AAAI 2021 Papers

The paper suggests a definition of "know who" as a modality using Grove-Halpern semantics of names. It also introduces a logical system that describes the interplay between modalities "knows who", "knows", and "for all agents". The main technical result is a completeness theorem for the proposed system.

Entities, as the essential elements in relation extraction tasks, exhibit certain structure. In this work, we formulate such entity structure as distinctive dependencies between mention pairs. We then propose SSAN, which incorporates these structural dependencies within the standard self-attention mechanism and throughout the overall encoding stage. Specifically, we design two alternative transformation modules inside each self-attention building block to produce attentive biases so as to adaptively regularize its attention flow. Our experiments demonstrate the usefulness of the proposed entity structure and the effectiveness of SSAN. It significantly outperforms competitive baselines, achieving new state-of-the-art results on three popular document-level relation extraction datasets. We further provide ablation and visualization to show how the entity structure guides the model for better relation extraction. Our code is publicly available.

Question generation is a challenging task and has attracted widespread attention in recent years. Although previous studies have made great progress, there are still two main shortcomings: First, previous work did not simultaneously capture the sequence information and structure information hidden in the context, which results in poor results of the generated questions. Second, the generated questions cannot be answered by the given context. To tackle these issues, we propose an entity guided question generation model with contextual structure information and sequence information capturing. We use a Graph Convolutional Network and a Bidirectional Long Short Term Memory Network to capture the structure information and sequence information of the context, simultaneously. In addition, to improve the answerability of the generated questions, we use an entity-guided approach to obtain question type from the answer, and jointly encode the answer and question type. Both automatic and manual metrics show that our model can generate comparable questions with state-of-the-art models. Our code is available at https://github.com/VISLANG-Lab/EGSS.

One significant factor we expect the video representation learning to capture, especially in contrast with the image representation learning, is the object motion. However, we found that in the current mainstream video datasets, some action categories are highly related with the scene where the action happens, making the model tend to degrade to a solution where only the scene information is encoded. For example, a trained model may predict a video as playing football simply because it sees the field, neglecting that the subject is dancing as a cheerleader on the field. This is against our original intention towards the video representation learning and may bring scene bias on a different dataset that can not be ignored. In order to tackle this problem, we propose to decouple the scene and the motion (DSM) with two simple operations, so that the model attention towards the motion information is better paid. Specifically, we construct a positive clip and a negative clip for each video. Compared to the original video, the positive/negative is motion-untouched/broken but scene-broken/untouched by Spatial Local Disturbance and Temporal Local Disturbance. Our objective is to pull the positive closer while pushing the negative farther to the original clip in the latent space. In this way, the impact of the scene is weakened while the temporal sensitivity of the network is further enhanced. We conduct experiments on two tasks with various backbones and different pre-training datasets, and find that our method surpass the SOTA methods with a remarkable 8.1% and 8.8% improvement towards action recognition task on the UCF101 and HMDB51 datasets respectively using the same backbone.

Previous work for text summarization in scientific domain mainly focused on the content of the input document, but seldom considering its citation network. However, scientific papers are full of uncommon domain-specific terms, making it almost impossible for the model to understand its true meaning without the help of the relevant research community. In this paper, we redefine the task of scientific papers summarization by utilizing their citation graph and propose a citation graph-based summarization model CGSum which can incorporate the information of both the source paper and its references. In addition, we construct a novel scientific papers summarization dataset Semantic Scholar Network (SSN) which contains 141K research papers in different domains and 661K citation relationships. The entire dataset constitutes a large connected citation graph. Extensive experiments show that our model can achieve competitive performance when compared with the pretrained models even with a simple architecture. The results also indicates the citation graph is crucial to better understand the content of papers and generate high-quality summaries.

Aiming at convex optimization under structural constraints, this work introduces and analyzes a variant of the Frank Wolfe (FW) algorithm termed ExtraFW. The distinct feature of ExtraFW is the pair of gradients leveraged per iteration, thanks to which the decision variable is updated in a prediction-correction (PC) format. Relying on no problem dependent parameters in the step sizes, the convergence rate of ExtraFW for general convex problems is shown to be ${cal O}(frac{1}{k})$, which is optimal in the sense of matching the lower bound on the number of solved FW subproblems. However, the merit of ExtraFW is its faster rate ${cal O}big(frac{1}{k^2} big)$ on a class of machine learning problems. Compared with other parameter-free FW variants that have faster rates on the same problems, ExtraFW has improved rates and fine-grained analysis thanks to its PC update. Numerical tests on binary classification with different sparsity-promoting constraints demonstrate that the empirical performance of ExtraFW is significantly better than FW, and even faster than Nesterov's accelerated gradient on certain datasets. For matrix completion, ExtraFW enjoys smaller optimality gap, and lower rank than FW.

Graph partition is a key component to achieve workload balance and reduce job completion time in parallel graph processing systems. Among the various partition strategies, edge partition has demonstrated more promising performance in power-law graphs than vertex partition and thereby has been more widely adopted as the default partition strategy by existing graph systems. The graph edge partition problem, which is to split the edge set into multiple balanced parts with the objective of minimizing the total number of copied vertices, has been widely studied from the view of optimization and algorithms. In this paper, we study local search algorithms for this problem to further improve the partition results from existing methods. More specifically, we propose two novel concepts, namely adjustable edges and blocks. Based on these, we develop a greedy heuristic as well as an improved search algorithm utilizing the property of max-flow model. To evaluate the performance of our algorithms, we first provide adequate theoretical analysis in terms of approximation quality. We significantly improve the previous known approximation ratio for this problem. Then we conduct extensive experiments on a large number of benchmark datasets and state-of-the-art edge partition strategies. The results show that our proposed local search framework can further improve the quality of graph partition by a wide margin.

The recent success of audio-visual representations learning can be largely attributed to their pervasive concurrency property, which can be used as a self-supervision signal and extract correlation information. While most recent works focus on capturing the shared associations between the audio and visual modalities, they rarely consider multiple audio and video pairs at once and pay little attention to exploiting the valuable information within each modality. To tackle this problem, we propose a novel audio-visual representation learning method dubbed self-supervised curriculum learning (SSCL) under the teacher-student learning manner. Specifically, taking advantage of contrastive learning, a two-stage scheme is exploited, which transfers the cross-modal information between teacher and student model as a phased process. The proposed SSCL approach regards the pervasive property of audiovisual concurrency as latent supervision and mutually distills the structure knowledge of visual to audio data. Notably, the SSCL method can learn discriminative audio and visual representations for various downstream applications. Extensive experiments conducted on both action video recognition and audio sound recognition tasks show the remarkably improved performance of the SSCL method compared with the state-of-the-art self-supervised audio-visual representation learning methods.

Deep neural networks are the default choice of learning models for computer vision tasks. Extensive work has been carried out in recent years on explaining deep models for vision tasks such as classification. However, recent work has shown that it is possible for these models to produce substantially different attribution maps even when two very similar images are given to the network, raising serious questions about trustworthiness. To address this issue, we propose a robust attribution training strategy to improve attributional robustness of deep neural networks. Our method carefully analyzes the requirements for attributional robustness and introduces two new regularizers that preserve a model's attribution map during attacks. Our method surpasses state-of-the-art attributional robustness methods by a margin of approximately 3% to 9% in terms of attribution robustness measures on several datasets including MNIST, FMNIST, Flower and GTSRB.

Recognizing ongoing events based on acoustic clues has been a critical yet challenging problem that has attracted significant research attention in recent years. Joint audio-visual analysis can improve the event detection accuracy but may not always be feasible as under many circumstances only audio recordings are available in real-world scenarios. To solve the challenges, we present a novel visual-assisted teacher-student mutual learning framework for robust sound event detection from audio recordings. Our model adopts a multi-modal teacher network based on both acoustic and visual clues, and a single-modal student network based on acoustic clues only. Conventional teacher-student learning performs unsatisfactorily for knowledge transfer from a multi-modality network to a single-modality network. We thus present a mutual learning framework by introducing a single-modal transfer loss and a cross-modal transfer loss to collaboratively learn the audio-visual correlations between the two networks. Our proposed solution takes the advantages of joint audio-visual analysis in training while maximizing the feasibility of the model in use cases. Our extensive experiments on the DCASE17 and the DCASE18 sound event detection datasets show that our proposed method outperforms the state-of-the-art audio tagging approaches.

Detection of redundant operators that can be safely removed from the planning task is an essential technique allowing to greatly improve performance of planners. In this paper, we employ structure-preserving maps on labeled transition systems (LTSs), namely endomorphisms well known from model theory, in order to detect redundancy. Computing endomorphisms of an LTS induced by a planning task is typically infeasible, so we show how to compute some of them on concise representations of planning tasks such as finite domain representations and factored LTSs. We formulate the computation of endomorphisms as a constraint satisfaction problem (CSP) that can be solved by an off-the-shelf CSP solver. Finally, we experimentally verify that the proposed method can find a sizeable number of redundant operators on the standard benchmark set.

End-to-end semantic role labeling (SRL) has been received increasing interest. It performs the two subtasks of SRL: predicate identification and argument role labeling, jointly. Recent work is mostly focused on graph-based neural models, while the transition-based framework with neural networks which has been widely used in a number of closely-related tasks, has not been studied for the joint task yet. In this paper, we present the first work of transition-based neural models for end-to-end SRL. Our transition model incrementally discovers all sentential predicates as well as their arguments by a set of transition actions. The actions of the two subtasks are executed mutually for full interactions. Besides, we suggest high-order compositions to extract non-local features, which can enhance the proposed transition model further. Experimental results on CoNLL09 and Universal Proposition Bank show that our final model can produce state-of-the-art performance, and meanwhile keeps highly efficient in decoding. We also conduct detailed experimental analysis for a deep understanding of our proposed model.

Recently, high dynamic range (HDR) image reconstruction based on the multiple exposure stack from a given single exposure utilizes a deep learning framework to generate high-quality HDR images. These conventional networks focus on the exposure transfer task to reconstruct the multi-exposure stack. Therefore, they often fail to fuse the multi-exposure stack into a perceptually pleasant HDR image as the inversion artifacts occur. We tackle the problem in stack reconstruction-based methods by proposing a novel framework with a fully differentiable high dynamic range imaging (HDRI) process. By explicitly using the loss, which compares the network's output with the ground truth HDR image, our framework enables a neural network that generates the multiple exposure stack for HDRI to train stably. In other words, our differentiable HDR synthesis layer helps the deep neural network to train to create multi-exposure stacks while reflecting the precise correlations between multi-exposure images in the HDRI process. In addition, our network uses the image decomposition and the recursive process to facilitate the exposure transfer task and to adaptively respond to recursion frequency. The experimental results show that the proposed network outperforms the state-of-the-art quantitative and qualitative results in terms of both the exposure transfer tasks and the whole HDRI process.

We propose a novel Transformer encoder-based architecture with syntactical knowledge encoded for intent detection and slot filling. Specifically, we encode syntactic knowledge into the Transformer encoder by jointly training it to predict syntactic parse ancestors and part-of-speech of each token via multi-task learning. Our model is based on self-attention and feed-forward layers and does not require external syntactic information to be available at inference time. Experiments show that on two benchmark datasets, our models with only two Transformer encoder layers achieve state-of-the-art results. Compared to the previously best performed model without pre-training, our models achieve absolute F1 score and accuracy improvement of 1.59 % and 0.85 % for slot filling and intent detection on the SNIPS dataset, respectively. Our models also achieve absolute F1 score and accuracy improvement of 0.1 % and 0.34 % for slot filling and intent detection on the ATIS dataset, respectively, over the previously best performed model. Furthermore, the visualization of the self-attention weights illustrates the benefits of incorporating syntactic information during training.

Deep reinforcement learning has been successful in a variety of tasks, such as game playing and robotic manipulation. However, attempting to learn tabula rasa disregards the logical structure of many domains as well as the wealth of readily available knowledge from domain experts that could help "warm start" the learning process. We present a novel reinforcement learning technique that allows for intelligent initialization of a neural network weights and architecture. Our approach permits the encoding domain knowledge directly into a neural decision tree, and improves upon that knowledge with policy gradient updates. We empirically validate our approach on two OpenAI Gym tasks and two modified StarCraft 2 tasks, showing that our novel architecture outperforms multilayer-perceptron and recurrent architectures. Our knowledge-based framework finds superior policies compared to imitation learning-based and prior knowledge-based approaches. Importantly, we demonstrate that our approach can be used by untrained humans to initially provide >80% increase in expected reward relative to baselines prior to training (p < 0.001), which results in a >60% increase in expected reward after policy optimization (p = 0.011).

Currently the unified semantic role labeling (SRL) that achieves predicate identification and argument role labeling in an end-to-end manner has received growing interests. Recent works show that leveraging the syntax knowledge significantly enhances the SRL performances. In this paper, we investigate a novel unified SRL framework based on the sequence-to-sequence architecture with double enhancement in both the encoder and decoder sides. In the encoder side, we propose a novel label-aware graph convolutional network (LA-GCN) to encode both the syntactic dependent arcs and labels into BERT-based word representations. In the decoder side, we creatively design a pointer-network-based model for detecting predicates, arguments and roles jointly. Our pointer-net decoder is able to make decisions by consulting all the input elements in a global view, and meanwhile it is syntactic-aware by incorporating the syntax information from LA-GCN. Besides, a high-order interacted attention is introduced into the decoder for leveraging previously recognized triplets to help the current decision. Empirical experiments show that our framework significantly outperforms all existing graph-based methods on the CoNLL09 and Universal Proposition Bank datasets. In-depth analysis demonstrates that our model can effectively capture the correlations between syntactic and SRL structures.

Recently, the vulnerability of deep neural network (DNN)-based audio systems to adversarial attacks has obtained increasing attention. However, the existing audio adversarial attacks allow the adversary to possess the entire user's audio input as well as granting sufficient time budget to generate the adversarial perturbations. These idealized assumptions, however, make the existing audio adversarial attacks mostly impossible to be launched in a timely fashion in practice (e.g., playing unnoticeable adversarial perturbations along with user's streaming input). To overcome these limitations, in this paper we propose fast audio adversarial perturbation generator (FAPG), which uses generative model to generate adversarial perturbations for the audio input in a single forward pass, thereby drastically improving the perturbation generation speed. Built on the top of FAPG, we further propose universal audio adversarial perturbation generator (UAPG), a scheme to craft universal adversarial perturbation that can be imposed on arbitrary benign audio input to cause misclassification. Extensive experiments on DNN-based audio systems show that our proposed FAPG can achieve high success rate with up to 214X speedup over the existing audio adversarial attack methods. Also our proposed UAPG generates universal adversarial perturbations that can achieve much better attack performance than the state-of-the-art solutions.

Much research effort in HRI has focused on how to enable robots to learn new skills from observations, demonstrations, and instructions. Less work, however, has focused on how skills can be corrected if they were learned incorrectly, adapted to changing circumstances, or generalized/specialized to different contexts. In this paper, a skill modification framework is introduced that allows users to modify a robot’s stored skills quickly through instructions to (1) reduce inefficiencies, (2) fix errors, and (3) enable generalizations, all in a way for modified skills to be immediately available for task performance. A thorough evaluation of the implemented framework shows the operation of the algorithms integrated in a cognitive robotic architecture on different fully autonomous robots in various HRI case studies. An additional online HRI user study verifies that subjects prefer to quickly modify robot knowledge in the way we proposed in the framework.

With the capability of trading accuracy for latency on-the-fly, the technique of adaptive early-exit inference has emerged as a promising line of research to accelerate the deep learning inference. However, studies in this line of research commonly use a group of thresholds to control the accuracy-latency trade-off, where a thorough and general methodology on how to determine these thresholds has not been conducted yet, especially with regard to the common requirements of average inference latency. To address this issue and enable latency-aware adaptive early-exit inference, in the present paper, we approximately formulate the threshold determination problem of finding the accuracy-maximum threshold setting that meets a given average latency requirement, and then propose a threshold determination method to tackle our formulated non-convex problem. Theoretically, we prove that, for certain parameter settings, our method finds an approximate stationary point of the formulated problem. Empirically, on top of various models across multiple datasets (CIFAR-10, CIFAR-100, ImageNet and two time-series datasets), we show that our method can well handle the average latency requirements, and consistently finds good threshold settings in negligible time.