Search

Adversarial examples about deep learning models have been paid much attention in recent years, including single-label adversarial examples and multi-label adversarial examples. In this paper, for the first time, an empirical study of generating a multi-label adversarial example to hide all labels in a multi-label example is presented. The objective of hiding all labels in a multi-label example is to make deep learning models know nothing about the environments. That is very worthy of studying because deep learning models will say there is nothing, although the real input has more than one label. In the empirical study, we use five state-of-the-art multi-label attack algorithms, i.e., ML-CW, ML-DP, FGSM, MI-FGSM, MLA-LP, four popular datasets, i.e., VOC2007, VOC2012, NUS-WIDE and COCO, and two typical models ML-GCN and ASL for evaluation. We conduct extensive experiments and report the attack success rates, the amount of perturbations of the adversarial examples generated by state-of-the-art multi-label attack algorithms. We also report the attack performance when a typical defending algorithm based on JPEG compression is used. The work in this paper is beneficial to the future study of generating multi-label adversarial examples as well as defending them.

One of the most important processes in big data analysis is outlier detection, where anomalies of observed data are detected and eliminated to improve the system performance. Many approaches to outlier detection have been proposed so far for classification and prediction purposes. This paper focus on the outlier detection under a practical circumstance such that multiple organizations possess different data sets of a specific task, while they cannot directly share with each other from a privacy point of view; that is, each organization is not allowed to provide their sensitive data to other entities but they are keen to cooperate in data analysis for some reasons. To address this issue, we present a new outlier detection approach to the data analysis for multiple organizations using a decision tree ensemble based on the so-called federated learning scheme. In this paper, we extend an existing outlier detection mode called Isolation Forest to the federated learning concept so that data protection of each entity can be achieved by introducing an additive homomorphic encryption. The experimental results for several benchmark data sets demonstrate that the proposed privacy-preserving Isolation Forest (pp-iForest) achieves stable classification performance, which is almost the same as that under a single-organization setting, even when the number of organization is increased.

This paper focuses on the word disentanglement-based approaches for text style transfer. The related systems first remove the style attribute words in the given texts and then generate the target sentences using the remained neutral templates. Previous work retrieves the style words using intuitive heuristics, which lacks in-depth analysis and can hardly provide a precise word detection. The resulting error will further affect the generation phase and leads to the failure of the style transformation. In this paper, we formalize the style detection task as a dynamic decision process; each word in the given text is classified as a style or neutral content sequentially leveraging the word property, the context, and the decision history. As no background labels indicating the attribute words, reinforcement learning is deployed in our model. An evaluation (reward) function is designed to quantify the style shift and content loss simultaneously, providing an accurate estimation for the decision process. The efficiency of the proposal is verified on two types of style transfer tasks, i.e., sentiment transfer and formality transfer. The experimental results demonstrate the model's competitive performance compared to the state-of-the-art baselines once equipped with an insertion-based generation approach.

Intrusion Detection System (IDS) is an important tool to ensure network security, which can detect and prevent malicious behavior in time. However, the noise and redundancy of data often reduce the detection performance of classifiers. The traditional model of intrusion detection system cannot effectively solve this problem. Therefore, in this paper, autoencoders (AEs) are firstly used to reduce the dimension of the original data, and a hybrid model combining particle swarm optimization (PSO) and gray wolf optimization (GWO) is proposed to optimize the support vector machine (SVM) parameters. This method combines the two optimization algorithms and selects the optimal parameter values according to the locally enhanced particles to train the classifier. In this paper, the NSL-KDD benchmark dataset and UNSW-NB15 dataset are used to evaluate the proposed model, and the model is compared with other classification methods separately. The experimental results show that our hybrid optimization model has better performance in detection accuracy and provides good detection rate and false alarm rate.

In recent years, deep learning has developed rapidly and has shown great performance on many challenging machine learning tasks, such as image classification, natural language processing, and speech recognition. However, researchers have recently discovered that deep learning models have security risks and are easily affected by adversarial examples. The adversarial example is a sample formed by deliberately adding subtle perturbation that is invisible to the human in the dataset. It can make the classification classify incorrectly with a high degree of confidence, which poses more challenges for deep learning research. In this paper, we propose a defense model based on the supervision mechanism. The model adds supervision layers to the original convolutional neural network and improves the robustness and defense ability of the model by improving the loss function. The LeNet-5 and VGG networks are used as the original network models. The experimental results on MNIST and CIFAR-10 confirm that the method proposed in this paper will effectively increase the difficulty of the attackers.

Scientific creativity refers to the development of new and innovative ideas that foster technological growth and advancement. This paper attempts to identify creative individuals in scientific domain by evaluating their visuo-spatial reasoning ability using EEG analysis system. The main objective of the present work is achieved by first procuring EEG signals from the scalp of subjects involved in a mental paper folding task. The acquired signals are pre-processed and transformed into a topological map using correlation analysis to investigate the connectivity patterns amongst different brain lobes pertaining to different degrees of spatial abilities (high and low). The acquired brain networks depict strong interconnections amongst bilateral supramarginal gyrus, right anterior prefrontal and right superior temporal lobes for subject possessing high spatial skills. Since, the correlation based brain network analysis provides a qualitative illustration of the connectivity patterns, centrality induced quantitative analysis have been performed. Such an evaluation has been conducted using two categories of centrality measures viz. degree centrality and betweenness centrality. The results obtained from centrality analysis also infer the active involvement of the aforesaid brain regions during mental paper folding activity. Finally, the centrality based network features are transferred to a novel Capsule Graph Neural Network (CapsGNN) based classifier to categorize the desired class labels. Performance analysis undertaken confirms the superior functionality of the proposed network with respect to the conventional models. Moreover, statistical evaluation conducted also assures the supercilious performance of the proposed classifier. Thus, the present methodology can be utilized for recruiting creative individuals to different sectors of industries associated with innovation such as architecture, game designing and the like.

This paper considers the unsupervised domain adaptation problem, in which we want to find a good prediction function on the unlabeled target domain, by utilizing the information provided in the labeled source domain. A common approach to the domain adaptation problem is to learn a representation space where the distributional discrepancy of the source and target domains is small. Existing methods generally tend to match the marginal distributions of the two domains, while the label information in the source domain is not fully exploited. In this paper, we propose a representation learning approach for domain adaptation, which is addressed as JODAWAT. We aim to adapt the joint distributions of the feature-label pairs in the shared representation space for both domains. In particular, we minimize the Wasserstein distance between the source and target domains, while the prediction performance on the source domain is also guaranteed. The proposed approach results in a minimax adversarial training procedure that incorporates a novel split gradient penalty term. A generalization bound on the target domain is provided to reveal the efficacy of representation learning for joint distribution adaptation. We conduct extensive evaluations on JODAWAT, and test its classification accuracy on multiple synthetic and real datasets. The experimental results justify that our proposed method is able to achieve superior performance compared with various domain adaptation methods.

Aspect-based sentiment analysis analyses the fine-grained sentiment polarity of a particular attribute in a sentence. In addition to the fact that most papers have applied attention mechanisms and neural networks to fine-grained sentiment analysis, some studies have reported outstanding performance from graph-structured neural networks in aspect-based sentiment analysis. In this paper, we propose the use of a graph attention network (GAT) to embed external knowledge and grammatical relations (KD-GAT). First, we employ a GAT to extract the nodes and edges related to sentences in the knowledge graph. Second, we consider the influence of conjunctions and parts of speech based on the aspect-oriented syntactic dependency tree. Finally, we emphasize the effect of aspect position in the mechanism of multiple attention. Experiments on the SemEval 2014 and Twitter datasets show that our model can improve the ability to analyze sentiment and that the graph structure method can better integrate grammatical relations to understand a sentence.

With the recent surge in threats to public safety, the security focus of several organizations has been moved towards enhanced intelligent screening systems. Conventional X-ray screening, which relies on the human operator is the best use of this technology, allowing for the more accurate identification of potential threats. This paper explores X-ray security imagery by introducing a novel approach that generates realistic synthesized data, which opens up the possibility of using different settings to simulate occlusion, radiopacity, varying textures, and distractors to generate cluttered scenes. The generated synthetic data is effective in the training of deep networks. It allows better generalization on training data to deal with domain adaptation in the real world. The extensive set of experiments in this paper provides evidence for the efficacy of synthetic datasets over human-annotated datasets for automated X-ray security screening. The proposed approach outperforms the state-of-the-art approach for a diverse threat object dataset on mean Average Precision (mAP) of region-based detectors and classification/regression-based detectors.

Multilingual chat systems are the need of the hour for organizations who render online conversational services to their customers. This can effectively be facilitated using the robust machine translation (MT) systems. Translating user-generated contents is regarded as one of the challenging tasks for MT. As for translating conversations, it is more challenging since the meaning of any particular utterance in a conversation usually depends on its context. Moreover, chats in conversational systems are usually informal, contain code-mixed (mix of more than one language), and other grammatical inconsistencies. In this paper, we use state-of-the-art Transformer models to build our MT systems and to translate conversations between the customer service agents and customers. We propose a novel method which effectively selects contextual information from the source text of conversation to be translated. We also employ a terminology-based pseudo in-domain corpus mining strategy for fine-tuning our translation model. We evaluate our methods on the German-English WMT20 Shared Task on Chat Translation dataset, and obtain 61.1 and 63.9 BLEU points on evaluation test sets for English-to-German and German-to-English, respectively, surpassing the present state-of-the-art MT systems by 0.7 BLEU and 1.5 BLEU points, respectively. In this paper, we use state-of-the-art Transformer models to build our MT systems and to translate conversations between the customer service agents and customers. We propose a novel method which effectively selects contextual information from the source text of conversation to be translated. We also employ a terminology-based pseudo in-domain corpus mining strategy for fine-tuning our translation model. We evaluate our methods on the German-English WMT20 Shared Task on Chat Translation dataset, and obtain 61.1 and 63.9 BLEU points on evaluation test sets for English-to-German and German-to-English, respectively, surpassing the present state-of-the-art MT systems by 0.7 BLEU and 1.5 BLEU points, respectively.

Transfer learning has been widely studied and gained increasing popularity to improve the accuracy of machine learning models by transferring some knowledge acquired in different training. However, no prior work has pointed out that transfer learning can strengthen privacy attacks on machine learning models. In this paper, we propose TransMIA (Transfer learning-based Membership Inference Attacks), which use transfer learning to perform membership inference attacks on the source model when the adversary is able to access the parameters of the transferred model. In particular, we propose a transfer shadow training technique, where an adversary employs the parameters of the transferred model to construct shadow models, to significantly improve the performance of membership inference when a limited amount of shadow training data is available to the adversary. We evaluate our attacks using two real datasets, and show that our attacks outperform the state-of-the-art that does not use our transfer shadow training technique. We also compare four combinations of the learning-based/entropy-based approach and the fine-tuning/freezing approach, all of which employ our transfer shadow training technique. Then we examine the performance of these four approaches based on the distributions of confidence values, and discuss possible countermeasures against our attacks., Comment: IJCNN 2021 conference paper

By using document-level contextual information, document-level neural machine translation can achieve better results than ordinary machine translation, but traditional document-level machine translation is difficult to focus on the contextual sentence articulation relations and deep positional relations within the discourse while utilizing document-level vocabulary, and the model can concentrate only on relatively shallow inter-sentential relations or positional information. In this paper, we consider that most adjacent sentences are connected in document translation, and such links help improve the quality of translation. We propose a document translation model that focuses more on inter-sentential relations based on the previous work, and propose two methods to strengthen the model's positional information input, and combine these two methods to enhance the traditional Transformer positional information input. This paper also proposes a method for inserting paragraph information to allow inter-sentential relations to be learned by the model, and uses the improved Transformer model for Chinese-Mongolian document translation. Experiments show that in the improved Transformer system, the BLEU scores are enhanced on the Chinese-Mongolian machine translation task after fusing positional information and inter-sentential relation information, and the translation achieves better performance.

In recent years, Convolutional Neural Networks (CNN s) have been widely applied in various applications due to its powerful learning capability. However, its lack of explainability hinders its further usage in tasks requiring high reliability. Therefore, interpretability technique is the key to the application and deployment of CNN models. As a typical interpretability technique for CNN, Class Activation Map (CAM) utilizing the gradient based weights and activation map is widely applied to traditional CNN models for offering visual interpretability. However, the activation map adopted by CAM cannot loyally quantify the relevance between input samples and activation values. Hence, in this paper, we propose a new interpretability approach called Salience-CAM employing salience scores to accurately measure the relevance between input samples and activation values. To evaluate the effectiveness of Salience-CAM, comprehensive experiments are conducted on 6 selected time series datasets. By leveraging an evaluation algorithm proposed in this paper, the experimental results show that our proposed Salience-CAM outperforms the baseline by discovering more discriminative features.

This paper concerns the cooperative spectrum sensing (CSS) for cognitive radio (CR) networks, where the secondary users (SUs) collaborate to detect the presence of the primary users (PUs). With CSS, the information from different SUs is first fused, then, the detection of the PU signal is implemented based on the fused information. Most of the previous works focus on the design of a mapping function that calculates the probability of the existence of the PU signal based on the fused information. In this paper, we study the fusion process which combines the information from all SUs based on the local property of each SU. A graph-based multi-view learning framework for CSS (GMCSS) is designed to better fuse the information from different SUs. In the proposed framework, the information from each SU is considered as a view of the state of the target wireless channel and is fused with the information from other SUs through a graph-based learning process. Simulation results demonstrate the effectiveness of our method.

Script event prediction is a significant task to extract event evolution rules from text. It requires to capture the intra-rationality and inter-relationship of events. The previous methods have made great progress, but they can not fully capture the inner features of events. This paper proposes an event representation layer with argument-level attention to capture the relationship between predicate and arguments, which benefit to assess the rationality of the event. Besides, we integrate the enhanced event representation with the multi-level event similarity scores to obtain more robust performance. The multi-level similarity consists of scene similarity and semantic consistency. We utilize the similarity between candidates and individual events to evaluate the scene similarity. Semantic consistency is evaluated by the hidden state of Gate Recurrent Unit (GRU). Unlike the previous work, we only use the last hidden state of GRU instead of all hidden states to reduce training time. In addition, we adopt adversarial training to improve the robustness and generalization of the model. The experimental results verify the effectiveness of our model compared to these baseline methods. This paper also demonstrates what the attention mechanism has learned to enhance the interpretability of our model.

This paper uses a neural architecture search (NAS) algorithm based on Bayesian optimization (BO), a version of neural architecture search adapted for network morphism, is proposed to automatically seek an optimized cell-based architecture approximation for CNN to classify the coal and gangue. Besides, we defined the cell-based search space, which has significantly reduced arithmetic operations and computing resources with fewer parameters. In hyperparameter optimization, BO has focused on Euclidean domains that only permit tuning scalar hyper-parameters. However, when BO applies to NAS, a suitable kernel function is needed to build GP, and a distance function is required to describe the differences between two neural network architectures. We define a kernel function suitable for the search space and describe the similarity of network architecture appropriately. We also apply the particle swarm optimization algorithm to optimize the acquisition function. Network morphism is combined with Bayesian optimization (BO), so the obtained architecture inherited the knowledge of the parent network, realizing a better trade-off between exploitation and exploration. The adaptability of the GP model is enhanced by self-optimization of network architecture, and dynamic variable model super parameters gradually fit the real data distribution. We can first obtain the cell architecture from a low-resolution and small dataset and then superimpose its number to expand to a more extensive neural network architecture suitable for the feature of coal from different coal washery. The network CellNet outperforms architectures manually designed that is validated on coal and gangue classification problem.

Open-domain question answering extract the answer for a question from a large scale corpus, which typically employs ranker to filter irrelevant paragraphs. As traditional retrieval approaches based on statistical characteristics, neural ranking model has gradually become a popular choice for filtering as its capacity of semantic features extraction. But massive corpus request high performance for filtering efficiency. Thus, acceleration of textual encoding by neural model is worthy of attention and there are many achievements on various of perspectives. Inspired by reading habit of human, this paper presents a novel neural ranker on sentence-level jump encoding which can better skip those irrelevant sentences than existing neural models. When people browsing text with questions, they tend to find the key sentence based on heuristic information of question and current reading. And our sentence-level jump encoder works in a similar way which can effectively and accurately locate key sentences for question. We evaluate our neural ranker on long paragraph setting of Quasar-T dataset, our method achieves significant improvements on both ranking acceleration and ranking accuracy, as well as surpass other question answering baselines on the whole question answering results.

Although matrix factorization and its variants have proved their effectiveness in learning user preferences, they violate the triangle inequality and may fail to capture the inner grained preference information. Thus, metric learning based models have attracted increasing interests in recommender systems. However, most of these models only measure the distance in Euclidean space, though it is known to limit the prediction performance. To address this problem, this paper proposes curvilinear collaborative metric learning (CCML) as a solution. Rather than learning from Euclidean distance, CCML measures the “path” (i.e. curvilinear distance) between the user and items. Therefore, it is able to better infer information (in geometrical representation) and user-item relations behind the observed data points. Under this setting, the relations between the user and items are not linearly reflected by the relation vector. In order to effectively learn the representation of items, we propose to fuse macro features of categories and micro representations of items to form their latent vectors. Our model can be less prone to be affected by the noise commonly presented in datasets by separately learning macro and micro features. Experimental results on eight public recommendation datasets demonstrate that CCML produces a outstanding performance compared with five competitive baselines. Finally, quantitative analysis is included to verify the efficiency and effectiveness of CCML and the macro-micro module.

Recognizing discourse relations is crucial for understanding semantic and logical connections between two discourse units in the text. The most of existing work only considers the single-level and single-perspective discourse relationship between discourse units, which will lead to inconsistencies between recognizing classes and sub-classes of the discourse relations and cannot distinguish very similar relations in semantics. Therefore, in this paper, we propose a Multi-perspective and Hierarchical Model (MHM) that can model hierarchical relationships and capture the implicit connections between discourse units from multi-perspective (i.e., rhetorical, co-referential, and temporal) to strengthen the distinction between different discourse relations. Specifically, we first build a hierarchical classification module with contrastive learning to mine the finer semantic in two-level discourse relations and recognize relations consistently. Then, we introduce recognizing co-referential and temporal relations as auxiliary tasks and build a mapping from rhetorical relations to them, modeling multi-perspective relationships between discourse units. The experimental results on MCDTB and CDTB show that our model achieves the best performance.

In few-shot learning, methods based on graph neural network (GNN) are proposed for full exploitation of relations between samples. However, most GNN-based methods only explore a very rough relation among instances, ignoring that the relation between two samples is more complicated than a similarity value can convey. In this paper, we propose a model named GNN-based Relations Extraction for local features (GNN-RE in short) to explore the complex relations between instances. Specifically, we represent each image with multiple local semantic features, with each representing one type of semantic information. By exploring the relations among local semantic features in different images, we can learn more complex relations between images. To evaluate the effectiveness of the proposed GNN-RE, we conduct extensive experiments on two benchmark datasets. Our method achieves better or competitive performance comparing with previous work.

Securing communications is an increasingly challenging problem. While communication channels can be secured using strong ciphers, attackers who gain access to the channel can still perform certain types of attacks. One way to mitigate such attacks is to verify the integrity of exchanging messages between two parties or more. While there are robust integrity check mechanisms currently, these lack variety and very few are based on machine learning. This paper presents a methodology for performing an integrity check inspired by recent advances in neural cryptography. We provide formal, mathematical functions and an optimization problem for training an adversarial neural cryptography architecture. The proposed neural architectures can adequately solve the problem. In our experiments, a receiver can verify if incoming messages are authentic or altered with an accuracy greater than 99%. This work expands the repertoire of integrity checking methodologies, provides a unique perspective based on neural networks, and facilitates data security and privacy.

For modern recommender systems, the issue of data sparsity still often arises in many cases. To address this challenge, Cross-Domain Recommendation (CDR) has been explored by transferring knowledge learned from the source domain to alleviate the data sparsity in the target domain, and existing methods typically assume intra-domain samples share more commonalities than variabilities. However, it can also cause negative effects when there exists a high diversity among intra-domain samples. In this paper, we propose a novel Mixed Gate Control (MGC) model to fulfill 1) the division of samples into groups according to their commonalities and 2) the transfer of knowledge based on groups. We stack up multiple MGC layers into Multiple Layers MGC (ML-MGC) and apply it to a large-scale bundle retrieval system in a listed internet company. We conduct experiments on its own commercial dataset and two public real-world datasets. Experimental results show the superiority of our model against state-of-the-art methods for the CDR task. Finally, we present a case study to illustrate the groups identified by our model.

Accurate and timely prediction of the Internet traffic flow is important for network performance improvement. Few prediction methods incorporate the topology information of networks. In this paper, we present a novel Internet traffic forecasting algorithm named TTGCN, which applies the graph neural networks for traffic flow prediction on each link of a backbone network. The topology of the network was represented by a novel adjacency matrix, which models the relationship between links. The design makes TTGCN capable of capturing both the temporal and topological information of the traffic flow. TTGCN is validated on UKERNA, a dataset captured from a real backbone network. The experimental results show that the average error of the proposed TTGCN over 90 minutes is 28.249 (Mbit/s), which is a significant improvement of conventional models, while that of ARIMA and GRU and STGCN are 44.955, 40.935, and 36.152, respectively. The proposed TTGCN, by encoding both temporal and topological information in a neural representation, can achieve better prediction performance on network traffic.

In elderly healthcare the risk of falling is elevated, and falling can have serious consequences. Early fall detection can greatly improve patient outcomes. Sensors including accelerometers can provide the data necessary for fall detection, but correctly distinguishing falls from non-falls is a difficult task. In this paper we survey recent advances including the use of recurrent and convolutional neural networks, and we propose a novel ResNet architecture and variants, adapted to multivariate time series data. We carry out a systematic experiment comparing our method against several baselines and recent state-of-the-art models, using the Farseeing dataset. We find that the proposed ResNet architecture is the best-performing model across metrics, and we investigate the internal features it learns to enable this. We also investigate different approaches to data augmentation in the context of the lack of labelled data.

Construction Cost Index (CCI) is a weighted aggregation index published by Engineering News-Record (ENR), a world-famous academic journal in the field of engineering and construction. CCI provides a long recorded construction costs that reflect the overall cost level of construction and engineering development status. Therefore, many companies and cost estimators have done a lot of work for CCI prediction in order to better plan investment and prepare bids. In this paper, in order to improve the accuracy of CCI prediction, unlike previous forecasting methods based on causal model and statistical analysis, we propose a network method based on the visibility graph to better forecast CCI. The main idea is to map CCI data into networks, so that each CCI data can be seen as a node in the network, and then use node visibility and network structures to predict CCI. The proposed method is applied to predict the short-term, mid-term and long-term CCI separately, and compared with other popular prediction models based on common evaluation indicators. The experimental results show that this method has good prediction performance and can effectively improve the CCI prediction accuracy.

The growing interest for adversarial examples, i.e. maliciously modified examples which fool a classifier, has resulted in many defenses intended to detect them, render them inoffensive or make the model more robust against them. In this paper, we pave the way towards a new approach to improve the robustness of a model against black-box transfer attacks. A removable additional neural network is included in the target model, and is designed to induce the luring effect, which tricks the adversary into choosing false directions to fool the target model. Training the additional model is achieved thanks to a loss function acting on the logits sequence order. Our deception-based method only needs to have access to the predictions of the target model and does not require a labeled data set. We explain the luring effect thanks to the notion of robust and non-robust useful features and perform experiments on MNIST, SVHN and CIFAR10 to characterize and evaluate this phenomenon. Additionally, we scale the luring effect to ImageNet, experiment practical use of it and discuss its complementarity with other defense schemes.

Knowledge distillation is a widely used neural network model compression technique. In general, the knowledge distillation transfer the knowledge from a large pre-trained teacher network with superior performance to a small student network enables the student network to achieve better performance. This paper proposes a triplet knowledge distillation framework (abbreviated as TKD), which introduces a smaller assistant network into the knowledge distillation structure. The performance of the assistant network is lower than that of the student network. During the training of the TKD, by minimizing the Mean Squared Error(MSE) loss function, the output of the student network will closer to the output of the teacher network and further from that of the assistant network. Therefore, the student network can learn more expressive knowledge from the teacher network while throwing away mistaken knowledge in the assistant network. Finally, the student network achieves a surprising performance even superior to the teacher network. We have demonstrated the effectiveness of TKD by extensive experiments on benchmark datasets(CIFAR-10, CIFAR-100, SVHN, STL-10). When using VGGNet as an experimental model, the student network VGGNet13 achieving 94.29%, 75.30%, 95.53%, and 87.61% accuracy on the CIFAR-10, CIFAR-100, SVHN, and STL-10 datasets, improved by 1.24%, 2.81%, 0.40%, and 2.32%, respectively.

Developers change software programs for various purposes (e.g., bug fixes, feature additions, and code refactorings), but the intents of code changes are often not recorded or are poorly documented. To automatically infer the change intent of each program commit (i.e., a set of code changes), existing work classifies commits based on commit messages and/or the sheer counts of edited files, lines, or abstract syntax tree (AST) nodes. However, none of these tools reason about the syntactic or semantic dependencies between co-applied changes, neither do they adopt any deep learning method. To better characterize program commits, in this paper, we present CClassifier—a new approach that classifies commits by (1) using advanced static program analysis to comprehend relationship between co-applied edits, (2) representing edits and their relationship via graphs, and (3) applying convolutional neural networks (CNN) to classify those graphs. Compared with prior work, CClassifier extracts a richer set of features from program changes; it is the first to classify program commits using CNN. For evaluation, we prepared a benchmark that contains 7,414 code changes from 5 open-source Java projects. On this benchmark, we empirically compared CClassifier and the state-of-the-art approach with five-fold cross validation. On average, when predicting bug-fixing commits within the same projects, CClassifier improved the prediction accuracy from 70% to 72%. More importantly, prior work seldom identifies feature-addition commits; CClassifier can successfully identify such commits in a lot more scenarios. Our evaluation shows that CClassifier outperforms prior work due to its usage of advanced program analysis and CNN.

A Convolutional neural network (CNN) has emerged as a widely used approach to computer vision tasks, including object classification and detection tasks. The high requirement for the model to be more computationally efficient on lower information and communication technology (ICT) resource, e.g., mobile terminals can benefit from model distillation. However, most existing distillation methods suffer from a significant accuracy reduction, which requires a large number of pre-training models or doesn't make good use of the more of the network information, e.g., in the middle layers, during the distillation. In this paper, we study how knowledge about traffic signs recognition could be transferred to smaller models by distillation while cutting channels. We present an optimized object detection network, which uses a Region Proposal Network (RPN) weighted loss and hard-soft distribution-wise distillation loss for structural differences between teacher and student networks. We validate the network on multiple real-world datasets, the experiments demonstrate that the classification accuracy can be improved by 9 % with about 16 times parameter reduction while the detection network performance could be increased by 10.6% using an optimized object detection network.

Learning and classification of imbalanced data is a quite common challenge in machine learning. In order to generate more high-quality minority samples in imbalanced data and help solve the classification process, this paper proposes a novel local tangent generative adversarial network (LT-GAN). In LT-GAN, a local tangent based generator is designed to generate realistic and diverse minority class samples by learning the local tangent space of original minority samples. Meanwhile, a two-function discriminator is explored to judge the authenticity of samples and distinguish the majority samples from the minority samples. With the synthesis of minority class samples, the generator and discriminator are trained together by using adversarial learning. Experiments and comparisons show that our proposed LT-GAN outperforms other techniques and significantly improves the classification performance of imbalanced data.

Substructures of drugs are important for drug-drug interaction (DDI) prediction because drugs with similar chemical structures are prone to share similar properties. There are common substructures (i.e., functional groups) that play significant roles in DDI prediction. However, the existing computational methods can't fully utilize common structural patterns between drugs for DDI prediction. In this paper, we develop a substructure-based framework named StructDDI which can fully utilize common structural patterns between drugs. A graph processing method based on the random walk is proposed to generate the representation of drugs. A novel feature extraction component that includes dual convolutional neural networks (CNNs) is proposed to automatically summarize structural and chemical representation. The proposed StructDDI was evaluated on two real-world datasets and performed better than state-of-the-art baselines.

Online anomalous taxi trajectory detection, which identifies anomalies from ongoing taxi trajectories, has become an important and fundamental concern in many real-world applications. Most of the existing studies define the anomalous trajectories as the ones deviating from the majority of routes or showing abnormal driving time and distance at the same time. However, due to the complexity of road conditions and the variety of passenger preferences, those methods have large false-positive rates, i.e., reporting many normal routes as anomalies. A high false-positive rate is harmful since false alarms will 1) bring unnecessary panic to passengers, 2) cause fiscally punishments to normal drivers, and 3) wastes human resource to deal with drivers' complaints. To this end, this paper proposes an online anomalous trajectory detection method, namely multidimensional criteria based anomalous trajectory (MCAT), to identify anomalous trajectories online. It judges anomalies by considering multidimensional criteria (similarity, time, distance) at the same time, reducing the false positives without sacrificing false negative rates. We evaluate the proposed method based on the real-world taxi data collected from Shanghai, China. The experimental results demonstrate that our method can outperform state-of-the-art methods in terms of accuracy, false-negative rate, and false-positive rate.

The task of reasoning over knowledge graph for factoid questions has received significant interest from the research community of natural language processing. Performing this task inevitably faces the issues of question complexity and reasoning efficiency. In this paper, we investigate modern reasoning approaches over knowledge graph to tackle complex factoid questions of diverse reasoning schemas with attractive speedup in computational efficiency. To this end, we propose two evidence retrieval strategies to generate concise and informative evidence graph of high semantic-relevance and factual coverage to the question. Then, we adopt DELFT, a graph neural networks based framework that takes the linguistic structure representation of a question and the evidence graph as input, to predict the answer by reasoning over the evidence graph. We evaluate the performance across several baselines in terms of effectiveness and efficiency on two real-world datasets, MOOCQA and MetaQA. The results show the superiority of message passing paradigm in delivering a robust reasoner with better answer quality and significantly improved computational efficiency.

Existing image captioning approaches fail to generate fine-grained captions due to the lack of rich encoding representation of an image. In this paper, we present an attentive contextual network (ACN) to learn the spatially transformed image features and dense multi-scale contextual information of an image to generate semantically meaningful captions. At first, we construct deformable network on intermediate layers of convolutional neural network (CNN) to cultivate spatial invariant features. And the multi-scale contextual features are produced by employing contextual network on top of last layers of CNN. Then, we exploit attention mechanism on contextual network to extract dense contextual features. Further, the extracted spatial and contextual features are combined to encode the holistic representation of an image. Finally, a multi-stage caption decoder with visual attention module is incorporated to generate fine-grained captions. The performance of the proposed approach is demonstrated on COCO dataset, the largest dataset for image captioning.

Spiking neural networks (SNNs) are promising to replicate the efficiency of the brain by utilizing a paradigm of spike-based computation. Training a deep SNN is of great importance for solving practical tasks as well as discovering the fascinating capability of spike-based computation. The biologically plausible scheme of local learning motivates many approaches that enable training deep networks in an efficient parallel way. However, most of the existing spike-based local learning approaches show relatively low performances on challenging tasks. In this paper, we propose a new spike-based temporal dependent local learning (TDLL) algorithm, where each hidden layer of a deep SNN is independently trained with an auxiliary trainable spiking projection layer, and temporal dependency is fully employed to construct local errors for adjusting parameters. We examine the performance of the proposed TDLL with various networks on the MNIST, Fashion-MNIST, SVHN and CIFAR-10 datasets. Experimental results highlight that our method can scale up to larger networks, and more importantly, achieves relatively high accuracies on all benchmarks, which are even competitive with the ones obtained by global backpropagation-based methods. This work therefore contributes to providing an effective and efficient local learning method for deep SNNs, which could greatly benefit the developments of distributed neuromorphic computing.

Adversarial attacks/perturbations are becoming important for NLP research, as it has been shown recently that text-attacking adversaries can degrade an NLP model's performance without the victim's knowledge. This has far-reaching implications, especially when an NLP system is deployed in critical applications such as health or finance. In fact, the robustness of state-of-the-art models such as NLP transformers have increasingly been scrutinised due to their vulnerability against adversarial perturbations such as TextFooler and BERT-Attack. These methods, however, focus on changing words, which at times, ruins the readability and semantics of the sample. This paper introduces Special Symbol Text Attacks ‘SSTA’, a technique to improve the performance of language adversarial perturbations using special symbols that have downstream task information associated with them even though that should not have been the case. Our tests show that introducing such symbols which are meaningless to a human within a sentence, can perturb the sample in a particular direction. When this technique is used with TextFooler, which is a recent benchmark for the creation of NLP adversaries, through the TextAttack framework, it can improve all main evaluation metrics on three sentiment classification tasks and fake news detection. A simple, novel and symbol-specific adversarial learning technique is then introduced to reduce the influence of such special symbols.

Generating text with high quality and sufficient diversity is a fundamental task in natural language generation. Although generative adversarial networks (GANs) achieve promising results in text generation, GAN-based language models suffer from mode collapse, i.e., the generator tends to sacrifice diversity and focus on limited text patterns with high quality. By contrast, maximum likelihood estimation (MLE) based language models could cover various text patterns and generate diversified samples with poor quality. This paper proposes a cooperative framework with GANs and entropic auto-encoders (EAEs), named GAN-EAE, to synthesize their advantages for text generation, where EAEs are powerful MLE-based generative models based on deterministic auto-encoders. By imitating the output distribution of EAEs, the generator shapes its output distribution closer to the real data distribution against mode collapse. Meanwhile, by learning the samples from the generator of GANs, EAEs subtly distribute probability mass on high quality patterns for improving generation quality. The similar samples obtained from the generator may raise mode collapse and should be downplayed during adversarial training. Thus, a sample re-weighting mechanism is adopted to improve diversity by measuring the inner distance of generated samples. Experimental results demonstrate that GAN-EAE could improve both GANs and EAEs to achieve state-of-the-art performance.

Prognostication of vehicle trajectories in unknown environments is intrinsically a challenging and difficult problem to solve. The behavior of such vehicles is highly influenced by surrounding traffic, road conditions, and rogue participants present in the environment. Moreover, the presence of pedestrians, traffic lights, stop signs, etc., makes it much harder to infer the behavior of various traffic agents. This paper attempts to solve the problem of spatio-temporal look-ahead trajectory prediction using a novel recurrent neural network called the Memory Neuron Network. The Memory Neuron Network (MNN) attempts to capture the input-output relationship between the past positions and the future positions of the traffic agents. The proposed prediction model is computationally less intensive and has a simple architecture as compared to other deep learning models that utilize LSTMs and GRUs. It is then evaluated on the publicly available NGSIM dataset and its performance is compared with several state-of-art algorithms. Additionally, the performance is also evaluated on a custom synthetic dataset generated from the CARLA simulator. It is seen that the proposed model outperforms the existing state-of-art algorithms. Finally, the model is integrated with the CARLA simulator to test its robustness in real-time traffic scenarios.

The interpretable charge prediction task is to predict the final charges according to the fact descriptions, at the same time generate the corresponding explanations. The task is often more difficult than people think, as the explanation and prediction are highly interdependent, which is ignored by existing methods. In this paper, we deal with the interpretable charge prediction task from the perspective of generating the prediction and explanation (court view) pair while considering their dependency. To this end, we propose a Joint Prediction and Generation Model, named JPGM, which includes a coarse-to-fine classifier and a keyword-aware generator. Specifically, firstly, the classifier predicts a group of charge labels given the fact description. Then, we select the most matchable keywords of every charge from the pre-defined charge-discriminative keywords by an attention mechanism. Furthermore, the generator generates explanation using both the fact description and the most matchable keywords. Finally, to refine the prediction, the generated explanation and fact description are fused for final charge prediction. The classifier and generator are trained under an alternating procedure, which alleviates the error propagation. The experimental results validate that our model can effectively address the dependency issue and predict the charge with more interpretability.

It is necessary for the development of self-driving to fulfill the requirements for safety, stability, and intelligence, especially in high-speed conditions. Therefore, the detection of pedestrians that may occur in highway scenarios during driving and understanding the meaning of their behaviors in advance are significantly important for the self-driving vehicle to make correct decisions. However, no existing datasets are available for behavior recognition in self-driving scenarios. In order to advance the task of interactive cognition between the vehicle and pedestrians, in this paper, we present a new dataset, called THU-IntrudBehavior, that collects lane intrusion behaviors of pedestrians that can be applied in real world highway scenarios. The dataset contains diverse behaviors of single or multiple pedestrians/cyclists that are simulated in different urban roads under various weather conditions. We describe annotations of each video and report several experimental results of baseline methods on our self-collected dataset. Our THU-IntrudBehavior dataset provides new support for behavior recognition in high-speed conditions for self-driving.

With the increase of data centers, the number of disks also grows rapidly. Therefore, the prediction of disk failures has become an important task for both academia and industry. Existing prediction schemes predict disk failure in the short prediction horizon or with a short time window. However, these schemes cannot achieve ideal performance for a long prediction horizon with a long time window. In this paper, we proposed a deep learning method that can effectively solve the above problems. We refine the Self-Monitoring, Analysis and Reporting Technology (SMART) attributes by using information entropy to select the most related attributes for prediction. Moreover, we proposed the Multiple Channel Convolutional Neural Network based LSTM (MCCNN-LSTM) model to predict whether disk failures will occur in a given disk in next few days. We further evaluate the MCCNN-LSTM model by comparing it with the state-of-the-art works. Extensive experiments show that our model can improve FDR (Fault Detection Rate) to 99.8% and reduce FAR (False Alarm Rate) to 0.2%.

Deep CNNs, though have achieved the state of the art performance in image classification tasks, remain a black-box to a human using them. There is a growing interest in explaining the working of these deep models to improve their trustworthiness. In this paper, we introduce a Posthoc Architecture-agnostic Concept Extractor (PACE) that automatically extracts smaller sub-regions of the image called concepts relevant to the black-box prediction. PACE tightly integrates the faithfulness of the explanatory framework to the black-box model. To the best of our knowledge, this is the first work that extracts class-specific discriminative concepts in a posthoc manner automatically. The PACE framework is used to generate explanations for two different CNN architectures trained for classifying the AWA2 and Imagenet-Birds datasets. Extensive human subject experiments are conducted to validate the human interpretability and consistency of the explanations extracted by PACE. The results from these experiments suggest that over 72% of the concepts extracted by PACE are human interpretable., Accepted at International Joint Conference on Neural Networks (IJCNN 2021)

Sequential recommendation that aims to predict user preference with historical user interactions becomes one of the most popular tasks in the recommendation area. The existing methods concentrated on user's sequential features among exposed items have achieved good performance. However, they only rely on single item prediction optimization to learn data representation, which ignores the association between context data and sequence data. In this paper, we propose a novel self-supervised learning based sequential recommendation network (SSLRN), which contrastively learns data correlation to promote data representation of users and items. We design two auxiliary contrastive learning tasks to regularize user and item representation based on mutual information maximization (MIM). In particular, the item contrastive learning captures sequential contrast feature with sequence-item MIM, and the user contrastive learning regularizes user latent representation with user-item MIM. We evaluate our model on five real-world datasets and the experimental results show that the proposed framework significantly and consistently outperforms state-of-the-art sequential recommendation techniques.

3D object detection based on LiDAR point cloud has gained more and more attention from industry and academia. Most of the previous works are carried out on the KITTI dataset, which has the gap with real-world scenes in terms of data size and distribution. Faced with large-scale datasets, existing methods have encountered with some problems, especially in the detection of difficult targets and heading prediction. In this paper, we design an integrated framework to improve 3D object detection accuracy targeting on large-scale datasets. We explore data augmentation methods from the perspective of data-rich scenes. Two well-designed auxiliary losses are introduced into our framework to relieve the problem caused by the diversity of points density and distribution. To promote the performance of heading prediction, the existing heading classifier is replaced by our novel dual heading classifier. Experiments on the largest and most realistic 3D object detection benchmark of Waymo Open Dataset show that our integrated framework boosts the detection accuracy for a large margin.

Aspect-based sentiment analIsis (ABSA) aims to detect the sentiment polaritI of a specific aspect in an opinionated sentence. Current work focuses on exploiting the sIntactic tree to shorten the distance between the aspect term and context words. However, the “hard-pruning” strategI on the sIntactic tree maI lead to the reduction of importa nt sIntactic information. In this paper, we propose a novel sInt actical distance attention guided graph convolutional network (SDGCN) for ABSA. Our model is capable of fullI exploiting the sIntactic knowledge with a “soft pruning” strategI and learning crucial fine-grain sIntactic distance info rmation. AdditionallI, an effective denselI connected graph convolutional laIer is applied to avoid the over-sm oothing problem of standard GCN. Experiments conducted on three benchmark datasets show that our model achieves promising results comparing to the baseline models.

Existing lexically constrained machine translation employs data augmentation and incorporates lexical constraints during decoding period, which requires a bilingual dictionary or costs much decoding time. In this paper, we propose a simple but effective method to leverage lexical constraints. We use separate encoders to encode source sentence and lexical constraints, with self-attention layer mask to disentangle the two encoding subtasks. Our method does not require bilingual dictionaries or modify decoding process. Experiments on WMT 2016 English-German (En-De) and IWSLT 2017 English-Chinese (En-Zh) datasets show that our method gains improvement compared to baseline models.

With the ever increasing complexity of the user queries in a multi-domain based task-oriented dialogue system, it is imperative to facilitate robust Spoken Language Understanding (SLU) modules that perform multiple tasks in an unified way. In this paper, we present a novel multi-task approach for the joint modelling of three tasks together, namely, Domain Classification, Intent Detection and Slot-Filling. We hypothesize with the intuition that the cross dependencies of all these three tasks mutually help each other towards their representations and classifications which further simplify the SLU module in a multi-domain scenario. Towards this end, we propose a BERT language model based multi-task framework utilizing capsule networks and conditional random fields for addressing the classification and sequence labeling problems, respectively, for different tasks. Experimental results indicate that the proposed multi-task model outperformed several strong baselines and its single task counterparts on three benchmark datasets of different domains and attained state-of-the-art results on different tasks.

Rumor detection is a task of identifying information that spread among people whose truth value is false or unverified, and it has been a great challenge due to the rapid development of social media. The traditional machine learning based detection methods can make full use of informative features but cannot extract high-level representations. Other methods involved deep learning neural networks exploit propagation structural information to achieve high accuracy, for example, Bi-Directional Graph Convolution Networks(BiGCN) achieved the best performance on rumor detection by operating on bottom-up and top-down structures. However, those deep learning methods ignore other useful features like content-based features. In this paper, we not only make full use of three aspects of features based on a new concept: kernel subtree, which focus more on informative features of influential nodes of an event, but also propose a new model, which consists of Separation Convolution blocks, Long Short Term Memory(LSTM) and Squeeze and Excitation Networks(SENet), to make comprehensive use of features extracted on the basis of kernel subtree. First, we utilize Separation Convolutions to learn more local information with different kernel size, then LSTM can learn high-level interactions among features and find more global information. After that, SENet applies attention mechanism to put more weights on informative channels of feature maps. Meanwhile, on test set, Gradient Boosting Decision Tree(GBDT) is used to assist our model with few events. The experiments on the PHEME dataset show that our approach can identify rumors with accuracy 95% which outperforms BiGCN by 10% at least.

Neural networks (NNs) have been applied in numerous real-life applications and services. These include the applications in disease and healthcare analytics, such as identification and predictive analytics of coronavirus disease 2019 (COVID-19). However, many existing NN-based solutions train the models based on data (e.g., computed tomography (CT) scan images, viral/molecular test results) that can be expensive to produce and/or not easily accessible. They also require large volumes of these data for training. However, partially due to privacy concerns and other factors, the volume of available COVID-19 data can be limited. Hence, in this paper, we present a solution for predictive analytics of COVID-19 with NNs. Our solution consists of three algorithms, which make good use of autoencoder and few-shot learning, to train the prediction model with only a few samples of more accessible and less expensive types of data (e.g., serology/antibody test results from blood samples). Evaluation results on a real-life Brazilian COVID-19 dataset demonstrate the effectiveness of our solution in predictive analytics of COVD-19 with NNs.

Text embedding has recently emerged as a very useful and successful method for semantic representation. Following initial word-level embedding methods such as Latent Semantic Analysis (LSA) and topic-based bag-of-words approaches like Latent Dirichlet Allocation (LDA), the focus has turned to language models and text encoders implemented as neural networks - ranging from word-level models to those embedding whole documents. The distinctive feature of these models is their ability to infer semantic spaces at all levels based purely on data, with no need for complexities such as syntactic analysis or ontology building. Many of these models are available pre-trained on enormous amounts of data, providing downstream applications with general-purpose semantic spaces. In particular, embedding models at the sentence level or higher are most useful in applications because the meaning of text only becomes clear at that level. Most text embedding methods produce text embeddings in high-dimensional spaces, with a dimensionality ranging from a few hundred to thousands. However, it is often useful to visualize semantic spaces in very low dimension, which requires the use of dimensionality reduction methods. It is not clear what language models and what method of dimensionality reduction would work well in these cases. In this paper, we compare four text embedding methods in combination with three methods of dimensionality reduction to map three related real-world datasets comprising textual descriptions of items in a particular domain (sports) to a 2-dimensional semantic visualization space. The results provide several insights into the utility of these methods for data of this type.