. We show that our proposed algorithm can significantly reduce the run-time both in simulated and semi-synthetic data experiments and have predictive performance that is comparable to the existing work that runs in exponential time.

Learning an effective global model on private and decentralized datasets has become an increasingly important challenge of machine learning when applied in practice. Existing distributed learning paradigms, such as Federated Learning, enable this via model aggregation which enforces a strong form of modeling homogeneity and synchronicity across clients. This is however not suitable to many practical scenarios. For example, in distributed sensing, heterogeneous sensors reading data from different views of the same phenomenon would need to use different models for different data modalities. Local learning therefore happens in isolation but inference requires merging the local models to achieve consensus. To enable consensus among local models, we propose a feature fusion approach that extracts local representations from local models and incorporates them into a global representation that improves the prediction performance. Achieving this requires addressing two non-trivial problems. First, we need to learn an alignment between similar feature components which are arbitrarily arranged across clients to enable representation aggregation. Second, we need to learn a consensus graph that captures the high-order interactions between local feature spaces and how to combine them to achieve a better prediction. This paper presents solutions to these problems and demonstrates them in real-world applications on time series data such as power grids and traffic networks.

In the emerging paradigm of Federated Learning (FL), large amount of clients such as mobile devices are used to train possibly high-dimensional models on their respective data. Combining (dimension-wise) adaptive gradient methods (e.g., Adam, AMSGrad) with FL has been an active direction, which is shown to outperform traditional SGD based FL in many cases. In this paper, we focus on the problem of training federated deep neural networks, and propose a novel FL framework which further introduces layer-wise adaptivity to the local model updates to accelerate the convergence of adaptive FL methods. Our framework includes two variants based on two recent locally adaptive federated learning algorithms. Theoretically, we provide a convergence analysis of our layer-wise FL methods, coined Fed-LAMB and Mime-LAMB, which match the convergence rate of state-of-the-art results in adaptive FL and exhibits linear speedup in terms of the number of workers. Experimental results on various datasets and models, under both IID and non-IID local data settings, show that both Fed-LAMB and Mime-LAMB achieve faster convergence speed and better generalization performance, compared to various recent adaptive FL methods.

We introduce the Kernel Calibration Conditional Stein Discrepancy test (KCCSD test), a nonparametric, kernel-based test for assessing the calibration of probabilistic models with well-defined scores. In contrast to previous methods, our test avoids the need for possibly expensive expectation approximations while providing control over its type-I error. We achieve these improvements by using a new family of kernels for score-based probabilities that can be estimated without probability density samples, and by using a Conditional Goodness of Fit criterion for the KCCSD test’s U-statistic. We demonstrate the properties of our test on various synthetic settings.

Cooperative multi-agent reinforcement learning (MARL), where agents coordinates with teammate(s) for a shared goal, may sustain non-stationary caused by the policy change of teammates. Prior works mainly concentrate on the policy change cross episodes, ignoring the fact that teammates may suffer from sudden policy change within an episode, which might lead to miscoordination and poor performance. We formulate the problem as an open Dec-POMDP, where we control some agents to coordinate with uncontrolled teammates, whose policies could be changed within one episode. Then we develop a new framework \textit{\textbf{Fas}t \textbf{t}eammates \textbf{a}da\textbf{p}tation (\textbf{Fastap})} to address the problem. Concretely, we first train versatile teammates’ policies and assign them to different clusters via the Chinese Restaurant Process (CRP). Then, we train the controlled agent(s) to coordinate with the sampled uncontrolled teammates by capturing their identifications as context for fast adaptation. Finally, each agent applies its local information to anticipate the teammates’ context for decision-making accordingly. This process proceeds alternately, leading to a robust policy that can adapt to any teammates during the decentralized execution phase. We show in multiple multi-agent benchmarks that Fastap can achieve superior performance than multiple baselines in stationary and non-stationary scenarios.

Client selection schemes are widely adopted to handle the communication-efficient problems in recent studies of Federated Learning (FL). However, the large variance of the model updates aggregated from the randomly-selected unrepresentative subsets directly slows the FL convergence. We present a novel clustering-based client selection scheme to accelerate the FL convergence by variance reduction. Simple yet effective schemes are designed to improve the clustering effect and control the effect fluctuation, therefore, generating the client subset with certain representativeness of sampling. Theoretically, we demonstrate the improvement of the proposed scheme in variance reduction. We also present the tighter convergence guarantee of the proposed method thanks to the variance reduction. Experimental results confirm the exceed efficiency of our scheme compared to alternatives.