IJCAI 2020 Papers

In this paper, we apply self-attention (SA) mechanism to boost the performance of deep metric learning. However, due to the pairwise similarity measurement, the cost of storing and manipulating the complete attention maps makes it infeasible for large inputs. To solve this problem, we propose a compressed self-attention with low-rank approximation (CSALR) module, which significantly reduces the computation and memory costs without sacrificing the accuracy. In CSALR, the original attention map is decomposed into a landmark attention map and a combination coefficient map with a small number of landmark feature vectors sampled from the input feature map by average pooling. Thanks to the efficiency of CSALR, we can apply CSALR to high-resolution shallow convolutional layers and implement a multi-head form of CSALR, which further boosts the performance. We evaluate the proposed CSALR on person reidentification which is a typical metric learning task. Extensive experiments shows the effectiveness and efficiency of CSALR in deep metric learning and its superiority over the baselines.

Despite a multitude of empirical studies, little consensus exists on whether neural networks are able to generalise compositionally. As a response to this controversy, we present a set of tests that provide a bridge between, on the one hand, the vast amount of linguistic and philosophical theory about compositionality of language and, on the other, the successful neural models of language. We collect different interpretations of compositionality and translate them into five theoretically grounded tests for models that are formulated on a task-independent level. To demonstrate the usefulness of this evaluation paradigm, we instantiate these five tests on a highly compositional data set which we dub PCFG SET, apply the resulting tests to three popular sequence-to-sequence models and provide an in-depth analysis of the results.

Predicate Invention in Meta-Interpretive Learning (MIL) is generally based on a top-down approach, and the search for a consistent hypothesis is carried out starting from the positive examples as goals. We consider augmenting top-down MIL systems with a bottom-up step during which the background knowledge is generalised with an extension of the immediate consequence operator for second-order logic programs. This new method provides a way to perform extensive predicate invention useful for feature discovery. We demonstrate this method is complete with respect to a fragment of dyadic datalog. We theoretically prove this method reduces the number of clauses to be learned for the top-down learner, which in turn can reduce the sample complexity. We formalise an equivalence relation for predicates which is used to eliminate redundant predicates. Our experimental results suggest pairing the state-of-the-art MIL system Metagol with an initial bottom-up step can significantly improve learning performance.

We study a competition among two contests, where each contest designer aims to attract as much effort as possible. Such a competition exists in reality, e.g., in crowd-sourcing websites. Our results are phrased in terms of the ``relative prize power'' of a contest, which is the ratio of the total prize offered by this contest designer relative to the sum of total prizes of the two contests. When contestants have a quasi-linear utility function that captures both a risk-aversion effect and a cost of effort, we show that a simple contest attracts a total effort which approaches the relative prize power of the contest designer assuming a large number of contestants. This holds regardless of the contest policy of the opponent, hence providing a ``safety level'' which is a robust notion similar in spirit to the max-min solution concept.

Community detection, aiming at partitioning a network into multiple substructures, is practically importance. Graph convolutional network (GCN), a new deep-learning technique, has recently been developed for community detection. Markov Random Fields (MRF) has been combined with GCN in the MRFasGCN method to improve accuracy. However, the existing GCN community-finding methods are semi-supervised, even though community finding is essentially an unsupervised learning problem. We developed a new GCN approach for unsupervised community detection under the framework of Autoencoder. We cast MRFasGCN as an encoder and then derived node community membership in the hidden layer of the encoder. We introduced a community-centric dual decoder to reconstruct network structures and node attributes separately in an unsupervised fashion, for faithful community detection in the input space. We designed a scheme of local enhancement to accommodate nodes to have more common neighbors and similar attributes with similar community memberships. Experimental results on real networks showed that our new method outperformed the best existing methods, showing the effectiveness of the novel decoding mechanism for generating links and attributes together over the commonly used methods for reconstructing links alone.

Constructing proper representations of molecules lies at the core of numerous tasks such as molecular property prediction and drug design. Graph neural networks, especially message passing neural network (MPNN) and its variants, have recently made remarkable achievements in molecular graph modeling. Albeit powerful, the one-sided focuses on atom (node) or bond (edge) information of existing MPNN methods lead to the insufficient representations of the attributed molecular graphs. Herein, we propose a Communicative Message Passing Neural Network (CMPNN) to improve the molecular embedding by strengthening the message interactions between nodes and edges through a communicative kernel. In addition, the message generation process is enriched by introducing a new message booster module. Extensive experiments demonstrated that the proposed model obtained superior performances against state-of-the-art baselines on six chemical property datasets. Further visualization also showed better representation capacity of our model.

Our architecture uses non-monotonic logical reasoning with incomplete commonsense domain knowledge, and incremental inductive learning, to guide the construction of deep network models from a small number of training examples. Experimental results in the context of a robot reasoning about the partial occlusion of objects and the stability of object configurations in simulated images indicate an improvement in reliability and a reduction in computational effort in comparison with an architecture based just on deep networks.

To assess the trustworthiness of an agent in a multi-agent system, one often combines two types of trust information: direct trust information derived from one's own interactions with that agent, and indirect trust information based on advice from other agents. This paper provides the first systematic study on when it is beneficial to combine these two types of trust as opposed to relying on only one of them. Our large-scale experimental study shows that strong methods for computing indirect trust make direct trust redundant in a surprisingly wide variety of scenarios. Further, a new method for the combination of the two trust types is proposed that, in the remaining scenarios, outperforms the ones known from the literature.

The problem of multi-armed bandit (MAB) with fairness constraint has emerged as an important research topic recently. For such problems, one common objective is to maximize the total rewards within a fixed round of pulls, while satisfying the fairness requirement of a minimum selection fraction for each individual arm in the long run. Previous works have made substantial advancements in designing efficient online selection solutions, however, they fail to achieve a sublinear regret bound when incorporating such fairness constraints. In this paper, we study a combinatorial MAB problem with concave objective and fairness constraints. In particular, we adopt a new approach that combines online convex optimization with bandit methods to design selection algorithms. Our algorithm is computationally efficient, and more importantly, manages to achieve a sublinear regret bound with probability guarantees. Finally, we evaluate the performance of our algorithm via extensive simulations and demonstrate that it outperforms the baselines substantially.

In this paper, we show that a simple coloring scheme can improve, both theoretically and empirically, the expressive power of Message Passing Neural Networks (MPNNs). More specifically, we introduce a graph neural network called Colored Local Iterative Procedure (CLIP) that uses colors to disambiguate identical node attributes, and show that this representation is a universal approximator of continuous functions on graphs with node attributes. Our method relies on separability, a key topological characteristic that allows to extend well-chosen neural networks into universal representations. Finally, we show experimentally that CLIP is capable of capturing structural characteristics that traditional MPNNs fail to distinguish, while being state-of-the-art on benchmark graph classification datasets.

Most solution concepts in collective decision making are defined assuming complete knowledge of individuals' preferences and of the mechanism used for aggregating them. This is often unpractical or unrealistic. Under incomplete knowledge, a solution advocated by many consists in quanrtifying over all completions of the incomplete preference profile (or all instantiations of the incompletely specified mechanism). Voting rules can be `modalized' this way (leading to the notions of possible and necessary winners), and also efficiency and fairness notions in fair division, stability concepts in coalition formation, and more. I give here a survey of works along this line.

Session-based recommendation becomes a research hotspot for its ability to make recommendations for anonymous users. However, existing session-based methods have the following limitations: (1) They either lack the capability to learn complex dependencies or focus mostly on the current session without explicitly considering collaborative information. (2) They assume that the representation of an item is static and fixed for all users at each time step. We argue that even the same item can be represented differently for different users at the same time step. To this end, we propose a novel solution, Collaborative Self-Attention Network (CoSAN) for session-based recommendation, to learn the session representation and predict the intent of the current session by investigating neighborhood sessions. Specially, we first devise a collaborative item representation by aggregating the embedding of neighborhood sessions retrieved according to each item in the current session. Then, we apply self-attention to learn long-range dependencies between collaborative items and generate collaborative session representation. Finally, each session is represented by concatenating the collaborative session representation and the embedding of the current session. Extensive experiments on two real-world datasets show that CoSAN constantly outperforms state-of-the-art methods.

Scene perceiving and understanding tasks including depth estimation, visual odometry (VO) and camera relocalization are fundamental for applications such as autonomous driving, robots and drones. Driven by the power of deep learning, significant progress has been achieved on individual tasks but the rich correlations among the three tasks are largely neglected. In previous studies, VO is generally accurate in local scope yet suffers from drift in long distances. By contrast, camera relocalization performs well in the global sense but lacks local precision. We argue that these two tasks should be strategically combined to leverage the complementary advantages, and be further improved by exploiting the 3D geometric information from depth data, which is also beneficial for depth estimation in turn. Therefore, we present a collaborative learning framework, consisting of DepthNet, LocalPoseNet and GlobalPoseNet with a joint optimization loss to estimate depth, VO and camera localization unitedly. Moreover, the Geometric Attention Guidance Model is introduced to exploit the geometric relevance among three branches during learning. Extensive experiments demonstrate that the joint learning scheme is useful for all tasks and our method outperforms current state-of-the-art techniques in depth estimation and camera relocalization with highly competitive performance in VO.

Detecting fraud users, who fraudulently promote certain target items, is a challenging issue faced by e-commerce platforms. Generally, many fraud users have different spam behaviors simultaneously, e.g. spam transactions, clicks, reviews and so on. Existing solutions have two main limitations: 1) the correlations among multiple spam behaviors are neglected; 2) large-scale computations are intractable when dealing with an enormous user set. To remedy these problems, this work proposes a collaboration based multi-label propagation (CMLP) algorithm. We first introduce a general-purpose version that involves collaboration technique to exploit label correlations. Specifically, it breaks the final prediction into two parts: 1) its own prediction part; 2) the prediction of others, i.e. collaborative part. Then, to accelerate it on large-scale e-commerce data, we propose a heterogeneous graph based variant that detects communities on the user-item graph directly. Both theoretical analysis and empirical results clearly validate the effectiveness and scalability of our proposals.

Multi-lingual contextualized embeddings, such as multilingual-BERT (mBERT), have shown success in a variety of zero-shot cross-lingual tasks. However, these models are limited by having inconsistent contextualized representations of subwords across different languages. Existing work addresses this issue by bilingual projection and fine-tuning technique. We propose a data augmentation framework to generate multi-lingual code-switching data to fine-tune mBERT, which encourages model to align representations from source and multiple target languages once by mixing their context information. Compared with the existing work, our method does not rely on bilingual sentences for training, and requires only one training process for multiple target languages. Experimental results on five tasks with 19 languages show that our method leads to significantly improved performances for all the tasks compared with mBERT.

Person re-identification aims at identifying a certain pedestrian across non-overlapping camera networks. Video-based person re-identification approaches have gained significant attention recently, expanding image-based approaches by learning features from multiple frames. In this work, we propose a novel Co-Saliency Spatio-Temporal Interaction Network (CSTNet) for person re-identification in videos. It captures the common salient foreground regions among video frames and explores the spatial-temporal long-range context interdependency from such regions, towards learning discriminative pedestrian representation. Specifically, multiple co-saliency learning modules within CSTNet are designed to utilize the correlated information across video frames to extract the salient features from the task-relevant regions and suppress background interference. Moreover, multiple spatial-temporal interaction modules within CSTNet are proposed, which exploit the spatial and temporal long-range context interdependencies on such features and spatial-temporal information correlation, to enhance feature representation. Extensive experiments on two benchmarks have demonstrated the effectiveness of the proposed method.

Research in both computational social choice and preference reasoning uses tools and techniques from computer science, generally algorithms and complexity analysis, to examine topics in group decision making. This has brought tremendous progress in the last decades, creating new avenues for research and results in areas including voting and resource allocation. I argue that of equal importance to the theoretical results are impacts in research and development from the empirical part of the computer scientists toolkit: data, system building, and human interaction. I highlight work by myself and others to establish data driven, application driven research in the computational social choice and preference reasoning areas. Along the way, I highlight interesting application domains and important results from the community in driving this area to make concrete, real-world impact.

Adaptive gradient methods, which adopt historical gradient information to automatically adjust the learning rate, despite the nice property of fast convergence, have been observed to generalize worse than stochastic gradient descent (SGD) with momentum in training deep neural networks. This leaves how to close the generalization gap of adaptive gradient methods an open problem. In this work, we show that adaptive gradient methods such as Adam, Amsgrad, are sometimes "over adapted". We design a new algorithm, called Partially adaptive momentum estimation method, which unifies the Adam/Amsgrad with SGD by introducing a partial adaptive parameter $p$, to achieve the best from both worlds. We also prove the convergence rate of our proposed algorithm to a stationary point in the stochastic nonconvex optimization setting. Experiments on standard benchmarks show that our proposed algorithm can maintain fast convergence rate as Adam/Amsgrad while generalizing as well as SGD in training deep neural networks. These results would suggest practitioners pick up adaptive gradient methods once again for faster training of deep neural networks.

Deep Infomax (DIM) is an unsupervised representation learning framework by maximizing the mutual information between the inputs and the outputs of an encoder, while probabilistic constraints are imposed on the outputs. In this paper, we propose Supervised Deep InfoMax (SDIM), which introduces supervised probabilistic constraints to the encoder outputs. The supervised probabilistic constraints are equivalent to a generative classifier on high-level data representations, where class conditional log-likelihoods of samples can be evaluated. Unlike other works building generative classifiers with conditional generative models, SDIMs scale on complex datasets, and can achieve comparable performance with discriminative counterparts. With SDIM, we could perform classification with rejection. Instead of always reporting a class label, SDIM only makes predictions when test samples' largest class conditional surpass some pre-chosen thresholds, otherwise they will be deemed as out of the data distributions, and be rejected. Our experiments show that SDIM with rejection policy can effectively reject illegal inputs, including adversarial examples and out-of-distribution samples.