AAAI 2023 论文列表

Graph anomaly detection (GAD) is a vital task in graph-based machine learning and has been widely applied in many real-world applications. The primary goal of GAD is to capture anomalous nodes from graph datasets, which evidently deviate from the majority of nodes. Recent methods have paid attention to various scales of contrastive strategies for GAD, i.e., node-subgraph and node-node contrasts. However, they neglect the subgraph-subgraph comparison information which the normal and abnormal subgraph pairs behave differently in terms of embeddings and structures in GAD, resulting in sub-optimal task performance. In this paper, we fulfill the above idea in the proposed multi-view multi-scale contrastive learning framework with subgraph-subgraph contrast for the first practice. To be specific, we regard the original input graph as the first view and generate the second view by graph augmentation with edge modifications. With the guidance of maximizing the similarity of the subgraph pairs, the proposed subgraph-subgraph contrast contributes to more robust subgraph embeddings despite of the structure variation. Moreover, the introduced subgraph-subgraph contrast cooperates well with the widely-adopted node-subgraph and node-node contrastive counterparts for mutual GAD performance promotions. Besides, we also conduct sufficient experiments to investigate the impact of different graph augmentation approaches on detection performance. The comprehensive experimental results well demonstrate the superiority of our method compared with the state-of-the-art approaches and the effectiveness of the multi-view subgraph pair contrastive strategy for the GAD task. The source code is released at https://github.com/FelixDJC/GRADATE.

Cold temperatures during fall and spring have the potential to cause frost damage to grapevines and other fruit plants, which can significantly decrease harvest yields. To help prevent these losses, farmers deploy expensive frost mitigation measures, such as, sprinklers, heaters, and wind machines, when they judge that damage may occur. This judgment, however, is challenging because the cold hardiness of plants changes throughout the dormancy period and it is difficult to directly measure. This has led scientists to develop cold hardiness prediction models that can be tuned to different grape cultivars based on laborious field measurement data. In this paper, we study whether deep-learning models can improve cold hardiness prediction for grapes based on data that has been collected over a 30-year time period. A key challenge is that the amount of data per cultivar is highly variable, with some cultivars having only a small amount. For this purpose, we investigate the use of multi-task learning to leverage data across cultivars in order to improve prediction performance for individual cultivars. We evaluate a number of multi-task learning approaches and show that the highest performing approach is able to significantly improve over learning for single cultivars and outperforms the current state-of-the-art scientific model for most cultivars.

We study online convex optimization with constraints consisting of multiple functional constraints and a relatively simple constraint set, such as a Euclidean ball. As enforcing the constraints at each time step through projections is computationally challenging in general, we allow decisions to violate the functional constraints but aim to achieve a low regret and cumulative violation of the constraints over a horizon of T time steps. First-order methods achieve an O(sqrt{T}) regret and an O(1) constraint violation, which is the best-known bound under the Slater's condition, but do not take into account the structural information of the problem. Furthermore, the existing algorithms and analysis are limited to Euclidean space. In this paper, we provide an instance-dependent bound for online convex optimization with complex constraints obtained by a novel online primal-dual mirror-prox algorithm. Our instance-dependent regret is quantified by the total gradient variation V_*(T) in the sequence of loss functions. The proposed algorithm works in general normed spaces and simultaneously achieves an O(sqrt{V_*(T)}) regret and an O(1) constraint violation, which is never worse than the best-known (O(sqrt{T}), O(1)) result and improves over previous works that applied mirror-prox-type algorithms for this problem achieving O(T^{2/3}) regret and constraint violation. Finally, our algorithm is computationally efficient, as it only performs mirror descent steps in each iteration instead of solving a general Lagrangian minimization problem.

Scene text image super-resolution (STISR) in the wild has been shown to be beneficial to support improved vision-based text recognition from low-resolution imagery. An intuitive way to enhance STISR performance is to explore the well-structured and repetitive layout characteristics of text and exploit these as prior knowledge to guide model convergence. In this paper, we propose a novel gradient-based graph attention method to embed patch-wise text layout contexts into image feature representations for high-resolution text image reconstruction in an implicit and elegant manner. We introduce a non-local group-wise attention module to extract text features which are then enhanced by a cascaded channel attention module and a novel gradient-based graph attention module in order to obtain more effective representations by exploring correlations of regional and local patch-wise text layout properties. Extensive experiments on the benchmark TextZoom dataset convincingly demonstrate that our method supports excellent text recognition and outperforms the current state-of-the-art in STISR. The source code is available at https://github.com/xyzhu1/TSAN.

Constrained Reinforcement Learning (CRL) burgeons broad interest in recent years, which pursues maximizing long-term returns while constraining costs. Although CRL can be cast as a multi-objective optimization problem, it is still facing the key challenge that gradient-based Pareto optimization methods tend to stick to known Pareto-optimal solutions even when they yield poor returns (e.g., the safest self-driving car that never moves) or violate the constraints (e.g., the record-breaking racer that crashes the car). In this paper, we propose Gradient-adaptive Constrained Policy Optimization (GCPO for short), a novel Pareto optimization method for CRL with two adaptive gradient recalibration techniques. First, to find Pareto-optimal solutions with balanced performance over all targets, we propose gradient rebalancing which forces the agent to improve more on under-optimized objectives at every policy iteration. Second, to guarantee that the cost constraints are satisfied, we propose gradient perturbation that can temporarily sacrifice the returns for costs. Experiments on the SafetyGym benchmarks show that our method consistently outperforms previous CRL methods in reward while satisfying the constraints.

Gradient estimation is often necessary for fitting generative models with discrete latent variables, in contexts such as reinforcement learning and variational autoencoder (VAE) training. The DisARM estimator achieves state of the art gradient variance for Bernoulli latent variable models in many contexts. However, DisARM and other estimators have potentially exploding variance near the boundary of the parameter space, where solutions tend to lie. To ameliorate this issue, we propose a new gradient estimator bitflip-1 that is lower variance at the boundaries of the parameter space. As bitflip-1 has complementary properties to existing estimators, we introduce an aggregated estimator, unbiased gradient variance clipping (UGC) that uses either a bitflip-1 or a DisARM gradient update for each coordinate. We theoretically prove that UGC has uniformly lower variance than DisARM. Empirically, we observe that UGC achieves the optimal value of the optimization objectives in toy experiments, discrete VAE training, and in a best subset selection problem.

Detecting objects as multiple keypoints is an important approach in the anchor-free object detection methods while corner pooling is an effective feature encoding method for corner positioning. The corners of the bounding box are located by summing the feature maps which are max-pooled in the x and y directions respectively by corner pooling. In the unidirectional max pooling operation, the features of the densely arranged objects of the same class are prone to occlusion. To this end, we propose a method named Gradient Corner Pooling. The spatial distance information of objects on the feature map is encoded during the unidirectional pooling process, which effectively alleviates the occlusion of the homogeneous object features. Further, the computational complexity of gradient corner pooling is the same as traditional corner pooling and hence it can be implemented efficiently. Gradient corner pooling obtains consistent improvements for various keypoint-based methods by directly replacing corner pooling. We verify the gradient corner pooling algorithm on the dataset and in real scenarios, respectively. The networks with gradient corner pooling located the corner points earlier in the training process and achieve an average accuracy improvement of 0.2%-1.6% on the MS-COCO dataset. The detectors with gradient corner pooling show better angle adaptability for arrayed objects in the actual scene test.

Positive-unlabeled learning is an essential problem in many real-world applications with only labeled positive and unlabeled data, especially when the negative samples are difficult to identify. Most existing positive-unlabeled learning methods will inevitably overfit the positive class to some extent due to the existence of unidentified positive samples. This paper first analyzes the overfitting problem and proposes to bound the generalization errors via Wasserstein distances. Based on that, we develop a simple yet effective positive-unlabeled learning method, GradPU, which consists of two key ingredients: A gradient-based regularizer that penalizes the gradient norms in the interpolated data region, which improves the generalization of positive class; An unnormalized upweighting mechanism that assigns larger weights to those positive samples that are hard, not-well-fitted and less frequently labeled. It enforces the training error of each positive sample to be small and increases the robustness to the labeling bias. We evaluate our proposed GradPU on three datasets: MNIST, FashionMNIST, and CIFAR10. The results demonstrate that GradPU achieves state-of-the-art performance on both unbiased and biased positive labeling scenarios.

It has been witnessed that masked image modeling (MIM) has shown a huge potential in self-supervised learning in the past year. Benefiting from the universal backbone vision transformer, MIM learns self-supervised visual representations through masking a part of patches of the image while attempting to recover the missing pixels. Most previous works mask patches of the image randomly, which underutilizes the semantic information that is beneficial to visual representation learning. On the other hand, due to the large size of the backbone, most previous works have to spend much time on pre-training. In this paper, we propose Attention-driven Masking and Throwing Strategy (AMT), which could solve both problems above. We first leverage the self-attention mechanism to obtain the semantic information of the image during the training process automatically without using any supervised methods. Masking strategy can be guided by that information to mask areas selectively, which is helpful for representation learning. Moreover, a redundant patch throwing strategy is proposed, which makes learning more efficient. As a plug-and-play module for masked image modeling, AMT improves the linear probing accuracy of MAE by 2.9% ~ 5.9% on CIFAR-10/100, STL-10, Tiny ImageNet, and ImageNet-1K, and obtains an improved performance with respect to fine-tuning accuracy of MAE and SimMIM. Moreover, this design also achieves superior performance on downstream detection and segmentation tasks.

Many applications of reinforcement learning can be formalized as goal-conditioned environments, where, in each episode, there is a "goal" that affects the rewards obtained during that episode but does not affect the dynamics. Various techniques have been proposed to improve performance in goal-conditioned environments, such as automatic curriculum generation and goal relabeling. In this work, we explore a connection between off-policy reinforcement learning in goal-conditioned settings and knowledge distillation. In particular: the current Q-value function and the target Q-value estimate are both functions of the goal, and we would like to train the Q-value function to match its target for all goals. We therefore apply Gradient-Based Attention Transfer (Zagoruyko and Komodakis 2017), a knowledge distillation technique, to the Q-function update. We empirically show that this can improve the performance of goal-conditioned off-policy reinforcement learning when the space of goals is high-dimensional. We also show that this technique can be adapted to allow for efficient learning in the case of multiple simultaneous sparse goals, where the agent can attain a reward by achieving any one of a large set of objectives, all specified at test time. Finally, to provide theoretical support, we give examples of classes of environments where (under some assumptions) standard off-policy algorithms such as DDPG require at least O(d^2) replay buffer transitions to learn an optimal policy, while our proposed technique requires only O(d) transitions, where d is the dimensionality of the goal and state space. Code and appendix are available at https://github.com/alevine0/ReenGAGE.

Goal-conditioned Reinforcement Learning (RL) aims at learning optimal policies, given goals encoded in special command inputs. Here we study goal-conditioned neural nets (NNs) that learn to generate deep NN policies in form of context-specific weight matrices, similar to Fast Weight Programmers and other methods from the 1990s. Using context commands of the form ``generate a policy that achieves a desired expected return,'' our NN generators combine powerful exploration of parameter space with generalization across commands to iteratively find better and better policies. A form of weight-sharing HyperNetworks and policy embeddings scales our method to generate deep NNs. Experiments show how a single learned policy generator can produce policies that achieve any return seen during training. Finally, we evaluate our algorithm on a set of continuous control tasks where it exhibits competitive performance. Our code is public.

The transferability of adversarial examples is the key property in practical black-box scenarios. Currently, numerous methods improve the transferability across different models trained on the same modality of data. The investigation of generating video adversarial examples with imagebased substitute models to attack the target video models, i.e., cross-modal transferability of adversarial examples, is rarely explored. A few works on cross-modal transferability directly apply image attack methods for each frame and no factors especial for video data are considered, which limits the cross-modal transferability of adversarial examples. In this paper, we propose an effective cross-modal attack method which considers both the global and local characteristics of video data. Firstly, from the global perspective, we introduce inter-frame interaction into attack process to induce more diverse and stronger gradients rather than perturb each frame separately. Secondly, from the local perspective, we disrupt the inherently local correlation of frames within a video, which prevents black-box video model from capturing valuable temporal clues. Extensive experiments on the UCF-101 and Kinetics-400 validate the proposed method significantly improves cross-modal transferability and even surpasses strong baseline using video models as substitute model. Our source codes are available at https://github.com/lwmming/Cross-Modal-Attack.

Data augmentation with Mixup has been proven an effective method to regularize the current deep neural networks. Mixup generates virtual samples and corresponding labels simultaneously by linear interpolation. However, the one-stage generation paradigm and the use of linear interpolation have two defects: (1) The label of the generated sample is simply combined from the labels of the original sample pairs without reasonable judgment, resulting in ambiguous labels. (2) Linear combination significantly restricts the sampling space for generating samples. To address these issues, we propose a novel and effective augmentation method, Global Mixup, based on global clustering relationships. Specifically, we transform the previous one-stage augmentation process into two-stage by decoupling the process of generating virtual samples from the labeling. And for the labels of the generated samples, relabeling is performed based on clustering by calculating the global relationships of the generated samples. Furthermore, we are no longer restricted to linear relationships, which allows us to generate more reliable virtual samples in a larger sampling space. Extensive experiments for CNN, LSTM, and BERT on five tasks show that Global Mixup outperforms previous baselines. Further experiments also demonstrate the advantage of Global Mixup in low-resource scenarios.

We hypothesize that deep network classifications of complex scenes can be explained using sets of relevant objects. We employ beam search and singular value decomposition to generate local and global explanations that summarize the deep model's interpretation of a class.

Self Attention has shown the excellent performance in tracking due to its global modeling capability. However, it brings two challenges: First, its global receptive field has less attention on local structure and inter-channel associations, which limits the semantics to distinguish objects and backgrounds; Second, its feature fusion with linear process cannot avoid the interference of non-target semantic objects. To solve the above issues, this paper proposes a robust tracking method named GdaTFT by defining the Global Dilated Attention (GDA) and Target Focusing Network (TFN). The GDA provides a new global semantics modeling approach to enhance the semantic objects while eliminating the background. It is defined via the local focusing module, dilated attention and channel adaption module. Thus, it promotes semantics by focusing local key information, building long-range dependencies and enhancing the semantics of channels. Subsequently, to distinguish the target and non-target objects both with rich semantics, the TFN is proposed to accurately focus the target region. Different from the present feature fusion, it uses the template as the query to build a point-to-point correlation between the template and search region, and finally achieves part-level augmentation of target feature in the search region. Thus, the TFN efficiently augments the target embedding while weakening the non-target objects. Experiments on challenging benchmarks (LaSOT, TrackingNet, GOT-10k, OTB-100) demonstrate that the GdaTFT outperforms many state-of-the-art trackers and achieves leading performance. Code will be available.

The actor-critic (AC) reinforcement learning algorithms have been the powerhouse behind many challenging applications. Nevertheless, its convergence is fragile in general. To study its instability, existing works mostly consider the uncommon double-loop variant or basic models with finite state and action space. We investigate the more practical single-sample two-timescale AC for solving the canonical linear quadratic regulator (LQR) problem, where the actor and the critic update only once with a single sample in each iteration on an unbounded continuous state and action space. Existing analysis cannot conclude the convergence for such a challenging case. We develop a new analysis framework that allows establishing the global convergence to an epsilon-optimal solution with at most an order of epsilon to -2.5 sample complexity. To our knowledge, this is the first finite-time convergence analysis for the single sample two-timescale AC for solving LQR with global optimality. The sample complexity improves those of other variants by orders, which sheds light on the practical wisdom of single sample algorithms. We also further validate our theoretical findings via comprehensive simulation comparisons.

Graph neural networks (GNNs) are highly effective on a variety of graph-related tasks; however, they lack interpretability and transparency. Current explainability approaches are typically local and treat GNNs as black-boxes. They do not look inside the model, inhibiting human trust in the model and explanations. Motivated by the ability of neurons to detect high-level semantic concepts in vision models, we perform a novel analysis on the behaviour of individual GNN neurons to answer questions about GNN interpretability. We propose a novel approach for producing global explanations for GNNs using neuron-level concepts to enable practitioners to have a high-level view of the model. Specifically, (i) to the best of our knowledge, this is the first work which shows that GNN neurons act as concept detectors and have strong alignment with concepts formulated as logical compositions of node degree and neighbourhood properties; (ii) we quantitatively assess the importance of detected concepts, and identify a trade-off between training duration and neuron-level interpretability; (iii) we demonstrate that our global explainability approach has advantages over the current state-of-the-art -- we can disentangle the explanation into individual interpretable concepts backed by logical descriptions, which reduces potential for bias and improves user-friendliness.

Measuring and alleviating the discrepancies between the synthetic (source) and real scene (target) data is the core issue for domain adaptive semantic segmentation. Though recent works have introduced depth information in the source domain to reinforce the geometric and semantic knowledge transfer, they cannot extract the intrinsic 3D information of objects, including positions and shapes, merely based on 2D estimated depth. In this work, we propose a novel Geometry-Aware Network for Domain Adaptation (GANDA), leveraging more compact 3D geometric point cloud representations to shrink the domain gaps. In particular, we first utilize the auxiliary depth supervision from the source domain to obtain the depth prediction in the target domain to accomplish structure-texture disentanglement. Beyond depth estimation, we explicitly exploit 3D topology on the point clouds generated from RGB-D images for further coordinate-color disentanglement and pseudo-labels refinement in the target domain. Moreover, to improve the 2D classifier in the target domain, we perform domain-invariant geometric adaptation from source to target and unify the 2D semantic and 3D geometric segmentation results in two domains. Note that our GANDA is plug-and-play in any existing UDA framework. Qualitative and quantitative results demonstrate that our model outperforms state-of-the-arts on GTA5->Cityscapes and SYNTHIA->Cityscapes.

The model identifiability is a considerable issue in the unsupervised learning of disentangled representations. The PCA inductive biases revealed recently for unsupervised disentangling in VAE-based models are shown to improve local alignment of latent dimensions with principal components of the data. In this paper, in additional to the PCA inductive biases, we propose novel geometric inductive biases from the manifold perspective for unsupervised disentangling, which induce the model to capture the global geometric properties of the data manifold with guaranteed model identifiability. We also propose a Geometric Disentangling Regularized AutoEncoder (GDRAE) that combines the PCA and the proposed geometric inductive biases in one unified framework. The experimental results show the usefulness of the geometric inductive biases in unsupervised disentangling and the effectiveness of our GDRAE in capturing the geometric inductive biases.