AAAI 2022 Papers

Graph Convolutional Networks (GCNs) have emerged as the state-of-the-art deep learning model for representation learning on graphs. However, it remains notoriously challenging to train and inference GCNs over large graph datasets, limiting their application to large real-world graphs and hindering the exploration of deeper and more sophisticated GCN graphs. This is because as the graph size grows, the sheer number of node features and the large adjacency matrix can easily explode the required memory and data movements. To tackle the aforementioned challenges, we explore the possibility of drawing lottery tickets when sparsifying GCN graphs, i.e., subgraphs that largely shrink the adjacency matrix yet are capable of achieving accuracy comparable to or even better than their full graphs. Specifically, we for the first time discover the existence of graph early-bird (GEB) tickets that emerge at the very early stage when sparsifying GCN graphs, and propose a simple yet effective detector to automatically identify the emergence of such GEB tickets. Furthermore, we advocate graph-model co-optimization and develop a generic efficient GCN early-bird training framework dubbed GEBT that can significantly boost the efficiency of GCN training by (1) drawing joint early-bird tickets between the GCN graphs and models and (2) enabling simultaneously sparsification of both the GCN graphs and models. Experiments on various GCN models and datasets consistently validate our GEB finding and the effectiveness of our GEBT, e.g., our GEBT achieves up to 80.2% ~ 85.6% and 84.6% ~ 87.5% savings of GCN training and inference costs while offering a comparable or even better accuracy as compared to state-of-the-art methods. Our source code and supplementary appendix are available at https://github.com/RICE-EIC/Early-Bird-GCN.

Tensor factorization and distanced based models play important roles in knowledge graph completion (KGC). However, the relational matrices in KGC methods often induce a high model complexity, bearing a high risk of overfitting. As a remedy, researchers propose a variety of different regularizers such as the tensor nuclear norm regularizer. Our motivation is based on the observation that the previous work only focuses on the “size” of the parametric space, while leaving the implicit semantic information widely untouched. To address this issue, we propose a new regularizer, namely, Equivariance Regularizer (ER), which can suppress overfitting by leveraging the implicit semantic information. Specifically, ER can enhance the generalization ability of the model by employing the semantic equivariance between the head and tail entities. Moreover, it is a generic solution for both distance based models and tensor factorization based models. Our experimental results indicate a clear and substantial improvement over the state-of-the-art relation prediction methods.

Emergency vehicles (EMVs) play a crucial role in responding to time-critical events such as medical emergencies and fire outbreaks in an urban area. The less time EMVs spend traveling through the traffic, the more likely it would help save people's lives and reduce property loss. To reduce the travel time of EMVs, prior work has used route optimization based on historical traffic-flow data and traffic signal pre-emption based on the optimal route. However, traffic signal pre-emption dynamically changes the traffic flow which, in turn, modifies the optimal route of an EMV. In addition, traffic signal pre-emption practices usually lead to significant disturbances in traffic flow and subsequently increase the travel time for non-EMVs. In this paper, we propose EMVLight, a decentralized reinforcement learning (RL) framework for simultaneous dynamic routing and traffic signal control. EMVLight extends Dijkstra's algorithm to efficiently update the optimal route for the EMVs in real-time as it travels through the traffic network. The decentralized RL agents learn network-level cooperative traffic signal phase strategies that not only reduce EMV travel time but also reduce the average travel time of non-EMVs in the network. This benefit has been demonstrated through comprehensive experiments with synthetic and real-world maps. These experiments show that EMVLight outperforms benchmark transportation engineering techniques and existing RL-based signal control methods.

This paper describes an energy-based learning method that predicts the activities of multiple agents simultaneously. It aims to forecast both upcoming actions and paths of all agents in a scene based on their past activities, which can be jointly formulated by a probabilistic model over time. Learning this model is challenging because: 1) it has a large number of time-dependent variables that must scale with the forecast horizon and the number of agents; 2) distribution functions have to contain multiple modes in order to capture the spatio-temporal complexities of each agent's activities. To address these challenges, we put forth a novel Energy-based Learning approach for Multi-Agent activity forecasting (ELMA) to estimate this complex model via maximum log-likelihood estimation. Specifically, by sampling from a sequence of factorized marginalized multi-model distributions, ELMA generates most possible future actions efficiently. Moreover, by graph-based representations, ELMA also explicitly resolves the spatio-temporal dependencies of all agents' activities in a single pass. Our experiments on two large-scale datasets prove that ELMA outperforms recent leading studies by an obvious margin.

Image-guided depth completion aims to generate dense depth maps with sparse depth measurements and corresponding RGB images. Currently, spatial propagation networks (SPNs) are the most popular affinity-based methods in depth completion, but they still suffer from the representation limitation of the fixed affinity and the over smoothing during iterations. Our solution is to estimate independent affinity matrices in each SPN iteration, but it is over-parameterized and heavy calculation.This paper introduces an efficient model that learns the affinity among neighboring pixels with an attention-based, dynamic approach. Specifically, the Dynamic Spatial Propagation Network (DySPN) we proposed makes use of a non-linear propagation model (NLPM). It decouples the neighborhood into parts regarding to different distances and recursively generates independent attention maps to refine these parts into adaptive affinity matrices. Furthermore, we adopt a diffusion suppression (DS) operation so that the model converges at an early stage to prevent over-smoothing of dense depth. Finally, in order to decrease the computational cost required, we also introduce three variations that reduce the amount of neighbors and attentions needed while still retaining similar accuracy. In practice, our method requires less iteration to match the performance of other SPNs and yields better results overall. DySPN outperforms other state-of-the-art (SoTA) methods on KITTI Depth Completion (DC) evaluation by the time of submission and is able to yield SoTA performance in NYU Depth v2 dataset as well.

Classical matrix completion methods focus on data with stationary latent structure and hence are not effective in missing value imputation when the latent structure changes with time. This paper proposes a dynamic nonlinear matrix completion (D-NLMC) method, which is able to recover the missing values of streaming data when the low-dimensional nonlinear latent structure of the data changes with time. The paper provides an efficient approach to updating the nonlinear model dynamically. D-NLMC incorporates the information of new data and remove the information of earlier data recursively. The paper shows that the missing data can be estimated if the change of latent structure is slow enough. Different from existing online or adaptive low-rank matrix completion methods, D-NLMC does not require the local low-rank assumption and is able to adaptively recover high-rank matrices with low-dimensional latent structures. Note that existing high-rank matrix completion methods have high-computational costs and are not applicable to streaming data with varying latent structures, which fortunately can be handled by D-NLMC efficiently and accurately. Numerical results show that D-NLMC outperforms the baselines in real applications.

Landslides refer to occurrences of massive ground movements due to geological (and meteorological) factors, and can have disastrous impact on property, economy, and even lead to loss of life. The advances of remote sensing provide accurate and continuous terrain monitoring, enabling the study and analysis of land deformation which, in turn, can be used for possible landslides forecast. Prior studies either rely on independent observations for displacement prediction or model static land characteristics without considering the subtle interactions between different locations and the dynamic changes of the surface conditions. We present DyLand -- Dynamic Manifold Learning with Normalizing Flows for Land deformation prediction -- a novel framework for learning dynamic structures of terrain surface and improving the performance of land deformation prediction. DyLand models the spatial connections of InSAR measurements and estimates conditional distributions of deformations on the terrain manifold with a novel normalizing flow-based method. Instead of modeling the stable terrains, it incorporates surface permutations and captures the innate dynamics of the land surface while allowing for tractable likelihood estimates on the manifold. Our extensive evaluations on curated InSAR datasets from continuous monitoring of slopes prone to landslides show that DyLand outperforms existing bechmarking models.

Knowledge-based visual question answering (VQA) is a vision-language task that requires an agent to correctly answer image-related questions using knowledge that is not presented in the given image. It is not only a more challenging task than regular VQA but also a vital step towards building a general VQA system. Most existing knowledge-based VQA systems process knowledge and image information similarly and ignore the fact that the knowledge base (KB) contains complete information about a triplet, while the extracted image information might be incomplete as the relations between two objects are missing or wrongly detected. In this paper, we propose a novel model named dynamic knowledge memory enhanced multi-step graph reasoning (DMMGR), which performs explicit and implicit reasoning over a key-value knowledge memory module and a spatial-aware image graph, respectively. Specifically, the memory module learns a dynamic knowledge representation and generates a knowledge-aware question representation at each reasoning step. Then, this representation is used to guide a graph attention operator over the spatial-aware image graph. Our model achieves new state-of-the-art accuracy on the KRVQR and FVQA datasets. We also conduct ablation experiments to prove the effectiveness of each component of the proposed model.

This paper introduces a simple yet effective data-centric approach for the task of improving persona-conditioned dialogue agents. Prior model-centric approaches unquestioningly depend on the raw crowdsourced benchmark datasets such as Persona-Chat. In contrast, we aim to fix annotation artifacts in benchmarking, which is orthogonally applicable to any dialogue model. Specifically, we augment relevant personas to improve dialogue dataset/agent, by leveraging the primal-dual structure of the two tasks, predicting dialogue responses and personas based on each other. Experiments on Persona-Chat show that our approach outperforms pre-trained LMs by an 11.7 point gain in terms of accuracy.

Pseudo bounding boxes from the self-training paradigm are inevitably noisy for semi-supervised object detection. To cope with that, a dual decoupling training framework is proposed in the present study, i.e. clean and noisy data decoupling, and classification and localization task decoupling. In the first decoupling, two-level thresholds are used to categorize pseudo boxes into three groups, i.e. clean backgrounds, noisy foregrounds and clean foregrounds. With a specially designed noise-bypass head focusing on noisy data, backbone networks can extract coarse but diverse information; and meanwhile, an original head learns from clean samples for more precise predictions. In the second decoupling, we take advantage of the two-head structure for better evaluation of localization quality, thus the category label and location of a pseudo box can remain independent of each other during training. The approach of two-level thresholds is also applied to group pseudo boxes into three sections of different location accuracy. We outperform existing works by a large margin on VOC datasets, reaching 54.8 mAP(+1.8), and even up to 55.9 mAP(+1.5) by leveraging MS-COCO train2017 as extra unlabeled data. On MS-COCO benchmark, our method also achieves about 1.0 mAP improvements averaging across protocols compared with the prior state-of-the-art.

With various facial manipulation techniques arising, face forgery detection has drawn growing attention due to security concerns. Previous works always formulate face forgery detection as a classification problem based on cross-entropy loss, which emphasizes category-level differences rather than the essential discrepancies between real and fake faces, limiting model generalization in unseen domains. To address this issue, we propose a novel face forgery detection framework, named Dual Contrastive Learning (DCL), which specially constructs positive and negative paired data and performs designed contrastive learning at different granularities to learn generalized feature representation. Concretely, combined with the hard sample selection strategy, Inter-Instance Contrastive Learning (Inter-ICL) is first proposed to promote task-related discriminative features learning by especially constructing instance pairs. Moreover, to further explore the essential discrepancies, Intra-Instance Contrastive Learning (Intra-ICL) is introduced to focus on the local content inconsistencies prevalent in the forged faces by constructing local region pairs inside instances. Extensive experiments and visualizations on several datasets demonstrate the generalization of our method against the state-of-the-art competitors. Our Code is available at https://github.com/Tencent/TFace.git.

The task of few-shot fine-grained recognition is to classify images belonging to subordinate categories merely depending on few examples. Due to the fine-grained nature, it is desirable to capture subtle but discriminative part-level patterns from limited training data, which makes it a challenging problem. In this paper, to generate fine-grained tailored representations for few-shot recognition, we propose a Dual Attention Network (Dual Att-Net) consisting of two dual branches of both hard- and soft-attentions. Specifically, by producing attention guidance from deep activations of input images, our hard-attention is realized by keeping a few useful deep descriptors and forming them as a bag of multi-instance learning. Since these deep descriptors could correspond to objects' parts, the advantage of modeling as a multi-instance bag is able to exploit inherent correlation of these fine-grained parts. On the other side, a soft attended activation representation can be obtained by applying attention guidance upon original activations, which brings comprehensive attention information as the counterpart of hard-attention. After that, both outputs of dual branches are aggregated as a holistic image embedding w.r.t. input images. By performing meta-learning, we can learn a powerful image embedding in such a metric space to generalize to novel classes. Experiments on three popular fine-grained benchmark datasets show that our Dual Att-Net obviously outperforms other existing state-of-the-art methods.

Existing permutation-invariant methods can be divided into two categories according to the aggregation scope, i.e. global aggregation and local one. Although the global aggregation methods, e. g., PointNet and Deep Sets, get involved in simpler structures, their performance is poorer than the local aggregation ones like PointNet++ and Point Transformer. It remains an open problem whether there exists a global aggregation method with a simple structure, competitive performance, and even much fewer parameters. In this paper, we propose a novel global aggregation permutation-invariant network based on dual MLP dot-product, called DuMLP-Pin, which is capable of being employed to extract features for set inputs, including unordered or unstructured pixel, attribute, and point cloud data sets. We strictly prove that any permutation-invariant function implemented by DuMLP-Pin can be decomposed into two or more permutation-equivariant ones in a dot-product way as the cardinality of the given input set is greater than a threshold. We also show that the DuMLP-Pin can be viewed as Deep Sets with strong constraints under certain conditions. The performance of DuMLP-Pin is evaluated on several different tasks with diverse data sets. The experimental results demonstrate that our DuMLP-Pin achieves the best results on the two classification problems for pixel sets and attribute sets. On both the point cloud classification and the part segmentation, the accuracy of DuMLP-Pin is very close to the so-far best-performing local aggregation method with only a 1-2% difference, while the number of required parameters is significantly reduced by more than 85% in classification and 69% in segmentation, respectively. The code is publicly available on https://github.com/JaronTHU/DuMLP-Pin.

Given a first-order sentence ? and a domain size n, how can one sample a model of ? on the domain {1, . . . , n} efficiently as n scales? We consider two variants of this problem: the uniform sampling regime, in which the goal is to sample a model uniformly at random, and the symmetric weighted sampling regime, in which models are weighted according to the number of groundings of each predicate appearing in them. Solutions to this problem have applications to the scalable generation of combinatorial structures, as well as sampling in several statistical-relational models such as Markov logic networks and probabilistic logic programs. In this paper, we identify certain classes of sentences that are domain-liftable under sampling, in the sense that they admit a sampling algorithm that runs in time polynomial in n. In particular, we prove that every sentence of the form ∀x∀y: ?(x, y) for some quantifier-free formula ?(x,y) is domain-liftable under sampling. We then further show that this result continues to hold in the presence of one or more cardinality constraints as well as a single tree axiom constraint.

Sketch-based 3D shape retrieval is a challenging task due to the large domain discrepancy between sketches and 3D shapes. Since existing methods are trained and evaluated on the same categories, they cannot effectively recognize the categories that have not been used during training. In this paper, we propose a novel domain disentangled generative adversarial network (DD-GAN) for zero-shot sketch-based 3D retrieval, which can retrieve the unseen categories that are not accessed during training. Specifically, we first generate domain-invariant features and domain-specific features by disentangling the learned features of sketches and 3D shapes, where the domain-invariant features are used to align with the corresponding word embeddings. Then, we develop a generative adversarial network that combines the domain-specific features of the seen categories with the aligned domain-invariant features to synthesize samples, where the synthesized samples of the unseen categories are generated by using the corresponding word embeddings. Finally, we use the synthesized samples of the unseen categories combined with the real samples of the seen categories to train the network for retrieval, so that the unseen categories can be recognized. In order to reduce the domain shift problem, we utilize unlabeled unseen samples to enhance the discrimination ability of the discriminator. With the discriminator distinguishing the generated samples from the unlabeled unseen samples, the generator can generate more realistic unseen samples. Extensive experiments on the SHREC'13 and SHREC'14 datasets show that our method significantly improves the retrieval performance of the unseen categories.

Feature attribution methods are popular in interpretable machine learning. These methods compute the attribution of each input feature to represent its importance, but there is no consensus on the definition of "attribution", leading to many competing methods with little systematic evaluation, complicated in particular by the lack of ground truth attribution. To address this, we propose a dataset modification procedure to induce such ground truth. Using this procedure, we evaluate three common methods: saliency maps, rationales, and attentions. We identify several deficiencies and add new perspectives to the growing body of evidence questioning the correctness and reliability of these methods applied on datasets in the wild. We further discuss possible avenues for remedy and recommend new attribution methods to be tested against ground truth before deployment. The code and appendix are available at https://yilunzhou.github.io/feature-attribution-evaluation/.

Clustering is important for domain adaptive person re-identification(re-ID). A majority of unsupervised domain adaptation (UDA) methods conduct clustering on the target domain and then use the generated pseudo labels for adaptive training. Albeit important, the clustering pipeline adopted by current literature is quite standard and lacks consideration for two characteristics of re-ID, i.e., 1) a single person has various feature distribution in multiple cameras. 2) a person’s occurrence in the same camera are usually temporally continuous. We argue that the multi-camera distribution hinders clustering because it enlarges the intra-class distances. In contrast, the temporal continuity prior is beneficial, because it offers clue for distinguishing some look-alike person (who are temporally far away from each other). These two insight motivate us to propose a novel Divide-And-Regroup Clustering (DARC) pipeline for re-ID UDA. Specifically, DARC divides the unlabeled data into multiple camera-specific groups and conducts local clustering within each camera. Afterwards, it regroups those local clusters potentially belonging to the same person into a unity. Through this divide-and-regroup pipeline, DARC avoids directly clustering across multiple cameras and focuses on the feature distribution within each individual camera. Moreover, during the local clustering, DARC uses the temporal continuity prior to distinguish some look-alike person and thus reduces false positive pseudo labels. Consequentially, DARC effectively reduces clustering errors and improves UDA. Importantly, we show that DARC is compatible to many pseudo label-based UDA methods and brings general improvement. Based on a recent UDA method, DARC advances the state of the art (e.g, 85.1% mAP on MSMT-to-Market and 83.1% mAP on PersonX-to-Market).

To interpret uncertainty estimates from differentiable probabilistic models, recent work has proposed generating a single Counterfactual Latent Uncertainty Explanation (CLUE) for a given data point where the model is uncertain. We broaden the exploration to examine δ-CLUE, the set of potential CLUEs within a δ ball of the original input in latent space. We study the diversity of such sets and find that many CLUEs are redundant; as such, we propose DIVerse CLUE (∇-CLUE), a set of CLUEs which each propose a distinct explanation as to how one can decrease the uncertainty associated with an input. We then further propose GLobal AMortised CLUE (GLAM-CLUE), a distinct, novel method which learns amortised mappings that apply to specific groups of uncertain inputs, taking them and efficiently transforming them in a single function call into inputs for which a model will be certain. Our experiments show that δ-CLUE, ∇-CLUE, and GLAM-CLUE all address shortcomings of CLUE and provide beneficial explanations of uncertainty estimates to practitioners.

Occlusion is common in the actual 3D scenes, causing the boundary ambiguity of the targeted object. This uncertainty brings difficulty for labeling and learning. Current 3D detectors predict the bounding box directly, regarding it as Dirac delta distribution. However, it does not fully consider such ambiguity. To deal with it, distribution learning is used to efficiently represent the boundary ambiguity. In this paper, we revise the common regression method by predicting the distribution of the 3D box and then present a distribution-aware regression (DAR) module for box refinement and localization quality estimation. It contains scale adaptive (SA) encoder and joint localization quality estimator (JLQE). With the adaptive receptive field, SA encoder refines discriminative features for precise distribution learning. JLQE provides a reliable location score by further leveraging the distribution statistics, correlating with the localization quality of the targeted object. Combining DAR module and the baseline VoteNet, we propose a novel 3D detector called DAVNet. Extensive experiments on both ScanNet V2 and SUN RGB-D datasets demonstrate that the proposed DAVNet achieves significant improvement and outperforms state-of-the-art 3D detectors.