IJCAI 2021 Papers

Securing networked infrastructures is important in the real world. The problem of deploying security resources to protect against an attacker in networked domains can be modeled as Network Security Games (NSGs). Unfortunately, existing approaches, including the deep learning-based approaches, are inefficient to solve large-scale extensive-form NSGs. In this paper, we propose a novel learning paradigm, NSG-NFSP, to solve large-scale extensive-form NSGs based on Neural Fictitious Self-Play (NFSP). Our main contributions include: i) reforming the best response (BR) policy network in NFSP to be a mapping from action-state pair to action-value, to make the calculation of BR possible in NSGs; ii) converting the average policy network of an NFSP agent into a metric-based classifier, helping the agent to assign distributions only on legal actions rather than all actions; iii) enabling NFSP with high-level actions, which can benefit training efficiency and stability in NSGs; and iv) leveraging information contained in graphs of NSGs by learning efficient graph node embeddings. Our algorithm significantly outperforms state-of-the-art algorithms in both scalability and solution quality.

Graph homomorphism problems involve finding adjacency-preserving mappings between two given graphs. Although theoretically hard, these problems can often be solved in practice using constraint programming algorithms. We show how techniques from the state-of-the-art in subgraph isomorphism solving can be applied to broader graph homomorphism problems, and introduce a new form of filtering based upon clique-finding. We demonstrate empirically that this filtering is effective for the locally injective graph homomorphism and subgraph isomorphism problems, and gives the first practical constraint programming approach to finding general graph homomorphisms.

Episodic memory lets reinforcement learning algorithms remember and exploit promising experience from the past to improve agent performance. Previous works on memory mechanisms show benefits of using episodic-based data structures for discrete action problems in terms of sample-efficiency. The application of episodic memory for continuous control with a large action space is not trivial. Our study aims to answer the question: can episodic memory be used to improve agent's performance in continuous control? Our proposed algorithm combines episodic memory with Actor-Critic architecture by modifying critic's objective. We further improve performance by introducing episodic-based replay buffer prioritization. We evaluate our algorithm on OpenAI gym domains and show greater sample-efficiency compared with the state-of-the art model-free off-policy algorithms.

Smart contracts hold digital coins worth billions of dollars, their security issues have drawn extensive attention in the past years. Towards smart contract vulnerability detection, conventional methods heavily rely on fixed expert rules, leading to low accuracy and poor scalability. Recent deep learning approaches alleviate this issue but fail to encode useful expert knowledge. In this paper, we explore combining deep learning with expert patterns in an explainable fashion. Specifically, we develop automatic tools to extract expert patterns from the source code. We then cast the code into a semantic graph to extract deep graph features. Thereafter, the global graph feature and local expert patterns are fused to cooperate and approach the final prediction, while yielding their interpretable weights. Experiments are conducted on all available smart contracts with source code in two platforms, Ethereum and VNT Chain. Empirically, our system significantly outperforms state-of-the-art methods. Our code is released.

In this paper, we present Skills2Graph, a tool that, starting from a set of users’ professional skills, identifies the most suitable jobs as they emerge from a large corpus of 2.5M+ Online Job Vacancies (OJVs) posted in three different countries (the United Kingdom, France, and Germany). To this aim, we rely both on co-occurrence statistics - computing a count-based measure of skill-relevance named Revealed Comparative Advantage (rca) - and distributional semantics - generating several embeddings on the OJVs corpus and performing an intrinsic evaluation of their quality. Results, evaluated through a user study of 10 labor market experts, show a high P@3 for the recommendations provided by Skills2Graph, and a high nDCG (0.985 and 0.984 in a [0,1] range), that indicates a strong correlation between the experts’ scores and the rankings generated by Skills2Graph.

We address the problem of deciding skeptical acceptance wrt. preferred semantics of an argument in abstract argumentation frameworks, i.e., the problem of deciding whether an argument is contained in all maximally admissible sets, a.k.a. preferred extensions. State-of-the-art algorithms solve this problem with iterative calls to an external SAT-solver to determine preferred extensions. We provide a new characterisation of skeptical acceptance wrt. preferred semantics that does not involve the notion of a preferred extension. We then develop a new algorithm that also relies on iterative calls to an external SAT-solver but avoids the costly part of maximising admissible sets. We present the results of an experimental evaluation that shows that this new approach significantly outperforms the state of the art. We also apply similar ideas to develop a new algorithm for computing the ideal extension.

We apply generative adversarial network (GAN) technology to build an event generator that simulates particle production in electron-proton scattering that is free of theoretical assumptions about underlying particle dynamics. The difficulty of efficiently training a GAN event simulator lies in learning the complicated patterns of the distributions of the particles physical properties. We develop a GAN that selects a set of transformed features from particle momenta that can be generated easily by the generator, and uses these to produce a set of augmented features that improve the sensitivity of the discriminator. The new Feature-Augmented and Transformed GAN (FAT-GAN) is able to faithfully reproduce the distribution of final state electron momenta in inclusive electron scattering, without the need for input derived from domain-based theoretical assumptions. The developed technology can play a significant role in boosting the science of existing and future accelerator facilities, such as the Electron-Ion Collider.

Given a knowledge base and an observation as a set of facts, ABox abduction aims at computing a hypothesis that, when added to the knowledge base, is sufficient to entail the observation. In signature-based ABox abduction, the hypothesis is further required to use only names from a given set. This form of abduction has applications such as diagnosis, KB repair, or explaning missing entailments. It is possible that hypotheses for a given observation only exist if we admit the use of fresh individuals and/or complex concepts built from the given signature, something most approaches for ABox abduction so far do not allow or only allow with restrictions. In this paper, we investigate the computational complexity of this form of abduction---allowing either fresh individuals, complex concepts, or both---for various description logics, and give size bounds on the hypotheses if they exist.

Recently, most siamese network based trackers locate targets via object classification and bounding-box regression. Generally, they select the bounding-box with maximum classification confidence as the final prediction. This strategy may miss the right result due to the accuracy misalignment between classification and regression. In this paper, we propose a novel siamese tracking algorithm called SiamRCR, addressing this problem with a simple, light and effective solution. It builds reciprocal links between classification and regression branches, which can dynamically re-weight their losses for each positive sample. In addition, we add a localization branch to predict the localization accuracy, so that it can work as the replacement of the regression assistance link during inference. This branch makes the training and inference more consistent. Extensive experimental results demonstrate the effectiveness of SiamRCR and its superiority over the state-of-the-art competitors on GOT-10k, LaSOT, TrackingNet, OTB-2015, VOT-2018 and VOT-2019. Moreover, our SiamRCR runs at 65 FPS, far above the real-time requirement.

We introduce an end-to-end model for participatory budgeting grounded in social choice theory. Our model accounts for the interplay between the two stages commonly encountered in real-life partici- patory budgeting. In the first stage participants pro- pose projects to be shortlisted, while in the second stage they vote on which of the shortlisted projects should be funded. Prior work of a formal nature has focused on analysing the second stage only. We in- troduce several shortlisting rules for the first stage and analyse them in both normative and algorith- mic terms. Our main focus is on the incentives of participants to engage in strategic behaviour during the first stage, in which they need to reason about how their proposals will impact the range of strate- gies available to everyone in the second stage.

3D human pose estimation is a fundamental problem in artificial intelligence, and it has wide applications in AR/VR, HCI and robotics. However, human pose estimation from point clouds still suffers from noisy points and estimated jittery artifacts because of handcrafted-based point cloud sampling and single-frame-based estimation strategies. In this paper, we present a new perspective on the 3D human pose estimation method from point cloud sequences. To sample effective point clouds from input, we design a differentiable point cloud sampling method built on density-guided attention mechanism. To avoid the jitter caused by previous 3D human pose estimation problems, we adopt temporal information to obtain more stable results. Experiments on the ITOP dataset and the NTU-RGBD dataset demonstrate that all of our contributed components are effective, and our method can achieve state-of-the-art performance.

If domain knowledge can be integrated as an appropriate constraint, it is highly possible that the generalization performance of a neural network model can be improved. We propose Sensitivity Direction Learning (SDL) for learning about the neural network model with user-specified relationships (e.g., monotonicity, convexity) between each input feature and the output of the model by imposing soft shape constraints which represent domain knowledge. To impose soft shape constraints, SDL uses a novel penalty function, Sensitivity Direction Error (SDE) function, which returns the squared error between coefficients of the approximation curve for each Individual Conditional Expectation plot and coefficient constraints which represent domain knowledge. The effectiveness of our concept was verified by simple experiments. Similar to those such as L2 regularization and dropout, SDL and SDE can be used without changing neural network architecture. We believe our algorithm can be a strong candidate for neural network users who want to incorporate domain knowledge.

We aim for mobile robots to function in a variety of common human environments, which requires them to efficiently search previously unseen target objects. We can exploit background knowledge about common spatial relations between landmark objects and target objects to narrow down search space. In this paper, we propose an active visual object search strategy method through our introduction of the Semantic Linking Maps (SLiM) model. SLiM simultaneously maintains the belief over a target object's location as well as landmark objects' locations, while accounting for probabilistic inter-object spatial relations. Based on SLiM, we describe a hybrid search strategy that selects the next best view pose for searching for the target object based on the maintained belief. We demonstrate the efficiency of our SLiM-based search strategy through comparative experiments in simulated environments. We further demonstrate the real-world applicability of SLiM-based search in scenarios with a Fetch mobile manipulation robot.

Few-shot classification aims to recognize new classes by learning reliable models from very few available samples. It could be very challenging when there is no intersection between the alreadyknown classes (base set) and the novel set (new classes). To alleviate this problem, we propose to evolve the network (for the base set) via label propagation and self-supervision to shrink the distribution difference between the base set and the novel set. Our network evolution approach transfers the latent distribution from the already-known classes to the unknown (novel) classes by: (a) label propagation of the novel/new classes (novel set); and (b) design of dual-task to exploit a discriminative representation to effectively diminish the overfitting on the base set and enhance the generalization ability on the novel set. We conduct comprehensive experiments to examine our network evolution approach against numerous state-of-the-art ones, especially in a higher way setup and cross-dataset scenarios. Notably, our approach outperforms the second best state-of-the-art method by a large margin of 3.25% for one-shot evaluation over miniImageNet.

Knowledge distillation is a simple but effective method for model compression, which obtains a better-performing small network (Student) by learning from a well-trained large network (Teacher). However, when the difference in the model sizes of Student and Teacher is large, the gap in capacity leads to poor performance of Student. Existing methods focus on seeking simplified or more effective knowledge from Teacher to narrow the Teacher-Student gap, while we address this problem by Student's self-boosting. Specifically, we propose a novel distillation method named Self-boosting Feature Distillation (SFD), which eases the Teacher-Student gap by feature integration and self-distillation of Student. Three different modules are designed for feature integration to enhance the discriminability of Student's feature, which leads to improving the order of convergence in theory. Moreover, an easy-to-operate self-distillation strategy is put forward to stabilize the training process and promote the performance of Student, without additional forward propagation or memory consumption. Extensive experiments on multiple benchmarks and networks show that our method is significantly superior to existing methods.

This paper proposes a novel pretext task for self-supervised video representation learning by exploiting spatiotemporal continuity in videos. It is motivated by the fact that videos are spatiotemporal by nature and a representation learned by detecting spatiotemporal continuity/discontinuity is thus beneficial for downstream video content analysis tasks. A natural choice of such a pretext task is to construct spatiotemporal (3D) jigsaw puzzles and learn to solve them. However, as we demonstrate in the experiments, this task turns out to be intractable. We thus propose Constrained Spatiotemporal Jigsaw (CSJ) whereby the 3D jigsaws are formed in a constrained manner to ensure that large continuous spatiotemporal cuboids exist. This provides sufficient cues for the model to reason about the continuity. Instead of solving them directly, which could still be extremely hard, we carefully design four surrogate tasks that are more solvable. The four tasks aim to learn representations sensitive to spatiotemporal continuity at both the local and global levels. Extensive experiments show that our CSJ achieves state-of-the-art on various benchmarks.

Weakly-supervised temporal action localization aims to locate intervals of action instances with only video-level action labels for training. However, the localization results generated from video classification networks are often not accurate due to the lack of temporal boundary annotation of actions. Our motivating insight is that the temporal boundary of action should be stably predicted under various temporal transforms. This inspires a self-supervised equivariant transform consistency constraint. We design a set of temporal transform operations, including naive temporal down-sampling to learnable attention-piloted time warping. In our model, a localization network aims to perform well under all transforms, and another policy network is designed to choose a temporal transform at each iteration that adversarially brings localization result inconsistent with the localization network's. Additionally, we devise a self-refine module to enhance the completeness of action intervals harnessing temporal and semantic contexts. Experimental results on THUMOS14 and ActivityNet demonstrate that our model consistently outperforms the state-of-the-art weakly-supervised temporal action localization methods.

Medication recommendation is a significant healthcare application due to its promise in effectively prescribing medications. Avoiding fatal side effects related to Drug-Drug Interaction (DDI) is among the critical challenges. Most existing methods try to mitigate the problem by providing models with extra DDI knowledge, making models complicated. While treating all patients with different DDI properties as a single cohort would put forward strict requirements on models' generalization performance. In pursuit of a valuable model for a safe recommendation, we propose the Self-Supervised Adversarial Regularization Model for Medication Recommendation (SARMR). SARMR obtains the target distribution associated with safe medication combinations from raw patient records for adversarial regularization. In this way, the model can shape distributions of patient representations to achieve DDI reduction. To obtain accurate self-supervision information, SARMR models interactions between physicians and patients by building a key-value memory neural network and carrying out multi-hop reading to obtain contextual information for patient representations. SARMR outperforms all baseline methods in the experiment on a real-world clinical dataset. This model can achieve DDI reduction when considering the different number of DDI types, which demonstrates the robustness of adversarial regularization for safe medication recommendation.

The vast majority of community detection algorithms assume that the networks are totally observed. However, in reality many networks cannot be fully observed. On such network is edges-missing network, where some relationships (edges) between two entities are missing. Recently, several works have been proposed to solve this problem by combining link prediction and community detection in a two-stage method or in a unified framework. However, the goal of link prediction, which is to predict as many correct edges as possible, is not consistent with the requirement for predicting the important edges for discovering community structure on edges-missing networks. Thus, combining link prediction and community detection cannot work very well in terms of detecting community structure for edges-missing network. In this paper, we propose a community self-guided generative model which jointly completes the edges-missing network and identifies communities. In our new model, completing missing edges and identifying communities are not isolated but closely intertwined. Furthermore, we developed an effective model inference method that combines a nested Expectation-Maximization (EM) algorithm and Metropolis-Hastings Sampling. Extensive experiments on real-world edges-missing networks show that our model can effectively detect community structures while completing missing edges.