Abstract: Deep learning (DL), especially Convolutional neural networks (CNN), has gained wide popularity in various image processing tasks. With the significant achievements obtained in DL, it has provided many successful solutions for real-world applications as well as in medical domain. Automated retinal images analysis has been widely applied to screening Diabetic retinopathy (DR), which can greatly help preventing the occurrence of complete blindness when used in the early screening. In this paper, we mainly focus on DL, and we will give an overview of the deep learning-based methods for DR screening. Finally, we will discuss the main issues encountered in the DR screening systems.
Abstract: With the development of artificial intelligence, machine learning has been applied in more and more domains. In order to improve the quality and efficiency of software, automatic program generation is becoming a research hotspot. In recent years, machine learning has also been gradually applied in automatic program generation. Decision trees, language models, and cyclic neural networks have been applied in code generation, code completion and code knowledge mining. The efficiency of software development has been improved to a certain extent using machine learning. Aimed at the automatic program generation, this paper analyzes and summarizes the models of machine learning, the modifications involved in the models and the application effects. The research direction is discussed from the aspects of programmer behavior and automatic program generation of machine learning.
Abstract: Wireless local area network (WLAN) fingerprint-based localization has become the most attractive and popular approach for indoor localization. However, the primary concern for its practical implementation is the laborious manual effort of calibrating sufficient location-labeled fingerprints. The Semi-supervised extreme learning machine (SELM) performs well in reducing calibration effort. Traditional SELM methods only use Received signal strength (RSS) information to construct the neighbor graph and ignores location information, which helps recognizing prior information for manifold alignments. We propose Composite SELM (CSELM) method by using both RSS signals and location information to construct composite graph. Besides, the issue of unlabeled RSS data quality has not been solved. We propose a novel approach called Composite semisupervised extreme learning machine with unlabeled RSS Quality estimation (CSELM-QE) that takes into account the quality of unlabeled RSS data and combines the composite neighbor graph, which considers location information in the semi-supervised extreme learning machine. Experimental results show that the CSELM-QE could construct a precise localization model, reduce the calibration effort for radio map construction and improve localization accuracy. Our quality estimation method can be applied to other methods that need to retain high quality unlabeled Received signal strength data to improve model accuracy.
Abstract: Multi-modal sentiment analysis (MSA) is increasingly becoming a hotspot because it extends the conventional Sentiment analysis (SA) based on texts to multi-modal content which can provide richer affective information. However, compared with textbased sentiment analysis, multi-modal sentiment analysis has much more challenges, because the joint learning process on multi-modal data requires both fine-grained semantic matching and effective heterogeneous feature fusion. Existing approaches generally infer sentiment type from splicing features extracted from different modalities but neglect the strong semantic correlation among cooccurrence data of different modalities. To solve the challenges, a multi-level deep correlative network for multimodal sentiment analysis is proposed, which can reduce the semantic gap by analyzing simultaneously the middlelevel semantic features of images and the hierarchical deep correlations. First, the most relevant cross-modal feature representation is generated with Multi-modal Deep and discriminative correlation analysis (Multi-DDCA) while keeping those respective modal feature representations to be discriminative. Second, the high-level semantic outputs from multi-modal deep and discriminative correlation analysis are encoded into attention-correlation cross-modal feature representation through a co-attention-based multimodal correlation submodel, and then they are further merged by multi-layer neural network to train a sentiment classifier for predicting sentimental categories. Extensive experimental results on five datasets demonstrate the effectiveness of the designed approach, which outperforms several state-of-the-art fusion strategies for sentiment analysis.
Abstract: We are concerned with a kind of iterative method for computing the Moore-Penrose inverse, which can be considered as a discrete-time form of recurrent neural networks. We study the momentum learning scheme of the method and discuss its semi-convergence when computing the Moore-Penrose inverse of a rankdeficient matrix. We prove the semi-convergence for our new acceleration algorithm and obtain the optimal momentum factor which makes the fastest semi-convergence. Numerical tests demonstrate the effectiveness of our new acceleration algorithm.
Abstract: To solve the poor performance of the single-server polling system in high traffic and the complex analysis of the multi-server polling system, a synchronous double-server polling system is proposed, and its performance is analyzed using a Backpropagation (BP) neural network prediction algorithm. Experimental data are processed and analyzed, and a three-layer multiinput single-output BP network model is constructed to predict the performance of the polling system under different arrival rates of information packets. In the prediction stage, first, the data are processed and the average queue length under different information arrival rates is used to form a sequence. Subsequently, a multiinput single-output BP neural network is constructed for prediction. Experimental results show that the algorithm can accurately predict the performance of the double-server polling system, thereby facilitating research regarding polling systems.
Abstract: The cost of misclassifying a malware program as normal is often higher than that of misclassifying a normal program as malware. Therefore, how to improve the detection accuracy of malware programs is a very important problem. This paper proposes a deep learning malware program detection algorithm based on attention mechanism. Word2Vec model is used to map the Application programming interface (API) into word vectors, and all word vectors of each sample are arranged into a matrix with the same size. On this basis, residual network is used to extract features of samples. The features are input into the attention mechanism to learn the similarity between samples. Then, the features are weighted with the similarity to obtain the new features with better robustness. The new features and the original features are added element by element to obtain the sample features more suitable for classification. Finally, samples are classified by classifier. Experiments show that the classification effect of the proposed method is better than that of the traditional machine learning method.
Abstract: Deep neural networks (DNNs) show great performance in lots of applications. Convolutional neural network (CNN) is one of the classic DNNs, and various modified CNNs have been brought up, such as DenseNet, GoogleNet, ResNet, etc. For diverse tasks, a unique structure of CNN may show its advantage. However, how to design an effective CNN model for a practical task is a puzzle. In this paper, we model the architecture optimization of CNN as an optimization problem and design a Genetic network programming based Fast evolutionary learning (GNP-FEL) to optimize CNN. GNP-FEL contains three main ideas: First GNP is adopted to optimize CNN architecture and hyperparameters, which can build diverse network structures and make network parameters selfevolve; Second multi-objective optimization is designed by balancing both CNN model efficiency and structure compactness; Last a novel incremental training method is proposed to train offspring CNN models in GNP, which is capable of reducing time complexity sharply. Experiments have validated that GNP-FEL can quickly evolve a CNN classifier with a sufficiently compact architecture. And the classifier has a comparable classification effect to state-ofthe-art CNN model.
Abstract: This paper proposes an effective image inpainting method using an improved deep convolutional auto-encoder network. By analogy with exiting methods of image inpainting based on auto-decoders, inpainting methods using the deep convolutional auto-encoder networks are significantly more effective in capturing high-level features than classical methods based on exemplar. However, the inpainted regions would appear blurry and global inconsistency. To alleviate the fuzzy problem, we improved the network model by adding skip connections between mirrored layers in encoder and decoder stacks, so that the generative process of the inpainting area can directly use the low-level features information of the processing image. For making the inpainted result look both more plausible and consistent with its surrounding contexts, the model is trained with a combination of standard pixel-wise reconstruction loss and two adversarial losses which ensures pixel-accurate and local-global contents consistency. With extensive experimental on the ImageNet and Paris Streetview datasets, we demonstrate qualitatively and quantitatively that our approach performs better than state of the art.
Abstract: Quaternion kernel Fisher discriminant analysis (QKFDA) is proposed for feature level multimodal biometric recognition. In quaternion division ring, QKFDA extracts the most discriminative information from the quaternion fusion feature sets by maximizing the betweenclass variance while minimizing the within-class variance. A complete two-phases framework of QKFDA is developed: Quaternion kernel principal component analysis (QKPCA) plus Quaternion linear discriminant analysis(QLDA). Two experiments are designed: experiment I fuses four different features of face and plamprint, experiment II fuses three different features of face, plamprint and signature. The experimental results show that QKFDA is superior to both traditional feature fusion methods (series rule and weighted sum rule)and other quaternion feature fusion methods (QPCA, QFDA, QLPP and QKPCA).
Abstract: Anomaly detection refers to identify the true anomalies from a given data set. We present an ensemble anomaly detection method called Relative mass and half-space tree based forest (RMHSForest), which detect anomalies, including global and local anomalies, based on relative mass estimation and halfspace tree. Different from density or distance based measure, RMHSForest utilizes a novel relative mass estimation to improve the detection of local anomaly. Meanwhile, half-space tree based on augmented mass can estimate a mass distribution efficiently without density or distance calculations or clustering. Our empirical results show that RMHSForest outperforms the current popular anomaly detection algorithms in terms of AUC and processing time in the test data sets.
Abstract: We propose a novel indoor head detection network using dual-stream information and multi-attention that can be used for indoor crowd counting. To solve the problem of object scale diversity in indoor human head detection, especially the problem of smallscale human head, we propose a dual-stream information flow structure to enrich the positioning and category semantic information of small-scale objects. We propose a kind of structure of the channel-attention mechanism which is used to enhance the ability of the network to identify small-scale objects. Our method has achieved a recall rate of 0.91 and an F1 score of 0.92 on SCUT-HEAD, which achieves the state-of-art performance in the field of indoor crowd detection.
Abstract: Student performance prediction plays an important role in improving education quality. Noticing that students' exercise-answering processes exhibit different characteristics according to their different performance levels, this paper aims to mine the performance-related information from students' exercising logs and to explore the possibility of predicting students' performance using such process-characteristic information. A formal model of student-shared exercising processes and its discovery method from students' exercising logs are presented. Several similarity measures between students' individual exercising behavior and student-shared exercising processes are presented. A prediction method of students' performance level considering these similarity measures is explored based on classification algorithms. An experiment on real-life exercise-answering event logs shows the effectiveness of the proposed prediction method.
Abstract: A classifier trained on the label-rich source dataset tends to perform poorly on the unlabeled target dataset because of the distribution discrepancy across different datasets. Unsupervised domain adaptation aims to transfer knowledge from the labeled source dataset to the unlabeled target dataset to solve this problem. Most of the existing unsupervised domain adaptation methods only concentrate on learning domain-invariant features across different domains, but they neglect the discriminability of the learned features to satisfy the cluster assumption. In this paper, we propose Semantic pairwise centroid alignment (SPCA), which is a point-wise method to learn both domain-invariant and discriminative features for homogeneous unsupervised domain adaptation. SPCA utilizes a novel semantic centroid loss to reduce the intraclass distance in feature space by using source data and target High-confidence centroid points (HCCPs). Then a classifier trained on source features is expected to generalize well on target features. Extensive experiments on visual recognition tasks verify the effectiveness of the proposed SPCA and also demonstrate that both domaininvariant and discriminative features learned by SPCA can significantly boost the performance of homogeneous unsupervised domain adaptation.
Abstract: Deep neural networks (DNNs) have achieved state-of-the-art performance in a number of domains but suffer intensive complexity. Network quantization can effectively reduce computation and memory costs without changing network structure, facilitating the deployment of DNNs on mobile devices. While the existing methods can obtain good performance, low-bit quantization without time-consuming training or access to the full dataset is still a challenging problem. In this paper, we develop a novel method named Compressorbased non-uniform quantization (CNQ) method to achieve non-uniform quantization of DNNs with few unlabeled samples. Firstly, we present a compressor-based fast nonuniform quantization method, which can accomplish nonuniform quantization without iterations. Secondly, we propose to align the feature maps of the quantization model with the pre-trained model for accuracy recovery. Considering the property difference between different activation channels, we utilize the weighted-entropy perchannel to optimize the alignment loss. In the experiments, we evaluate the proposed method on image classification and object detection. Our results outperform the existing post-training quantization methods, which demonstrate the effectiveness of the proposed method.
Abstract: Graph convolution networks are extremely efficient on the graph-structure data, which both consider the graph and feature information. Most existing models mainly focus on redefining the complicated network structure, while ignoring the negative impact of lowquality input data during the aggregation process. This paper utilizes the revised Laplacian matrix to improve the performance of the original model in the preprocessing stage. The comprehensive experimental results testify that our proposed model performs significantly better than other off-the-shelf models with a lower computational complexity, which gains relatively higher accuracy and stability.
Abstract: Vertices in the same group tend to connect densely, and usually share common attributes. Groups of different sizes reflect the relations of vertices in different ranges, and also reflect the features of different orders of the network. In this work, we propose a novel network representation learning algorithm by introducing group features of vertices of different orders to learn more discriminative network representations, named as network representation learning algorithm using Hierarchical structure embedding (HSNR). HSNR algorithm firstly constructs hierarchical relations of network structures of different orders based on greedy algorithm and modularity. In order to introduce hierarchical features into the network representation learning model, HSNR algorithm then introduces the idea of multi-relational modeling from knowledge representation, and converts the hierarchical relations into the triplet form between vertices. Finally, HSNR proposes a joint learning model embedding vertex triplets into the network representations. The experimental results show that the HSRN algorithm presented has an excellent performance in network vertex classification task on three real-world datasets.
Abstract: Hand gesture recognition on the depth videos is a promising approach for automotive interfaces because it is less sensitive to light variation and more accurate than other traditional methods. However, video gestures recognition is still a challenging task since lots of interferences are induced by the uncorrelated gesture factors. Considering that if the displays are more relevant, the results will more accurate, so ResNext, a kind of compact and efficient neural network, is firstly used as feature extractor, then an improved weighted frame unification method is adopted to obtain the key frame samples, finally the Discriminant correlation analysis (DCA) is employed to fuse features for static data and dynamic data after conducting Feature embedding branch (FEB) on static data. The public dataset named Depth based gesture recognition database (DGRD) is used in this paper, but the dataset is a little small and the class distribution is largely imbalance, and we find the performance of ResNext degrades badly in the condition of imbalance problem although it achieves excellent result at sufficient training data. In order to conquer the disadvantages of limited dataset, a special loss function scheme combining the softmax loss and dice loss is proposed. Evaluation of the algorithm performances in comparison with other state-of-the-art methods indicates that the proposed method is more practical for gesture recognition and may be widely adopted by automotive interfaces.
Abstract: Portable document format (PDF) files are increasingly used to launch cyberattacks due to their popularity and increasing number of vulnerabilities. Many solutions have been developed to detect malicious files, but their accuracy decreases rapidly in face of new evasion techniques. We explore how to improve the robustness of classifiers for detecting adversarial attacks in PDF files. Content replacement and the n-gram are implemented to extract robust features using proposed guiding principles. In the two-stage machine learning model, the objects are divided based on their types, and the anomaly detection model is first trained for each type individually. The former detection results are organized into tree-like information structure and treated as inputs to convolutional neural network. Experimental results show that the accuracy of our classifier is nearly 100% and the robustness against evasive samples is excellent. The object features also enable the identification of different vulnerabilities exploited in malicious PDF files.
Abstract: In the training process of Neural networks (NNs), the selection of hyper-parameters is crucial, which determines the final training effect of the model. Among them, the Learning rate decay (LRD) can improve the learning speed and accuracy; the Weight decay (WD) improves the over-fitting to varying degrees. However, the decay methods still have problems such as hysteresis and stiffness of parameter adjustment, so that the final model will be inferior. Based on the Quantum contextuality (QC) theory, we propose a Quantum contextuality constraint (QCC) to constrain the weights of nodes in NNs to further improve the training effect. In the simplest classification model, we combine this constraint with different methods of LRD and WD to verify that QCC can further improve the training effect on the decay method. The performance of the experiments shows that QCC can significantly improve the convergence and accuracy of the model.
Abstract: We propose a novel expression from manifolds to define Convolutional neural network (CNN). The layered structure is proceeded by integration in limited space continuously, with weights adjusted including value and direction in neural manifolds. Status transfer functions are proposed to simulate the kernel dynamics as a control matrix. We theoretically analyze the stability and controllability of kernel-based CNNs, and verify our findings by numerical experiments.
Abstract: Aiming at the lack of effective quantitative model to support the analysis of terrorist attacks, a multilayer depth Neural network (NN) Graph convolutional networks (GCN) model (NNGCN) was put forward to realize the classification and early warning of terrorist attacks. The proposed model optimized the traditional GCN with the help of complex NN. The concept of link index was introduced into the NNGCN model. It is combined with the important information between event nodes. The information includes the similarity of events and link probability. Compared with the original unoptimized model, the improved model increased the classification accuracy of terrorist attacks. Because the model uses the node's feature information and the link relationship of graph structure, it can also warn the sudden terrorist attacks effectively.
Abstract: The existing Goodness of fit (GoF) test based spectrum sensing algorithms mostly use samples or energies as observations to make decisions, which can hardly achieve satisfactory performance especially when the Primary user (PU) signals are highly correlated. Meanwhile, the eigenvalue of covariance matrix can reflect signal correlations well. Motivated by this, we study the distribution of eigenvalue and propose an eigenvalue based GoF spectrum sensing algorithm. In the proposed scheme, we use the ratios of maximum to minimum eigenvalue as observations and thus it can bring performance improvements through capturing correlation of PU signals. We also provide the related theoretical analysis for the proposed method. Simulation results show that the proposed method overcomes the problem of noise uncertainty and achieves performance improvement over the classical samples-based GoF test.
Abstract: Virtual machine (VM) consolidation offers a promising approach to saving energy and improving resource utilization in cloud data centers. However, the aggressive consolidation of VMs may lead to Servicelevel agreement (SLA) violations, which are essential for data centers and their users. Therefore, it is essential to find a tradeoff between the reduction in SLA violation level and energy costs. In this paper, we improve and expand our Host state 3rd-order Markov chain (HS3MC) model proposed in our previous research comprehensively: we propose a new VM selection algorithm, an improved VM placement algorithm, and an HS3MC-based VM consolidation algorithm based on our improved HS3MC model for the SLA-aware and energy-efficient consolidation of VMs in cloud data centers. We evaluate our proposed algorithms on an extended Cloudsim simulator using the PlanetLab workload and a random workload. Specifically, we improve and expand the OpenStack Neat experimental platform, and evaluate the performance of our algorithms in a real cloud environment based on OpenStack. The experimental results show that our proposed model can significantly reduce the SLA violation rates while maintaining energy efficiency.