AAAI 2023 论文列表

In this paper, we improve the kernel alignment regret bound for online kernel learning in the regime of the Hinge loss function. Previous algorithm achieves a regret of O((A_TT ln T)^{1/4}) at a computational complexity (space and per-round time) of O((A_TT ln T)^{1/2}), where A_T is called kernel alignment. We propose an algorithm whose regret bound and computational complexity are better than previous results. Our results depend on the decay rate of eigenvalues of the kernel matrix. If the eigenvalues of the kernel matrix decay exponentially, then our algorithm enjoys a regret of O((A_T)^{1/2}) at a computational complexity of O((ln T)^2). Otherwise, our algorithm enjoys a regret of O((A_TT)^{1/4}) at a computational complexity of O((A_TT)^{1/2}). We extend our algorithm to batch learning and obtain a O(T^{-1}(E[A_T])^{1/2}) excess risk bound which improves the previous O(T^{-1/2}) bound.

The Maximum Satisfiability (MAXSAT) problem is an optimization version of the Satisfiability problem (SAT) in which one is given a CNF formula with n variables and needs to find the maximum number of simultaneously satisfiable clauses. Recent works achieved significant progress in proving new upper bounds on the worst-case computational complexity of MAXSAT. All these works reduce general MAXSAT to a special case of MAXSAT where each variable appears a small number of times. So, it is important to design fast algorithms for (n,k)-MAXSAT to construct an efficient exact algorithm for MAXSAT. (n,k)-MAXSAT is a special case of MAXSAT where each variable appears at most k times in the input formula. For the (n,3)-MAXSAT problem, we design a O*(1.1749^n) algorithm improving on the previous record running time of O*(1.191^n). For the (n,4)-MAXSAT problem, we construct a O*(1.3803^n) algorithm improving on the previous best running time of O*(1.4254^n). Using the results, we develop a O*(1.0911^L) algorithm for the MAXSAT where L is a length of the input formula which improves previous algorithm with O*(1.0927^L) running time.

Submodular maximization has attracted extensive attention due to its numerous applications in machine learning and artificial intelligence. Many real-world problems require maximizing multiple submodular objective functions at the same time. In such cases, a common approach is to select a representative subset of Pareto optimal solutions with different trade-offs among multiple objectives. To this end, in this paper, we investigate the regret ratio minimization (RRM) problem in multi-objective submodular maximization, which aims to find at most k solutions to best approximate all Pareto optimal solutions w.r.t. any linear combination of objective functions. We propose a novel HS-RRM algorithm by transforming RRM into HittingSet problems based on the notions of ε-kernel and δ-net, where any α-approximation algorithm for single-objective submodular maximization is used as an oracle. We improve upon the previous best-known bound on the maximum regret ratio (MRR) of the output of HS-RRM and show that the new bound is nearly asymptotically optimal for any fixed number d of objective functions. Experiments on real-world and synthetic data confirm that HS-RRM achieves lower MRRs than existing algorithms.

Physics-informed neural networks (PINNs) have effectively been demonstrated in solving forward and inverse differential equation problems, but they are still trapped in training failures when the target functions to be approximated exhibit high-frequency or multi-scale features. In this paper, we propose to employ implicit stochastic gradient descent (ISGD) method to train PINNs for improving the stability of training process. We heuristically analyze how ISGD overcome stiffness in the gradient flow dynamics of PINNs, especially for problems with multi-scale solutions. We theoretically prove that for two-layer fully connected neural networks with large hidden nodes, randomly initialized ISGD converges to a globally optimal solution for the quadratic loss function. Empirical results demonstrate that ISGD works well in practice and compares favorably to other gradient-based optimization methods such as SGD and Adam, while can also effectively address the numerical stiffness in training dynamics via gradient descent.

Bilevel Optimization Programming is used to model complex and conflicting interactions between agents, for example in Robust AI or Privacy preserving AI. Integrating bilevel mathematical programming within deep learning is thus an essential objective for the Machine Learning community. Previously proposed approaches only consider single-level programming. In this paper, we extend existing single-level optimization programming approaches and thus propose Differentiating through Bilevel Optimization Programming (BiGrad) for end-to-end learning of models that use Bilevel Programming as a layer. BiGrad has wide applicability and can be used in modern machine learning frameworks. BiGrad is applicable to both continuous and combinatorial Bilevel optimization problems. We describe a class of gradient estimators for the combinatorial case which reduces the requirements in terms of computation complexity; for the case of the continuous variable, the gradient computation takes advantage of the push-back approach (i.e. vector-jacobian product) for an efficient implementation. Experiments show that the BiGrad successfully extends existing single-level approaches to Bilevel Programming.

New information in the context of real life settings usually is accompanied by some kind of supplementary information that indicates context, reliability, or expertise of the information's source. Bounded Revision (BR) displays an iterated belief revision mechanism that takes as input a new information accompanied by a reference sentence acting as supplementary information, which specifies the depth with which the new input shall be integrated in the posterior belief state. The reference sentence specifies which worlds in the prior belief state are affected by the change mechanism. We show that Bounded Revision can be characterized by three simple, yet elegant postulates and corresponds to a special case of a lexicographic revision, which inherits all relevant features of BR. Furthermore, we present methodological implementations of BR including conditional revision with c-revisions, making it directly usable for conditional revision tools.

Adding perturbations via utilizing auxiliary gradient information or discarding existing details of the benign images are two common approaches for generating adversarial examples. Though visual imperceptibility is the desired property of adversarial examples, conventional adversarial attacks still generate traceable adversarial perturbations. In this paper, we introduce a novel Adversarial Attack via Invertible Neural Networks (AdvINN) method to produce robust and imperceptible adversarial examples. Specifically, AdvINN fully takes advantage of the information preservation property of Invertible Neural Networks and thereby generates adversarial examples by simultaneously adding class-specific semantic information of the target class and dropping discriminant information of the original class. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet-1K demonstrate that the proposed AdvINN method can produce less imperceptible adversarial images than the state-of-the-art methods and AdvINN yields more robust adversarial examples with high confidence compared to other adversarial attacks. Code is available at https://github.com/jjhuangcs/AdvINN.

Label distribution covers a certain number of labels, representing the degree to which each label describes an instance. The learning process on the instances labeled by label distributions is called Label Distribution Learning (LDL). Although LDL has been applied successfully to many practical applications, one problem with existing LDL methods is that they are limited to data with balanced label information. However, annotation information in real-world data often exhibits imbalanced distributions, which significantly degrades the performance of existing methods. In this paper, we investigate the Imbalanced Label Distribution Learning (ILDL) problem. To handle this challenging problem, we delve into the characteristics of ILDL and empirically find that the representation distribution shift is the underlying reason for the performance degradation of existing methods. Inspired by this finding, we present a novel method named Representation Distribution Alignment (RDA). RDA aligns the distributions of feature representations and label representations to alleviate the impact of the distribution gap between the training set and the test set caused by the imbalance issue. Extensive experiments verify the superior performance of RDA. Our work fills the gap in benchmarks and techniques for practical ILDL problems.

ImageNet pre-training has enabled state-of-the-art results on many tasks. In spite of its recognized contribution to generalization, we observed in this study that ImageNet pre-training also transfers adversarial non-robustness from pre-trained model into fine-tuned model in the downstream classification tasks. We first conducted experiments on various datasets and network backbones to uncover the adversarial non-robustness in fine-tuned model. Further analysis was conducted on examining the learned knowledge of fine-tuned model and standard model, and revealed that the reason leading to the non-robustness is the non-robust features transferred from ImageNet pre-trained model. Finally, we analyzed the preference for feature learning of the pre-trained model, explored the factors influencing robustness, and introduced a simple robust ImageNet pre-training solution. Our code is available at https://github.com/jiamingzhang94/ImageNet-Pretraining-transfers-non-robustness.

Graph contrastive learning (GCL) has attracted a surge of attention due to its superior performance for learning node/graph representations without labels. However, in practice, the underlying class distribution of unlabeled nodes for the given graph is usually imbalanced. This highly imbalanced class distribution inevitably deteriorates the quality of learned node representations in GCL. Indeed, we empirically find that most state-of-the-art GCL methods cannot obtain discriminative representations and exhibit poor performance on imbalanced node classification. Motivated by this observation, we propose a principled GCL framework on Imbalanced node classification (ImGCL), which automatically and adaptively balances the representations learned from GCL without labels. Specifically, we first introduce the online clustering based progressively balanced sampling (PBS) method with theoretical rationale, which balances the training sets based on pseudo-labels obtained from learned representations in GCL. We then develop the node centrality based PBS method to better preserve the intrinsic structure of graphs, by upweighting the important nodes of the given graph. Extensive experiments on multiple imbalanced graph datasets and imbalanced settings demonstrate the effectiveness of our proposed framework, which significantly improves the performance of the recent state-of-the-art GCL methods. Further experimental ablations and analyses show that the ImGCL framework consistently improves the representation quality of nodes in under-represented (tail) classes.

Majority illusion occurs in a social network when the majority of the network vertices belong to a certain type but the majority of each vertex's neighbours belong to a different type, therefore creating the wrong perception, i.e., the illusion, that the majority type is different from the actual one. From a system engineering point of view, this motivates the search for algorithms to detect and, where possible, correct this undesirable phenomenon. In this paper we initiate the computational study of majority illusion in social networks, providing NP-hardness and parametrised complexity results for its occurrence and elimination.

Access to a representative sample from the population is an assumption that underpins all of machine learning. Selection effects can cause observations to instead come from a subpopulation, by which our inferences may be subject to bias. It is therefore important to know whether or not a sample is affected by selection effects. We study under which conditions we can identify selection bias and give results for both parametric and non-parametric families of distributions. Based on these results we develop two practical methods to determine whether or not an observed sample comes from a distribution subject to selection bias. Through extensive evaluation on synthetic and real world data we verify that our methods beat the state of the art both in detecting as well as characterizing selection bias.

Event causality identification (ECI) aims to identify the causal relationship between events, which plays a crucial role in deep text understanding. Due to the diversity of real-world causality events and difficulty in obtaining sufficient training data, existing ECI approaches have poor generalizability and struggle to identify the relation between seldom seen events. In this paper, we propose to utilize both external knowledge and internal analogy to improve ECI. On the one hand, we utilize a commonsense knowledge graph called ConceptNet to enrich the description of an event sample and reveal the commonalities or associations between different events. On the other hand, we retrieve similar events as analogy exam- ples and glean useful experiences from such analogous neigh- bors to better identify the relationship between a new event pair. By better understanding different events through exter- nal knowledge and making an analogy with similar events, we can alleviate the data sparsity issue and improve model gener- alizability. Extensive evaluations on two benchmark datasets show that our model outperforms other baseline methods by around 18% on the F1-value on average

We propose novel identification conditions and a statistical estimation method for the probabilities of potential outcome types using covariate information in randomized trials in which the treatment assignment is randomized but subject compliance is not perfect. Different from existing studies, the proposed identification conditions do not require strict assumptions such as the assumption of monotonicity. When the probabilities of potential outcome types are identifiable through the proposed conditions, the problem of estimating the probabilities of potential outcome types is reduced to that of singular models. Thus, the probabilities cannot be evaluated using standard statistical likelihood-based estimation methods. Rather, the proposed identification conditions show that we can derive consistent estimators of the probabilities of potential outcome types via the method of moments, which leads to the asymptotic normality of the proposed estimators through the delta method under regular conditions. We also propose a new statistical estimation method based on the bounded constrained augmented Lagrangian method to derive more efficient estimators than can be derived through the method of moments.

Recent advancements in Generative Adversarial Networks (GANs) have made it possible to obtain high-quality face images of synthetic identities. These networks see large amounts of real faces in order to learn to generate realistic looking synthetic images. However, the concept of a synthetic identity for these images is not very well-defined. In this work, we verify identity leakage from the training set containing real images into the latent space and propose a novel method, IdProv, that uses image composition to trace the source of identity signals in the generated image.

Unsupervised foreground-background segmentation aims at extracting salient objects from cluttered backgrounds, where Generative Adversarial Network (GAN) approaches, especially layered GANs, show great promise. However, without human annotations, they are typically prone to produce foreground and background layers with non-negligible semantic and visual confusion, dubbed "information leakage", resulting in notable degeneration of the generated segmentation mask. To alleviate this issue, we propose a simple-yet-effective explicit layer independence modeling approach, termed Independent Layer Synthesis GAN (ILSGAN), pursuing independent foreground-background layer generation by encouraging their discrepancy. Specifically, it targets minimizing the mutual information between visible and invisible regions of the foreground and background to spur interlayer independence. Through in-depth theoretical and experimental analyses, we justify that explicit layer independence modeling is critical to suppressing information leakage and contributes to impressive segmentation performance gains. Also, our ILSGAN achieves strong state-of-the-art generation quality and segmentation performance on complex real-world data.

This paper presents an inverse kinematic optimization layer (IKOL) for 3D human pose and shape estimation that leverages the strength of both optimization- and regression-based methods within an end-to-end framework. IKOL involves a nonconvex optimization that establishes an implicit mapping from an image’s 3D keypoints and body shapes to the relative body-part rotations. The 3D keypoints and the body shapes are the inputs and the relative body-part rotations are the solutions. However, this procedure is implicit and hard to make differentiable. So, to overcome this issue, we designed a Gauss-Newton differentiation (GN-Diff) procedure to differentiate IKOL. GN-Diff iteratively linearizes the nonconvex objective function to obtain Gauss-Newton directions with closed form solutions. Then, an automatic differentiation procedure is directly applied to generate a Jacobian matrix for end-to-end training. Notably, the GN-Diff procedure works fast because it does not rely on a time-consuming implicit differentiation procedure. The twist rotation and shape parameters are learned from the neural networks and, as a result, IKOL has a much lower computational overhead than most existing optimization-based methods. Additionally, compared to existing regression-based methods, IKOL provides a more accurate mesh-image correspondence. This is because it iteratively reduces the distance between the keypoints and also enhances the reliability of the pose structures. Extensive experiments demonstrate the superiority of our proposed framework over a wide range of 3D human pose and shape estimation methods. Code is available at https://github.com/Juzezhang/IKOL

Given an untrimmed video, temporal sentence localization (TSL) aims to localize a specific segment according to a given sentence query. Though respectable works have made decent achievements in this task, they severely rely on dense video frame annotations, which require a tremendous amount of human effort to collect. In this paper, we target another more practical and challenging setting: one-shot temporal sentence localization (one-shot TSL), which learns to retrieve the query information among the entire video with only one annotated frame. Particularly, we propose an effective and novel tree-structure baseline for one-shot TSL, called Multiple Hypotheses Segment Tree (MHST), to capture the query-aware discriminative frame-wise information under the insufficient annotations. Each video frame is taken as the leaf-node, and the adjacent frames sharing the same visual-linguistic semantics will be merged into the upper non-leaf node for tree building. At last, each root node is an individual segment hypothesis containing the consecutive frames of its leaf-nodes. During the tree construction, we also introduce a pruning strategy to eliminate the interference of query-irrelevant nodes. With our designed self-supervised loss functions, our MHST is able to generate high-quality segment hypotheses for ranking and selection with the query. Experiments on two challenging datasets demonstrate that MHST achieves competitive performance compared to existing methods.

In this paper, hypernetworks are trained to generate behaviors across a range of unseen task conditions, via a novel TD-based training objective and data from a set of near-optimal RL solutions for training tasks. This work relates to meta RL, contextual RL, and transfer learning, with a particular focus on zero-shot performance at test time, enabled by knowledge of the task parameters (also known as context). Our technical approach is based upon viewing each RL algorithm as a mapping from the MDP specifics to the near-optimal value function and policy and seek to approximate it with a hypernetwork that can generate near-optimal value functions and policies, given the parameters of the MDP. We show that, under certain conditions, this mapping can be considered as a supervised learning problem. We empirically evaluate the effectiveness of our method for zero-shot transfer to new reward and transition dynamics on a series of continuous control tasks from DeepMind Control Suite. Our method demonstrates significant improvements over baselines from multitask and meta RL approaches.