Search

Convolutional neural networks (CNNs) have shown their great power in multiple computer vision tasks. However, many recent works improve their performance by adding more layers and parameters, which lead to computational redundancy in many application scenarios, making it harder to implement on low-end devices. To solve this problem, model pruning methods are proposed, which aim to lower the computational and memory requirements of CNNs. In this paper, we propose ScoringNet, a neural network (NN) based pruning criteria for structured pruning procedure. ScoringNet generates a set of scores for each output channel in a model, which is used to reconstruct a pruned model later in a structured pruning way. ScoringNet can also use the gradient information to generate better scores, making the pruned model perform better. By using NNs, there are fewer hyperparameters, making it easier to implement. Experiment results demonstrate that the proposed ScoringNet can outperform or achieve competitive results compared to many state-of-the-art methods in both postpruning and pruning-at-initialization setups.

Non-maximum suppression (NMS) is an indispensable post-processing step in object detection. After a picture is detected by the target, a large number of redundant anchors will be obtained, so we need to go through NMS to process these repeated anchors. Its essence is to extract the target detection frame with high confidence and suppress the false detection frame with low confidence. The biggest problem in the NMS algorithm is that it completely removes adjacent low-confidence detection frames. In some images with dense targets, this is very likely to cause missed detection and false detection. Therefore, we proposed the Simple-NMS algorithm, which adds two new thresholds to the original NMS, and sets a constraint condition different from the original based on the new threshold, and does not change the complexity of the traditional NMS algorithm, which can be very good Improve the effectiveness of NMS missed targets. The new NMS algorithm is improved on the standard data sets PASAL VOC2007 and MS COCO. In addition, the algorithm is simple and efficient, and can be easily integrated into any other object detection process.

As the amount of user data increase, the computer performance and I/O speed required for data processing and analysis are getting higher and higher. Distributed file system has become the primary option for big data storage and query. According to the characteristics of high dimensionality and sparseness of data, this paper uses the distributed storage idea of CMD (coordinate modulo distribution) to store data in blocks. We only need to use cheap storage devices to form a distributed storage system, which solves the problem of big data disk I/O read performance to a certain extent. We have improved range query function under the CMD storage method; at the same time, the optimized B+ tree index technology has been used to solve the precise search problem of sparse data. Finally, in view of the unbalanced distribution of different sub-node data, we propose a new data rebalancing method on the CMD storage method.

Nowadays, the scene text detection approaches based on deep learning have already achieved promising performances. However, due to the insufficient use of the intermediate feature layer information of the neural network, and the learning of the indistinguishable scale and hard examples, it is difficult to improve the accuracy and the recall. Therefore, based on the Fully Convolutional Networks (FCNs), we propose a hierarchical scene text detector concerning hard examples. Specifically, our proposed detector HST-DHE directly predicts text lines of arbitrary orientations in full images without pre-defined anchor boxes. And the feature pyramid is constructed to predict the features at different levels to make full use of the feature information in the middle layers. At the same time, the loss function is redesigned to make the network focus on hard examples and further improve the precision of text detection. Experiments on standard datasets including ICDAR 2013, ICDAR 2015 and MSRA-TD500 demonstrate that the proposed algorithm significantly outperforms state-of-the-art detectors. According to the results, our proposed method has stronger robustness in multi-scale and multi-direction natural scene images.

Ranging from Matrix Factorization to deep learning based methods, recommender systems usually obtain user’s (or item’s) embedding via mapping from existed features. However, the collaborative signal is not encoded in the embedding process. Graph Neural Networks (GNNs) have been proved to be powerful in learning on graph data. But building recommender systems on GNNs faces challenges: the user-item graph encodes both interactions and opinions; user item interactions have heterogeneous strengths. To address these challenges, we proposed a novel framework on collaborative filtering recommendation via attention mechanism (GCFA), which models heterogeneous user-item interaction strengths and exploits the user-item graph structure by propagating embeddings on it. Extensive experiments on two real-world datasets demonstrate the effectiveness of the proposed framework GCFA.

Aiming at the problem of moving objects occluded in the field of automatic driving, in this paper, we propose a framework for occlusion relationship recognition based on distance measurement of the moving objects in video. The framework performs moving objects detection by YOLOv3 and occlusion recognition of moving objects by the information entropy. When the occlusion occurs, the framework will obtain the distance between the moving objects and the camera by binocular stereo vision distance measurement. Finally, the framework completes occlusion relationship recognition using a principle derived from this paper. The experimental results show that distance measurement of moving objects in video is effective and feasible for occlusion relationship recognition, which can provide some help for visual technology in the field of automatic driving.

With the development of big data, emphasis on data privacy and security has become a global trend. To address data privacy and security issues, Google has proposed a federal learning framework designed to make effective and accurate cross-organizational use of data within the industry. Suppose two organizations, A and B, each maintaining a private record of different sets of features for a common entity. If there is no coupling between the features, the processing is relatively simple. Assuming strong coupling between features, how to deal with it is relatively complicated, both to ensure privacy and to obtain coupling relationships, such as inner product calculation results. We designed and implemented a framework that guarantees privacy and efficient coupling. The experimental results show that the scheme is feasible, the data privacy is guaranteed, the feature coupling between different organizations is calculated, the result is accurate, and the execution time is acceptable.

Feature selection is an important part of data preprocessing. Selecting effective feature subsets can effectively reduce feature redundancy and reduce irrelevant features, and reduce training costs. Based on the theory of feature clusters, this paper proposes a feature selection strategy based on the graph structure. Considering a feature as a node in the graph, using the idea of graph message propagation, integrating the first-order neighbor information of each node, and then selecting the key point of the local maximum score as the selected feature, this can effectively reduce the feature redundancy and reduce features that are not related to the label. Finally, in order to verify the anti-interference of this novel method, the noise dimension was added in the UCI data set, and the comparison test was again performed. The experimental results show that the proposed algorithm can effectively improve the classification accuracy in a specific data set, and the anti-interference is better than other feature selection algorithms.

The deployment of nodes in 3D wireless sensor networks has always been the focus of research. In this paper, two algorithms for achieving full coverage are proposed, which are applicable to the even distribution and the uneven distribution of target points. Among them, the characteristics of the spherical and regular tetrahedrons are rationally utilized to achieve the purpose of covering more space with the same nodes. For the three-dimensional space with uneven target distribution, it is proposed to use the DBSCAN clustering method for target point processing, and then perform the same processing in each cluster to avoid wasting more sensor nodes and increasing energy consumption. Finally, considering the problem of energy saving, some nodes are put to sleep and waiting for the next call to reduce energy consumption and extend the working time of the network. Simulation experiments show that the algorithm has improved the number of nodes used and energy consumption.

In this paper, an algorithm of real-time eye blink monitoring using ordinary camera is proposed. This algorithm is based on pixel fluctuation ratio (FRP) of eye region image to detect eye blink. Position the eye area according to the facial landmark algorithm and then convert the image to a grayscale image. According to the point detected by the eye region in the grayscale image, the eye region is divided into three segments from the corner of the eye to the end of the eye. Calculate the value of the pixels in each segment, and find the partial image pixel value fluctuation ratio (FRP), and train the selected training samples with the support vector machine (SVM). We tested it on two data sets, and the blink detection algorithm achieved good results on both data sets.

With the development of the Human Genome Project, more and more biological sequence data are generated, and the analysis and processing of these sequence data have promoted the development of bioinformatics. Sequence similarity analysis is the basis of bioinformatics, through which we can use the information of known sequences to study the structure, function and evolutionary relationship of unknown new sequences. This paper performs data compression and retrieval on the genome database based on the dbSNP information of DNA. According to the rule of determining a protein by three bases, the amino acid characters are determined, and the redundant information is removed by using the dbSNP information. It is the first time to propose the construction of a new compressed form of biological sequence structure, which can reflect the strong correlation between the SNP location information and SNP in each sample in the genome. Finally, this paper constructs a complete biological sequence approximate neighbor query system, which can not only greatly reduce the storage and computing overhead, but also improve the query efficiency under the condition of ensuring the retrieval accuracy. The accuracy and scalability of this method are verified by experiments on a large data set of gene database.

In scientific research, bioinformatics, Internet applications, e-commerce and many other application fields, the amount of data is growing at an extremely fast rate. To analyze and utilize these huge data resources, it is necessary to rely on effective data analysis technology. Big data application is a process of mining effective information from big data by using data analysis methods, providing auxiliary decisions for users and realizing the value of big data. In the process of data calculation and analysis, we will often find that the data from the same source will show multiple function images in the same coordinate system, which will make the same variable correspond to multiple values in data prediction analysis. We call it multivalued function here. This paper provides a method to identify multivalued functions. By using the maximum information coefficient (MIC) theory proposed by David n. Reshef, and using the data sampling method to calculate the function identifiability of the data set, then, the spectral clustering method is used for recognition, segment and mark the images of different functions. Finally, the regression function equations with different marks are obtained by Gauss-newton iteration method. The results can be used for data prediction and analysis to assist decision-makers to make reasonable judgments.

Nowadays, the face recognition direction of unobstructed has been greatly developed, but there is still a huge deficiency in face recognition with occlusion. In order to improve the recognition rate under occlusion, this paper proposes a global and local occlusion recognition algorithm using deep learning for feature learning. Due to occlusion, local significant features will be missing, the impact will cause the accuracy of global recognition to decrease. Therefore, SIFT and SVM algorithms are used to discriminate partial face occlusion, and cut off the occluded local face to form a global face. Global facial features and unmasked local feature are extracted together. In the feature extraction, the LeNet-5 lightweight convolutional neural network and the improved LeNet-5 lightweight convolutional neural network are combined with SoftMax multi-classification. The final classification results use the improved hierarchical voting method to make final decisions on the face, which speeds up the training and prediction speed of the multi-classifier. This experiment will perform occlusion face verification on the AR dataset. The final experimental results show that the method greatly improves the accuracy of occlusion face recognition and has strong robustness to occlusion.

Data integration provides users a uniform interface for multiple heterogonous data sources. This problem has attracted a large amount of attention from both research and industry areas. In this paper, we overview the state-of-art approaches in data integration which are roughly divided into five parts: schema matching, entity resolution, data fusion, integration system, and new problems arisen.

To recognize accessible regions successfully and accurately is an important part of mobile robot navigation in outdoor environment. This thesis studies that the modified Mean Shift algorithm is applied for region partition on outdoor images obtained by robot vision. Then all parts of these pictures are recognized through the combination BOW model and the Opponent SIFT feature. And finally the results of region recognition are optimized through SVM and the way of using MS-OS-BOW (Mean Shift-Opponent SIFT-Bag Of Words) model to solve the problem of outdoor image region recognition is proposed. Specifically, outdoor working robots need to analyze received visual images to search for accessible path. For the outdoor images with broad angle of view and simple layout, whether it is accessible path or not can be decided by the semantics of this region. Therefore, the understanding of outdoor images can be finished by region recognition. These final experimental results demonstrate the effectiveness and practicability of outdoor image region recognition based on MS-OS-BOW model.

As location-based social network (LBSN) services become more popular in people’s lives, Point of Interest (POI) recommendation has become an important research topic.POI recommendation is to recommend places where users have not visited before. There are two problems in POI recommendation: sparsity and precision. Most users only check-in a few POIs in an LBSN. To tackle the sparse problem in a certain extent, we compute the similarity between the check-in datasets of different times. For the precision problem, we incorporate temporal information and geographical information. The temporal information will influence how the user chooses and allow the user to visit different distance point on different day. The geographical information is also used as a control for points which are too far away from the user’s check-in data. Our experimental results on real life LBSN datasets show that the proposed approach outperforms the other POI recommendation methods substantially.

Large-scale genome-wide association studies (GWAS) have become feasible recently because of the development of bead and chip technology. However, the success of GWAS partially depends on the statistical methods that are able to manage and analyze this sort of large-scale data. Currently, the commonly used tests for GWAS include the Cochran–Armitage trend test, the allelic χ2 test, the genotypic χ2 test, the haplotypic χ2 test, and the multi-marker genotypic χ2 test among others. From a methodological point of view, it is a great challenge to improve the power of commonly used tests, since these tests are commonly used precisely because they are already among the most powerful tests. In this article, we propose an improved score test that is uniformly more powerful than the score test based on the generalized linear model. Since the score test based on the generalized linear model includes the aforementioned commonly used tests as its special cases, our proposed improved score test is thus uniformly more powerful than these commonly used tests. We evaluate the performance of the improved score test by simulation studies and application to a real data set. Our results show that the power increases of the improved score test over the score test cannot be neglected in most cases. Genet. Epidemiol. 2011. © 2011 Wiley-Liss, Inc. 35:350-359, 2011

Summary In this paper, we propose a two-stage approach based on 17 biologically plausible models to search for two-locus combinations that have significant joint effects on the disease status in genome-wide association (GWA) studies. In the two-stage analyses, we only test two-locus joint effects of SNPs that show modest marginal effects. We use simulation studies to compare the power of our two-stage analysis with a single-marker analysis and a two-stage analysis by using a full model. We find that for most plausible interaction effects, our two-stage analysis can dramatically increase the power to identify two-locus joint effects compared to a single-marker analysis and a two-stage analysis based on the full model. We also compare two-stage methods with one-stage methods. Our simulation results indicate that two-stage methods are more powerful than one-stage methods. We applied our two-stage approach to a GWA study for identifying genetic factors that might be relevant in the pathogenesis of sporadic Amyotrophic Lateral Sclerosis (ALS). Our proposed two-stage approach found that two SNPs have significant joint effect on sporadic ALS while the single-marker analysis and the two-stage analysis based on the full model did not find any significant results.

Complex diseases are presumed to be the results of interactions of several genes and environmental factors, with each gene only having a small effect on the disease. Thus, the methods that can account for gene-gene interactions to search for a set of marker loci in different genes or across genome and to analyze these loci jointly are critical. In this article, we propose an ensemble learning approach (ELA) to detect a set of loci whose main and interaction effects jointly have a significant association with the trait. In the ELA, we first search for “base learners” and then combine the effects of the base learners by a linear model. Each base learner represents a main effect or an interaction effect. The result of the ELA is easy to interpret. When the ELA is applied to analyze a data set, we can get a final model, an overall P-value of the association test between the set of loci involved in the final model and the trait, and an importance measure for each base learner and each marker involved in the final model. The final model is a linear combination of some base learners. We know which base learner represents a main effect and which one represents an interaction effect. The importance measure of each base learner or marker can tell us the relative importance of the base learner or marker in the final model. We used intensive simulation studies as well as a real data set to evaluate the performance of the ELA. Our simulation studies demonstrated that the ELA is more powerful than the single-marker test in all the simulation scenarios. The ELA also outperformed the other three existing multi-locus methods in almost all cases. In an application to a large-scale case-control study for Type 2 diabetes, the ELA identified 11 single nucleotide polymorphisms that have a significant multi-locus effect (P-value=0.01), while none of the single nucleotide polymorphisms showed significant marginal effects and none of the two-locus combinations showed significant two-locus interaction effects. Genet. Epidemiol. © 2008 Wiley-Liss, Inc.

Background— Numerous experimental studies suggest that B-type natriuretic peptide (BNP) is cardioprotective; however, in clinical studies, higher plasma BNP concentrations have been associated with incident cardiovascular disease and higher left ventricular mass. Genetic association studies may allow us to determine the true causal directions without confounding by compensatory mechanisms. Methods and Results— We performed a meta-analysis of 2 genome-wide association results from a total of 2790 blacks. We assumed an additive genetic model in an association analysis of imputed 2.5 million single-nucleotide polymorphism dosages with residuals generated from multivariable-adjusted logarithmically transformed BNP controlling for relevant covariates and population stratification. Two loci were genome-wide significant, a candidate gene locus NPPB (rs198389, P =1.18×10 −09 ) and a novel missense variant in the KLKB1 locus (rs3733402, P =1.75×10 −11 ) that explained 0.4% and 1.9% of variation in log BNP concentration, respectively. The observed increase in BNP concentration was proportional to the number of effect allele copies, and an average of 8.1 pg/mL increase was associated with 2 allele copies. In a companion study, single-nucleotide polymorphisms in this loci were cross-checked with genome-wide association results for the aldosterone/renin ratio in individuals of European ancestry, and rs3733402 was genome-wide significant ( P −8 ), suggesting possible shared genetic architecture for these 2 pathways. Other statistically significant relations for these single-nucleotide polymorphisms included the following: rs198389 with systolic blood pressure in blacks (COGENT consortium) and rs198389 and rs3733402 with left ventricular mass in whites (EchoGEN consortium). Conclusions— These findings improve our knowledge of the genetic basis of BNP variation in blacks, demonstrate a possible shared allelic architecture for BNP with aldosterone-renin ratio, and motivate further studies of underlying mechanisms.

Motivation: Identifying microRNAs associated with diseases (disease miRNAs) is helpful for exploring the pathogenesis of diseases. Because miRNAs fulfill function via the regulation of their target genes and because the current number of experimentally validated targets is insufficient, some existing methods have inferred potential disease miRNAs based on the predicted targets. It is difficult for these methods to achieve excellent performance due to the high false-positive and false-negative rates for the target prediction results. Alternatively, several methods have constructed a network composed of miRNAs based on their associated diseases and have exploited the information within the network to predict the disease miRNAs. However, these methods have failed to take into account the prior information regarding the network nodes and the respective local topological structures of the different categories of nodes. Therefore, it is essential to develop a method that exploits the more useful information to predict reliable disease miRNA candidates. Results: miRNAs with similar functions are normally associated with similar diseases and vice versa. Therefore, the functional similarity between a pair of miRNAs is calculated based on their associated diseases to construct a miRNA network. We present a new prediction method based on random walk on the network. For the diseases with some known related miRNAs, the network nodes are divided into labeled nodes and unlabeled nodes, and the transition matrices are established for the two categories of nodes. Furthermore, different categories of nodes have different transition weights. In this way, the prior information of nodes can be completely exploited. Simultaneously, the various ranges of topologies around the different categories of nodes are integrated. In addition, how far the walker can go away from the labeled nodes is controlled by restarting the walking. This is helpful for relieving the negative effect of noisy data. For the diseases without any known related miRNAs, we extend the walking on a miRNA-disease bilayer network. During the prediction process, the similarity between diseases, the similarity between miRNAs, the known miRNA-disease associations and the topology information of the bilayer network are exploited. Moreover, the importance of information from different layers of network is considered. Our method achieves superior performance for 18 human diseases with AUC values ranging from 0.786 to 0.945. Moreover, case studies on breast neoplasms, lung neoplasms, prostatic neoplasms and 32 diseases further confirm the ability of our method to discover potential disease miRNAs. Availability and implementation: A web service for the prediction and analysis of disease miRNAs is available at http://bioinfolab.stx.hk/midp/ . Contact: guoyahong_hlju@163.com or lixia@hrbmu.edu.cn Supplementary information: Supplementary data are available at Bioinformatics online.

High blood pressure (BP) is more prevalent and contributes to more severe manifestations of cardiovascular disease (CVD) in African Americans than in any other United States ethnic group. Several small African-ancestry (AA) BP genome-wide association studies (GWASs) have been published, but their findings have failed to replicate to date. We report on a large AA BP GWAS meta-analysis that includes 29,378 individuals from 19 discovery cohorts and subsequent replication in additional samples of AA (n = 10,386), European ancestry (EA) (n = 69,395), and East Asian ancestry (n = 19,601). Five loci (EVX1-HOXA, ULK4, RSPO3, PLEKHG1, and SOX6) reached genome-wide significance (p < 1.0 × 10(-8)) for either systolic or diastolic BP in a transethnic meta-analysis after correction for multiple testing. Three of these BP loci (EVX1-HOXA, RSPO3, and PLEKHG1) lack previous associations with BP. We also identified one independent signal in a known BP locus (SOX6) and provide evidence for fine mapping in four additional validated BP loci. We also demonstrate that validated EA BP GWAS loci, considered jointly, show significant effects in AA samples. Consequently, these findings suggest that BP loci might have universal effects across studied populations, demonstrating that multiethnic samples are an essential component in identifying, fine mapping, and understanding their trait variability.