Search

Meta Reinforcement Learning (MRL) enables an agent to learn from a limited number of past trajectories and extrapolate to a new task. In this paper, we attempt to improve the robustness of MRL. We build upon model-agnostic meta-learning (MAML) and propose a novel method to generate adversarial samples for MRL by using Generative Adversarial Network (GAN). That allows us to enhance the robustness of MRL to adversal attacks by leveraging these attacks during meta training process.

In abstractive single-document summarization task, generated summaries always suffer from fabricated and less informative content. An intuitive way to alleviate this problem is to merge external semantic knowledge into the model framework. In this paper, we incorporate explicit graphs based on semantic knowledge, including term frequency, discourse information, and entities with their relations, into neural abstractive summarization for the problem. We propose a novel model for abstractive single-document Summarization based on Semantic Knowledge Graphs (SKGSUM), which regards sentences and entities as nodes, captures the relations between units in different textual levels, and focuses on salient content in the source documents to guide the summary generation process. To the best of our knowledge, we are the first to exploit different textual-unit levels explicit graph representations in a unified framework for the neural abstractive summarization task. Results show that our model achieves significant improvements on both XSum and CNN/Daily Mail datasets over some strong baselines. Human evaluations further indicate that our model can generate informative and coherent summaries.

There is no doubt that the use of machine learning is increasing every day. Its applications include self-driving cars, malware detection, recommendation systems and many other fields. Although the broad scope of this technology highlights the importance of its reliability, it has been shown that machine learning models can be vulnerable to adversarial attacks. In this paper, we study a property of these attacks called transferability across different architectures and models, measuring how these attacks transfer based on a specific number of parameters among three adversarial attacks: Fast Gradient Sign Method, Projected Gradient Descent and HopSkipJumpAttack.

Significant progress has been made in the area of text classification and natural language processing. However, like many other datasets from across different domains, text-based datasets may suffer from class-imbalance. This problem leads to model's bias toward the majority class instances. In this paper, we present a new approach to handle class-imbalance in text data by means of unsupervised learning algorithms. We present class-decomposition using two different unsupervised methods, namely k-means and Density-Based Spatial Clustering of Applications with Noise, applied to two different sentiment analysis data sets. The experimental results show that utilizing clustering to find within-class similarities can lead to significant improvement in learning algorithm's performances as well as reducing the dominance of the majority class instances without causing information loss.

Deep learning has achieved great results in the last decade, however, it is sensitive to so called adversarial attacks - small perturbations of the input that cause the network to classify incorrectly. In the last years a number of attacks and defenses against these attacks were described. Most of the defenses however focus on defending against gradient-based attacks. In this paper, we describe an evolutionary attack and show that the adversarial examples produced by the attack have different features than those from gradient-based attacks. We also show that these features mean that one of the state-of-the-art defenses fails to detect such attacks.

Recently, meta learning methods able to provide multiple initializations have drawn much attention due to its capability of handling multi-modal tasks. However, the modal differences in multi-modal distributions aggravate the catastrophic forgetting. In this paper, we augment multi-initial meta learning for continual learning, which alleviates the forgetting issue while allowing beneficial transfer to previous tasks from multi-modal distributions. Specifically, we propose a multi-modal meta continual learning (M3CL) framework, which is able to mitigate forgetting via gradient episodic memory (GEM). Modulation network in M3CL has the capability of identifying the mode of each image from online multimodal distributions. In addition, in order to take advantage of both modulation network and GEM, we consider the entropy-based regularization to learn effective modulation network and penalize overfitting to a specific task. Experiments on multimodal datasets identify that the modal differences indeed aggravate the catastrophic forgetting, and demonstrate the strong performance of M3CL in comparison to the state-of-the-art.

Adversarial training is one of the most effective methods for defending against adversarial attacks, but it is computationally costly. In this paper, we propose Saliency Adversarial Defense (SAD), an efficient defense algorithm that avoids adversarial training by minor modification of a deployed model. The saliency map is added to the input to enhance those pixels that are important for making decisions. This process causes a distribution shift to the original data. Interestingly, we find that this shift can be effectively fixed by only updating the statistics of batch normalization with the processed data without further training. We verify that SAD enlarges the average distance between processed data and the updated decision boundary, and significantly smooths the landscape. Hence the model can defend against stronger attacks and improve robustness. We show in experiments that the results hold for complex models and datasets. Our results demonstrate that SAD is superior in defending various attacks, including both white-box and black-box ones.

The newly proposed Broad Learning System (BLS) offers an alternative way to deep learning which saves a time-consuming training process and powerful computing resources. However, the randomly generated feature nodes and lots of enhancement nodes in BLS may reduce the performance of the final classifier. Aiming at the problems in randomly generated feature nodes which suffer from unpredictability and need guidance, this paper proposes an adaptive feature nodes evolutionary algorithm (AFNE) to extract better features; While broad learning neural network often requires a large number of enhancement nodes parameters to ensure its performance, which easily leads to the redundancy or dependency between features, as well as the performance degradation of the final model. Therefore, this article also proposes a new criterion based on node sensitivity to prune the network enhancement layer nodes to remove redundant nodes, reduce the network scale, and increase the generalization ability. The proposed algorithm ADP-BLS can improve the accuracy and generalization performance of the final classifier through the evolution of feature nodes and the pruning of enhancement nodes. Extensive comparative experiments on real world data sets verify the effectiveness of the proposed ADP-BLS. At the same time, when verifying the module validity of the innovative algorithms added in this article, experiments also show that the AFNE and pruning method integrated into BLS can improve the model to a certain extent.

Chinese semantic matching is a fundamental task in natural language processing, which is critical and yet challenging for a series of downstream tasks. Although recent work on text representation learning has shown its potential in improving the performance on semantic matching, relatively limited work has been done on exploring the relevant interactive information between two granularity of Chinese text, i.e., character and word. Existing methods usually focus on capturing the interactive features from single granularity, which lead to inefficient text representation. Also, they typically fail to consider the fusion of features from different granularity. As a result, they only achieve limited performance improvement. This paper proposes a novel Chinese semantic matching model based on multi-granularity alignment and feature fusion (MAFFo). To be specific, we first encode the texts from different granularity, which are further handled with soft-alignment attention mechanism to extract relevant interactive information between texts on different granularity. In addition, we devise a feature fusion structure to merge the features from different granularity to generate an ideal representation for the pair of input text sequences, followed by a sigmoid function to judge the semantic matching degree. Extensive experiments on the publicly available dataset BQ demonstrate that our model can effectively improve the performance of semantic matching task and achieve comparable performance with BERT-based methods.

A major challenge in the field of object detection is to design a model with high precision and high speed, which is necessary in real applications. In this paper, we study the problem of object detection, and propose a multi-layer feature fusion model for object detection, called MLFF. MLFF not only takes into account the rich location information and detail information of low-level features, but also considers the rich spatial information of high-level features. In MLFF, we develop three different components, Backbone, Neck and Head, to gradually extract and fuse low-level feature maps and high-level feature maps. In addition, CBAM attention mechanism is also used in feature fusion, which enables us to achieve more accurate feature extraction. Extensive experiments on PASCAL VOC2007 and VOC2012 datasets show that our MLFF model not only has the speed of one-stage object detection model but also has the detection precision of two-stage object detection model. Considering both precision and speed, MLFF outperforms state-of-the-art detection models significantly.

Length of stay (LoS) in the intensive care unit (ICU) is a general outcome measure used as an indicator of both quality of care and resource use. Existing LoS prediction methods usually model the problem as a binary classification or regression problem. This paper proposes to address this problem under a sequential learning framework using an embedded representation of patients. First, we exploit electronic health records (EHR) to learn patient distributed representation. Second, we formulate the LoS prediction task as a Deep Bayesian contextual bandit problem, which sequentially predicts LoS based on contextual distributed information about patients, with the goal of maximizing overall prediction outcomes. Experimental results show that our model has superior performance than other sequential learning approaches in the LoS task using the publicly available MIMIC-III ICU dataset. We also evaluate the robustness of our framework under different patient representations.

Accurate 3D object detection has recently aroused interest in the context of emerging autonomous driving technologies. Existing approaches predominantly use LiDAR-Camera fusion method to fulfill this challenging task, while neglecting the fact that LiDAR and camera data are spatially correlated, and cannot well retain the edge information. To solve these problems, in this paper, we propose a novel LiDAR-Camera 3D object detection method, namely the Multi-scale Spatial Transformer Network (MST-Net). The proposed method exploits an innovative spatial alignment scheme based on the projection transformer network (PTN) to mitigate the effects of the perspective view caused by sensors. In the process of generating 3D bounding boxes, the Atrous Spatial Pyramid Pool (ASPP) is applied to spatially aligned fusion features in order to preserve edge information to the greatest extent. Extensive experiments are conducted on the popular dataset KITTI, and the results can demonstrate the superiority of the proposed method. In addition, the effectiveness of these two strategies has been illustrated in ablation studies.

As the mainstream in network traffic classification (NTC), machine learning (ML) based methods suffer performance degradation due to the imbalance distribution of Internet traffic. Data augmentation methods including the traditional oversampling techniques and the Generative Adversarial Network (GAN) based generation methods are most commonly used to counter the imbalance problem in NTC. However, the former is prone to overfitting and introducing noise. The latter overcomes the above weaknesses, but the quality of the generated traffic samples is difficult to judge. Besides, these methods all divide the imbalanced traffic classification problem into two subproblems, which cannot guarantee the global optimality. In this paper, we propose a GAN based Traffic Augmentation (TA-GAN) for imbalanced traffic classification. TA-GAN is an end-to-end framework that integrates the generation of the minority traffic samples with the training of the target classifier. We design the feedback mechanism to better guide the direction of the sample generation and simultaneously indicate the quality of the synthesized samples. Moreover, the existing deep learning-based NTC methods can be easily adapted to imbalance scenarios with TA-GAN. Comprehensive experiments on the public ISCXVPN2016 dataset demonstrate that TA-GAN effectively mitigates the influence of traffic imbalance (a maximum 14.64% improvement to the minority class' $F_{1}$ score) and outperforms the state-of-the-art methods.

In open-domain multi-turn dialogue, capturing related information from conversational history is key to generating relevant responses. In this paper, we propose a method to extract and fuse related context information at different levels of granularity with a word-level encoder, an utterance-level encoder, and a fusion decoder. The word-level encoder aims to obtain a context-aware representation for the last utterance, called a post, by attending the post to the context utterances at the word level. The utterance-level encoder is used to model the correlation among context utterances with the self-matching attention mechanism. Finally, the multi-granularity context representations from the two encoders are fused via the fusion decoder to improve the relevance of generated responses. Experimental results on the Ubuntu Dialog Corpus and Cornell Movie Dialog Corpus show that our methods significantly outperform all the strong baselines.

This paper addresses the practical issue of lacking training data when building named entity recognition (NER) systems. To this aim, we introduce a new active learning method for reducing the number of training samples required by the underlying NER system. Different from prior work that only focuses on training data, we define a new loss function that when estimating loss and uncertainty scores of training samples for selection, it takes also into account the uncertainty of the $K$ unlabelled test instances most similar to the unlabelled training instances. Experimental results on both general domain and clinical benchmark datasets show that the proposed active learning method allows to train the NER system with between 5% to 7% less training data compared to state of the art uncertainty sampling methods, while retaining high NER effectiveness.

Trajectory clustering is an important task in trajectory data mining. Grouping a trajectory dataset into clusters based on the similarity between vehicle trajectories is conducive to revealing the movement pattern of the vehicles. For trajectory clustering in a real urban road network environment, the existing methods have some deficiencies, including high time complexity in measuring the distance between trajectory sequences and poor clustering performance. This paper proposes a method for clustering vehicle traj ectories in the urban road network environment based on a word vector model. First, the original Global Positioning System (GPS) trajectory data of vehicles are converted into road segment sequences by using the urban road segment information contained in the GPS trajectory data (after map-matching). Then, the spatial properties, such as the geographical location, time and moving direction of vehicles in the original trajectory sequence, are expressed by road segments and their order. Next, the road segment sequences are converted into traj ectory segment eigenvectors with fixed dimensions using the doc2vec model, in this way, the calculation efficiency of similarity between traj ectory segments of different lengths is improved. Finally, the vehicle tracks are clustered according to the distances between the traj ectory segment eigenvectors using the hierarchical clustering method. The result of a simulation based on taxi trajectory data gathered in a real urban road network shows that the proposed method was superior to the traditional clustering methods based on trajectory space-time distance, improving the silhouette index by 10%-25%, and reducing the clustering time by two orders of magnitude.

Intelligent flood forecasting systems provide an effective means to forecast flood disaster. Accurate flood flow value prediction is a huge challenge since it's influenced by both spatial and temporal relationship among flood factors. Popular deep learning structures like Long Short-Term Memory (LSTM) network lacks abilities of modeling the spatial correlations of hydrological data, thus cannot yield satisfactory prediction results. Moreover, not all the temporal information is always valuable for flood forecasting. In this paper, we proposed a novel spatial and temporal aware Graph Convolution Network (ST-GCN) for flood prediction, which is capable to extract spatial-temporal information from raw flood data. Moreover, a temporal attention mechanism is introduced to weight the importance of different time steps, thus involving global temporal information to improve flood prediction accuracy. Compared with the existing methods, results on two self-collected datasets show that ST-GCN greatly improves the prediction performance.

Multiple Object Tracking (MOT) has been a useful yet challenging task in many real-world applications such as video surveillance, intelligent retail, and smart city. The challenge is how to model long-term temporal dependencies in an efficient manner. Some recent works employ Recurrent Neural Networks (RNN) to obtain good performance, which, however, requires a large amount of training data. In this paper, we proposed a novel tracking method that integrates the auto-tuning Kalman method for prediction and the Gated Recurrent Unit (GRU), and achieves a near-optimum with a small amount of training data. Experimental results show that our new algorithm can achieve competitive performance on the challenging MOT benchmark, with higher efficiency and more robustness compared to the state-of-the-art RNN-based online MOT algorithms.

In this paper we investigate deep learning architectures combined with low-discrepancy sampling as surrogate models. The aim is to provide a quick estimate of the outcome of an expensive process when many evaluations are needed, e.g., for optimization purposes. The simulation of urban traffic, routinely employed for the design and performance evaluation of policies in urban mobility scenarios, is an example of procedure characterized by very complex dynamics, large dimensionality and intensive computational requirements. In this context, the need for a surrogate model arises in many forms, e.g., for origin-destination demand calibration, traffic light optimization, strategic planning. In order to cope with this complexity and large dimensionality we take advantage of the excellent approximating capabilities of deep neural networks as learning models. Then, we employ low-discrepancy sequences as sampling designs for the simulation runs required to create the training set for the deep surrogate model. This kind of sampling guarantees deterministically a good covering of the input space, and is able to exploit possible regularities of the simulation outcome. Extensive experimental tests are presented involving the popular SUMO microsimulator, showing the advantages of the proposed surrogate modeling solution under various performance measures.

An important task in unsupervised learning is the detection of communities in networks. Although many community detection techniques have been proposed, there are still some challenge problems, such as unbalanced community detection and the low efficiency. In this paper, we propose a community detection technique combining the sequential signal propagation of the Particle Competition model and the parallel propagation inspired by Self-Orgnizing Map (SOM). As a result, the model presents two salient features: 1) It can detect unbalanced communities. 2) It is much more efficient than the original particle competition model due to the introduction of parallel propagation. Still in this work, we analyze functional brain network by identifying the modules (communities) using the proposed technique. Our results show that there is a strong correlation between brain functions and brain regions and a big decrease of intra-strength measure among communities from the Control Network to the Schizophrenia Network, indicating that the functional correlation of brain regions is weakened in the disease network.

In this paper, we propose an unsupervised approach for knowledge graph (KG) creation from conversational data. We make use of intent classification and slot-filling, the two important components of any dialogue agent, exploit their interconnectedness, and finally construct a KG. We build a supervised intent classifier to extract the intent classes, and then on top of this we run our occlusion based slot-information extraction algorithm. Our algorithm is able to make use of supervised training of intent classifiers for extracting the relevant slot-information in an unsupervised way. To test the effectiveness of our system, we perform both automatic and manual evaluation of our intent-classifier and slot-filling system on three dialog datasets. Finally, we construct a knowledge graph from the dialogue conversation using an algorithm that makes use of our occlusion based slot-information extraction module. Empirical evaluation shows that our occlusion based method is able to successfully extract slot information from conversations, resulting in a high-quality KG.

Deep Neural Networks (DNNs) suffer from a rapid decrease in performance when trained on a sequence of tasks where only data of the most recent task is available. This phenomenon, known as catastrophic forgetting, prevents DNNs from accumulating knowledge over time. Overcoming catastrophic forgetting and enabling continual learning is of great interest since it would enable the application of DNNs in settings where unrestricted access to all the training data at any time is not always possible, e.g. due to storage limitations or legal issues. While many recently proposed methods for continual learning use some training examples for rehearsal, their performance strongly depends on the number of stored examples. In order to improve performance of rehearsal for continual learning, especially for a small number of stored examples, we propose a novel way of learning a small set of synthetic examples which capture the essence of a complete dataset. Instead of directly learning these synthetic examples, we learn a weighted combination of shared components for each example that enables a significant increase in memory efficiency. We demonstrate the performance of our method on commonly used datasets and compare it to recently proposed related methods and baselines., Comment: Paper accepted for publication at IJCNN2021

Human activity recognition has a wide range of application prospects and research significance in intelligent monitoring, assisted driving and human-computer interaction, such as intelligent monitoring of the elderly living alone, warning of dangerous behaviors of drivers and development of somatosensory games. Traditionally, human activity recognition is realized by cameras or wearable devices. However, in privacy-sensitive areas such as wards and cars, users may not be willing to share too many private videos. In this paper, we use millimeterwave radar to collect point clouds of human activities, design a novel graph neural network MMPoint-GNN with dynamic edges for the first time to process sparse point clouds, and combine it with Bidirectional LSTM to build a human activity recognition framework. We transform the logic operation into a differentiable function by edge selection network, and achieve the dynamic edge selection in MMPoint-GNN. Finally, we evaluate our method by comparing it with other methods on MMActivity dataset and MMGesture dataset. The results show that MMPoint-GNN outperforms all other baselines. The code is available at https://github.com/gongpx20069/mmRadar_for_HAR_VS

Neural network robustness has been a hot research topic since the discovery of adversarial examples. With the increasing threats posed by adversarial examples in safety-critical AI applications, the research problem of how to boost the robustness of neural networks has drawn even more attention in recent years. In this paper, we argue that existing methods for neural network robustness optimization primarily focus on increasing the minimum adversarial perturbation of individual datum while neglecting the purpose of the target machine learning task, i.e., classification. Consequently, we propose a neural network retraining method which implicitly enhances a model's capability in maximizing the minimum distance from data instances of all classes to the decision boundary. As demonstrated by the experiments, our method succeeds in improving the robustness of pre-trained neural networks within only a few retraining epochs while imposing little negative effect on classification accuracy, thus being applicable for online optimization of neural networks.

The next generation attackers often generate malware variants with Artificial Intelligence (AI) weapons, which are deliberately designed to evade antivirus engines. Security defenders propose many AI-based approaches to detect the massive number of malware variants. However, most AI-based malware detection approaches only output a label to users, and these labels are mainly unexplainable. The lack of transparency has introduced many black-box attacks. Malware developers can develop adversarial examples to evade these AI-based malware detection systems. In this paper, we propose an effective approach for malware detection and explanation, which can locate malicious code snippets by explaining the malware classifier decision result. To this end, firstly, we get the system call number sequence of the target sample with instrumentation tools in an elaborated sandbox. Secondly, we feed the mapped system call number sequence into a deep learning model to make a decision on whether the target sample is benign or malicious. Thirdly, we adopt the Layer-wise Relevance Propagation algorithm to find which slice of a sequence makes the greatest contribution in the decision. Our evaluation demonstrates that our approach achieves high classification accuracy (97.39%), reduces the neural network size by 20 times, and saves the malware analyst time to locate malicious code snippets.

In task-oriented dialogue systems, dialogue state tracking (DST) is responsible for estimating the current belief state of a dialogue. Recent research tends to utilize historical information to predict the states. However, most of them lack an efficient attention mechanism to comprehend the utterances well. Besides, existing methods usually ignore the relevance between the current utterances with the earlier ones, which determines whether or not the states change. In this paper, we propose an HDST-MRC (Hierarchical Dialogue State Tracking with Machine Reading Comprehension) model to tackle these issues. Within HSDT-MRC, we introduce bi-directional attention flow to extract a context span as the evidence of the ground truth and leverage another copy-augmented generator to predict the states. Experimental results on MultiWoz 2.0 and MultiWoz 2.1 demonstrate that our model achieves significant improvement compared with the baselines.

Social media has became a critical manner for people to acquire information in daily life. Despite the great convenience, fake news can be widely spread through social networks, causing various adverse effects on people's lives. Detecting these fake news or misinformations has proved to be a critical task and draws attentions from both governments and individuals. Recently, many methods have been proposed to solve this problem, but most of them rely on the body content of the news, ignoring the social context information such as the comments. We argue that the comments of a specific news contain common judgements of the whole society and could be extremely useful for detecting fake news. In this paper, we propose a new method TRANSFAKE which jointly models the body content and comments of news systemically, and detects fake news with multi-task learning framework. TRANSFAKE model is a Transformer-based model. It takes different modalities as input and employs multiple tasks, i.e. rumor score prediction and event classification, as intermediate tasks for extracting useful hidden relationships across various modalities. These intermediate tasks promote each other and encourage TRANSFAKE making the right decision. Extensive experiments on two standard real-life datasets demonstrate that TRANSFAKE outperforms state-of-the-art methods. It improves the detection accuracy by margins as large as ~12.6% and F1 scores as large as ~15%.

In this paper, we consider the embedding problem of Chinese new word with respect to its conceptual definition or description, which is especially important for understanding specialty documents. We present a two-stage model to learn the Chinese new word embedding, where the first encodes the information of character components and context, and the second aggregates the semantics of multiple texts. We perform extensive experiments to verify the proposed method and the results outperform the state of art methods on both direct semantics verification and advanced NLP tasks. Comparing with previous methods that require a corpus or an elaborately designed dataset for learning a new word embedding, our method requires only a few pieces of text and supports the evolution of meanings. We also experimentally verify the effects of different parts of model, the number and types of conceptual texts. Finally, we present some biology texts to illustrate whether the specialty semantics are encoded in the word embedding.

Protein-protein interaction (PPI) plays essential roles in nearly all biological processes of living organisms. Identification of PPI helps in understanding the cellular pathways and complex structure of proteins. Most of the works on the prediction of PPI utilized one type of information, mainly sequence-based. With the recent development in deep learning technologies, capturing the diverse and sparse biological dataset's relevant features is possible. This paper proposes a framework that incorporates a multi-modal representation of proteins to predict the interaction between them. Current work utilizes 3D structure and Gene ontology (GO) information to generate the vector representations of proteins. The task of PPI is divided into two phases: feature generation and prediction. We use deep learning algorithms in both stages. We validate our approach on two datasets: Human and S. cerevisiae. The trained model achieves accuracy values of 97.94% and 95.33% on the human and S. cerevisiae test sets, respectively. The results obtained over these datasets illustrate the superiority of the proposed method as compared to state-of-the-art methods.

With the rapid growth of multimodal data in social medias and the huge requirement of short but abundant information. Multimodal summarization has drawn much attention in both industry and academia. It usually obtains textual summary from multiple sources by computer vision or nature language processing technologies. However, there are also two challenges in modeling such task: 1) The feature representation is limited by the non-alignment among multimodal data; 2) Massive parallel data is required during training, which is time-consuming and laborious. In this paper, we introduce an unsupervised architecture (Compress-then-Reconstruct, CtnR) to generate the summary in an end-to-end manner and a Cross-Modal Transformer module (CMTrans) to fuse the multimodal non-alignment information. Comprehensive experiments show that the proposed CtnR framework with CMTrans outperforms mainstream unsupervised approaches in terms of BLEU, ROUGE and relevance scores on MSMO and Youtube News dataset, which increase 8.82% and 11.01% on average respectively.

Recently, the character-based model that incorporates potential word information has proven effective for Chinese named entity recognition (NER). However, due to the independence of the pre-trained character model and the lexicon, it will cause the embedding space to be misaligned and cannot be combined well. Chinese pre-trained encoders usually process text as characters. It ignores the information carried by the larger granular information, so the encoder cannot easily adapt to certain character combinations. Because large-grained information is ignored and Chinese does not have clear character boundaries, this will lead to the loss of important semantic information, which is an important problem for Chinese. In this paper, we propose PTWA: pre-training with word attention for Chinese named entity recognition. PTWA uses multi-head word attention to form a word vector from multiple word vectors, and proposes a word length prediction task to better integrate the word vector into pre-training. With the powerful capabilities of the transformer, PTWA can explicitly make full use of potential word information without adding an external lexicon, and can coexist with pre-trained models that implicitly use word information (such as BERT-WWM, and ERNIE). Experiments conducted on four Chinese NER datasets show that the performance of PTWA is better than other word-word models and Chinese pre-training models.

Rumor propagation is becoming easier and leads to severe consequences for society in several minutes or hours due to the rapid development of social networks. Thus, detecting rumors in early time is necessary and urgent for the community. In recent studies, machine learning approaches are widely applied in detecting rumors based on various features extracted from content, user characteristics, and propagation structure. Some studies have shown that features extracted from users are more valuable for rumor early detection. Whereas existing studies utterly utilize various user features, and all of them are deemed as equally, which ignore the inter-feature and temporal difference. Therefore, in this paper, we analyze the effectiveness of six frequently used user features with several representative deep learning models to learn more about such difference. And we propose a novel annealing attention model based on the analysis. The proposed model learns feature-attention and temporal-attention with multi-layer perceptron and parameterized annealing function to capture the difference and enhance the original user features in rumor early detection. Experimental results on the real-world dataset demonstrate that the proposed model detects rumors with an accuracy of 93.6% on Weibo in 15 minutes, which outperforms the state-of-art methods.

Time series classification (TSC) is widely used in various real-world applications such as human activity recognition, smart city governance, etc. Unfortunately, due to different reasons, only part of time series could be collected which may obviously degrade the performance of time series classifiers. To alleviate this problem, time series augmentation aims to generate synthetic time series by learning useful features from collected time series. As the popular generative model, generative adversarial networks (GAN) is regarded as a promising model for time series augmentation. However, applying GAN to the time series data suffers from a challenge in which the generated instances hold low quality but the model has gotten saturation. In this paper, for time series augmentation, we proposed TSA-GAN which is a robust GAN model with a self-adaptive recovering strategy to solve this problem. On 85 datasets of the UCR 2015 archive, our proposed TSA-GAN helps time series classifiers achieve performance improvements ranging from 8.3% to 12.5%, which is far better than the baseline.

Highly-reverberate underwater environments pose challenges for conventional localization techniques due to the highly non-linear nature of reflective surfaces, multi-path, and scattering fields. In this paper, we compare different machine learning methods for passive localization and tracking of single, non-stationary, underwater acoustic sources using multiple underwater acoustic vector sensors. We incorporate ordinal classification for localization in a novel approach to acoustic localization and compare the results with other standard methods. Realworld experiments demonstrate that both categorical and ordinal classification using deep LSTM networks significantly reduce localization error.

Data stream mining is an essential task in today's scientific community. It allows machine learning models to be updated over time as new data becomes available. Three pillars should be accounted for when selecting an appropriate algorithm for data stream mining: accuracy, processing time, and memory consumption. To develop and assess machine learning models in streaming scenarios, different tools have been developed, where the Massive Online Analysis, written in Java, and scikit-multiflow, written in Python, are in the spotlight. Despite the ease of use of both tools, neither are focused on performance, which puts in jeopardy the usage of the computational resources. In this paper, we show that with the right tools, Python libraries reach performance comparable to C/C++. More specifically, we show how optimized implementations in scikit-multiflow using low-level languages, i.e., C++, C++ with Intel Intrinsics, and Rust; with bindings to Python vastly overcome existing tools in computational resources usage while keeping predictive performance intact.

Authorship attribution identifies the true author of an unknown document. Authorship attribution plays a crucial role in plagiarism detection and blackmailer identification, however, the existing studies on authorship attribution in Bengali are limited. In this paper, we propose an instance-based deep authorship attribution model, called DAAB, to identify authors in Bengali. Our DAAB model fuses features from convolutional neural networks and another set of features from an artificial neural network to learn the stylometry of an author for authorship attribution. Extensive experiments with three real benchmark datasets such as Bengali-Quora and two online Bengali Corpus demonstrate the superiority of our authorship attribution model.

Emotion analysis from texts has emerged as an important area of research in the field of Natural Language Processing (NLP) in the past few years. Several benchmark datasets have been released for this task, however most of the datasets are open domain in nature and for resource-rich language like English. These datasets may not always be sufficient for capturing domain specific emotions, and may tend to skew towards an emotion based on the domain. In this paper, we provide a framework for multilingual emotion detection to deal with the crisis situations. We collect and annotate disaster domain social media tweets and news data in two languages, namely English and Hindi. We derive 6 emotions from Plutchik's wheel of emotions suitable for disaster domain, and annotate the data using these disaster specific emotions. In total we create four emotionally enriched datasets i.e. 2 tweets datasets (English and Hindi) and 2 news dataset (English and Hindi). We also establish strong baselines on the dataset using two popular deep learning algorithms, stacked Bi-LSTM+CNN and BERT. Evaluation shows that the best model achieves the averaged accuracy of 68% across the four different datasets.

Emotion recognition in conversation (ERC) is a challenging task due to the complexity of emotions and dynamics in dialogues. Current studies for emotion recognition mostly focus on the modeling of a single utterance in dialogue, which neglects self and inter-speaker influence. This paper presents a contextual attention neural network based on the multimodal framework that leverages the conversational information from both target and the other speaker for utterance-level emotion detection. Specifically, we utilize recurrent neural networks based on contextual attention for modeling the transaction and dependence between speakers. Further, the feature fusion is proposed to unite the important modal information extracted from multiple modalities, including audio, text and video, hence providing more useful and comprehensive knowledge for emotion recognition. The proposed approach shows its superiority in extracting contexts for self and inter-speaker influence and synthesizing them as global features that are beneficial to detect individual emotion state. Experiment result on the IEMOCAP corpus reports an accuracy of 64.6%, demonstrating the superiority of the proposed method in emotion recognition comparing to the state-of-the-arts.

Feature selection for deep learning architectures is one of the important and challenging steps in developing an efficient image parsing application. In this paper, a novel image parsing architecture which makes use of unique feature selection is proposed. It introduces the idea of weighted relationship awareness to reduce the redundancy of features and optimally select an efficient subset of feature representations. The proposed architecture is evaluated on Cam Vid benchmark dataset. A comparison with state-of-the-art methods was conducted which showed significant improvements in terms of segmentation and classification accuracy.

Image-level weakly supervised semantic segmentation (WSSS) reduces the cost of semantic segmentation significantly as only category labels are required. In the WSSS pipeline, the initially obtained seed areas are imprecise and should be refined. AffinityNet is a widely used seed area refinement method which learns pixel-level affinity between coordinates to adjust the object regions. However, the capabilities of AffinityNet are limited due to the complex simultaneous exploration of the affinities of all object classes, as the classes have different characteristics. This paper proposes a class-specific AffinityNet (CSANet) in which the affinities of each class are learnt separately by the class-specific affinity layers. Moreover, as the learning task is simple because the affinity of only one class is focused, the information contained by the uncertain neutral regions is extracted to provide additional information in the training of the class-specific affinity layers. The empirical results demonstrate the effectiveness of our method, which captures more precise boundaries and significantly improves segmentation results on the PASCAL VOC 2012 segmentation benchmark.