Search

Using a microelectromechanical system (MEMS) inertial measurement unit (IMU) array composed of multiple low-cost IMUs can reduce the measurement errors of inertial sensors, and improve its navigation performance. However, there is doubt about the benefit of the IMU array in the scene of GNSS/INS dynamic navigation . Therefore, to evaluate the navigation performance of the arrays, we developed four groups of IMU arrays, each containing 16 MEMS IMUs. Each IMU was accurately calibrated and compensated to improve the performance. The field experiments proved that the navigation accuracy of the IMU array improved by 3.4 times statistically over a single IMU, which is close to the theoretical limit, i.e., 4 times. Comparison data processing indicated that the individual IMU calibration reduced the horizontal position error of the array by 54% on average, which confirms that the precise calibration of each IMU, especially the cross-axis effect and mounting angles, is crucial to the array’s navigation performance. This research provides firm experimental support for the application of IMU arrays in the field of navigation.

Multispectral photoacoustic tomography enables the resolution of spectral components of a tissue or sample at high spatiotemporal resolution. With the availability of commercial instruments, the acquisition of data using this modality has become consistent and standardized. However, the analysis of such data is often hampered by opaque processing algorithms, which are challenging to verify and validate from a user perspective. Furthermore, such tools are inflexible, often locking users into a restricted set of processing motifs, which may not be able to accommodate the demands of diverse experiments. To address these needs, we have developed a Reconstruction, Analysis, and Filtering Toolbox to support the analysis of photoacoustic imaging data. The toolbox includes several algorithms to improve the overall quantification of photoacoustic imaging, including non-negative constraints and multispectral filters. We demonstrate various use cases, including dynamic imaging challenges and quantification of drug effect, and describe the ability of the toolbox to be parallelized on a high performance computing cluster.

In this study, we investigated a means to improve the robustness of deep network training on visual recognition tasks without sacrificing accuracy. The contribution of this work can reduce the dependence on model decay to gain a strong defense against malicious attacks, especially from adversarial samples. There are two major challenges in this study. First, the model defense capability should be strong and improved over the training stage. The other is that the degrading of the model performance must be minimized to ensure visual recognition performance. To tackle these challenges, we propose active dropblock (ActDB) by incorporating active learning into a dropblock. Dropblock effectively perturbs the feature maps, thus enhancing the invulnerability of gradient-based adversarial attacks. In addition, it selects an optimal perturbation solution to minimize the objective loss function, thereby reducing the model degradation. The proposed organic integration successfully solved the model robustness and accuracy simultaneously. We validated our approach using extensive experiments on various datasets. The results showed significant gains compared to state-of-the-art methods.

For visual tracking methods based on reinforcement learning, action space determines the ability of exploration, which is crucial to model robustness. However, most trackers adopted simple strategies with action space, which will suffer local optima problem. To address this issue, a novel reinforcement learning based tracker called AEVRNet is proposed with non-convex optimization and effective action space exploration. Firstly, inspired by combinatorial upper confidence bound, we design an adaptive exploration strategy leveraging temporal and spatial knowledge to enhance effective action exploration and jump out of local optima. Secondly, we define the tracking problem as a non-convex problem and incorporate non-convex optimization in stochastic variance reduced gradient as backward propagation of our model, which can converge faster with lower loss. Thirdly, different from existing reinforcement learning based trackers using classification method to train model, we define a regression based action-reward loss function, which is more sensitive to aspects of the target states, e.g., the width and height of the target to further improve robustness. Extensive experiments on six benchmark datasets demonstrate that our proposed AEVRNet achieves favorable performance against the state-of-the-art reinforcement learning based methods.

The walnut varieties in China are rich, the planting area in the country is wide, and the yield ranks at the forefront of the world. Walnut kernel is the most important application part of the walnut fruit. An in-depth study found that the by-products of walnut, such as green husk and walnut shell, also have great application potential and are cheap raw materials for the extraction of important medical ingredients and the production of industrial products. However, the by-products are often burned or discarded as waste during processing, which not only wastes resources but also causes environmental pollution. To realize the high value-added application of the walnut fruit, a deep processing of each part of the walnut should be considered. Preliminary processing is the key link before walnuts enter the field of intensive processing and consumption. The advanced level of the required technological equipment can help to determine the quality of the walnut products. The preliminary processing of walnuts in China is mainly divided into six steps: green husk removal, walnut drying, walnut size classification, walnut shell-breaking, walnut shell–kernel separation, and walnut kernel skin removal. This paper starts with a presentation of the importance of each link and the existing bottleneck. Then, the paper systematically discusses the analysis of the current situation and the development of devices required for each link. The working mechanism of each link type and its influence on the design of a corresponding device are summarized. On the basis of the corresponding working mechanism, this study classifies and summarizes the characteristics of the core mechanism of the devices for each preliminary process link; then, it evaluates and analyzes the existing typical mechanical devices according to their types. Finally, the influence rule of the various devices for each link in the preliminary processing is analyzed as a means of ensuring high-quality walnuts.

Classical decline methods, such as Arps yield decline curve analysis, have advantages of simple principles and convenient applications, and they are widely used for yield decline analysis. However, for carbonate reservoirs with high initial production, rapid decline, and large production fluctuations, with most wells having no stable production period, the adaptability of traditional decline methods is inadequate. Hence, there is an urgent need to develop a new decline analysis method. Although machine learning methods based on multiple regression and deep learning have been applied to unconventional oil reservoirs in recent years, their application effects have been unsatisfactory. For example, prediction errors based on multiple regression machine learning methods are relatively large, and deep learning sample requirements and the actual conditions of reservoir management do not match. In this study, a new equal probability gene expression programming (EP-GEP) method was developed to overcome the shortcomings of the conventional Arps decline model in the production decline analysis of carbonate reservoirs. Through model validation and comparative analysis of prediction effects, it was proven that the EP-GEP model exhibited good prediction accuracy, and the average relative error was significantly smaller than those of the traditional Arps model and existing machine learning methods. The successful application of the proposed method in the production decline analysis of carbonate reservoirs is expected to provide a new decline analysis tool for field reservoir engineers.

Visible light communication (VLC) is considered an important complementary technology for extremely high sixth-generation (6G) data transmission and has become part of a hybrid 6G indoor network architecture with an ultradense deployment of VLC access points (APs) that presents severe challenges to user mobility. An adaptive handover mechanism, which includes a seamless handover protocol and a selection algorithm optimized with a deep reinforcement learning (DRL) method, is proposed to overcome these challenges. Experimental simulation results reveal that the average downlink data rate with the proposed algorithm is up to 48% better than those with traditional RL algorithms and that this algorithm also outperforms the deep Q-network (DQN), Sarsa and Q-learning algorithms by 8%, 13% and 13%, respectively.

Siamese network is highly regarded in the visual object tracking filed because of its unique advantages of pairwise input and pairwise training. It can measure the similarity between two image patches, which coincides with the principle of the matching-based tracking algorithm. In this paper, a variant Siamese network based tracker is proposed to introduce attention module into traditional Siamese network, and relocate the object with some auxiliary relocation methods, when the proposed tracker runs under an untrusted state. Firstly, a novel attention shake layer is proposed to replace the max pooling layer in Siamese network. This layer could introduce and train two different kinds of attention modules at the same time, which means the proposed attention shake layer could also help to improve the expression power of Siamese network without increasing the depth of the network. Secondly, an auxiliary relocation branch is proposed to assist in object relocation and tracking. According to the prior assumptions of visual object tracking, some weights are involved in the auxiliary relocation branch, such as structure similarity weight, motion similarity weight, motion smoothness weight and object saliency weight. Thirdly, a novel response map based switch function is proposed to monitor the tracking process and control the effect of auxiliary relocation branch. Furthermore, in order to discuss the effect of pooling layer in Siamese network, 9 pooling and attention architectures are proposed and discussed in this paper. Some empirical results are shown in the experiment part. Comparing with the state-of-the-art trackers, the proposed tracker could achieve comparable performance in multiple benchmarks.

As the most important source of carbon emissions, the power industry has huge potential for carbon emission reduction.Based on the traditional dispatching model, this paper comprehensively considers power generation-side resources and low-carbon demand side resources, such as demand response(DR), electric vehicle(EV) and distributed generation(DG), to coordinate dispatching. Simultaneously carbon capture equipment is introduced to establish carbon capture unit which participates dispatching in power system.Basd on above, the cost of carbon trading is introduced into the low-carbon economic dispatching goals. And the low-carbon economic dispatching model including demand side low-carbon resources and carbon capture equipment is established under the carbon trading environment. A simulation example is used to analyze the impact of the introduction of demand side low-carbon resources and carbon capture equipment on the economic and low-carbon benefits of the power system. The results of the simulation example have verified the feasibility and superiority of the proposed model.

In the era of data explosion, a growing number of data-intensive computing frameworks, such as Apache Hadoop and Spark, have been proposed to handle the massive volume of unstructured data in parallel. Since programming models provided by these frameworks allow users to specify complex and diversified user-defined functions (UDFs) with predefined operations, the grand challenge of tuning up entire system performance arises if programmers do not fully understand the semantics of code, data, and runtime systems. In this paper, we design a holistic semantics-aware optimization for data-intensive applications using hybrid program analysis} (SODA) to assist programmers to tune performance issues. SODA is a two-phase framework: the offline phase is a static analysis that analyzes code and performance profiling data from the online phase of prior executions to generate a parameterized and instrumented application; the online phase is a dynamic analysis that keeps track of the application's execution and collects runtime information of data and system. Extensive experimental results on four real-world Spark applications show that SODA can gain up to 60%, 10%, 8%, faster than its original implementation, with the three proposed optimization strategies, i.e., cache management, operation reordering, and element pruning, respectively., Comment: 2021 IEEE International Conference on Cloud Computing

Detecting nodal activities in dynamic social networks has strategic importance in many applications, such as online marketing campaigns and homeland security surveillance. How peer-to-peer exchanges in social media can facilitate nodal activity detection is not well explored. Existing models assume network nodes to be static in time and do not adequately consider features from social theories. This research developed and validated two theory-based models, Random Interaction Model (RIM) and Preferential Interaction Model (PIM), to characterize temporal nodal activities in social media networks of human agents. The models capture the network characteristics of randomness and preferential interaction due to community size, human bias, declining connection cost, and rising reachability. The models were compared against three benchmark models (abbreviated as EAM, TAM, and DBMM) using a social media community consisting of 790,462 users who posted over 3,286,473 tweets and formed more than 3,055,797 links during 2013–2015. The experimental results show that both RIM and PIM outperformed EAM and TAM significantly in accuracy across different dates and time windows. Both PIM and RIM scored significantly smaller errors than DBMM did. Structural properties of social networks were found to provide a simple and yet accurate approach to predicting model performances. These results indicate the models’ strong capability of accounting for user interactions in real-world social media networks and temporal activity detection. The research should provide new approaches for temporal network activity detection, develop relevant new measures, and report new findings from large social media datasets.

Due to its speed and ease of use, Spark has become a popular tool amongst data scientists to analyze data in various sizes. Counter-intuitively, data processing workloads in industrial companies such as Google, Facebook, and Yahoo are dominated by short-running applications, which is due to the majority of applications being mostly consisted of simple SQL-like queries (Dean, 2004, Zaharia et al, 2008). Unfortunately, the current version of Spark is not optimized for such kinds of workloads. In this paper, we propose a novel framework, called Meteor, which can dramatically improve the performance for short-running applications. We extend Spark with three additional operating modes: one-thread, one-container, and distributed. The one-thread mode executes all tasks on just one thread; the one-container mode runs these tasks in one container by multi-threading; the distributed mode allocates all tasks over the whole cluster. A new framework for submitting applications is also designed, which utilizes a fine-grained Spark performance model to decide which of the three modes is the most efficient to invoke upon a new application submission. From our extensive experiments on Amazon EC2, one-thread mode is the optimal choice when the input size is small, otherwise the distributed mode is better. Overall, Meteor is up to 2 times faster than the original Spark for short applications.

There is a growing interest in designing models that can deal with images from different visual domains. If there exists a universal structure in different visual domains that can be captured via a common parameterization, then we can use a single model for all domains rather than one model per domain. A model aware of the relationships between different domains can also be trained to work on new domains with less resources. However, to identify the reusable structure in a model is not easy. In this paper, we propose a multi-domain learning architecture based on depthwise separable convolution. The proposed approach is based on the assumption that images from different domains share cross-channel correlations but have domain-specific spatial correlations. The proposed model is compact and has minimal overhead when being applied to new domains. Additionally, we introduce a gating mechanism to promote soft sharing between different domains. We evaluate our approach on Visual Decathlon Challenge, a benchmark for testing the ability of multi-domain models. The experiments show that our approach can achieve the highest score while only requiring 50% of the parameters compared with the state-of-the-art approaches.

Nonconvex and nonsmooth problems have recently attracted considerable attention in machine learning. However, developing efficient methods for the nonconvex and nonsmooth optimization problems with certain performance guarantee remains a challenge. Proximal coordinate descent (PCD) has been widely used for solving optimization problems, but the knowledge of PCD methods in the nonconvex setting is very limited. On the other hand, the asynchronous proximal coordinate descent (APCD) recently have received much attention in order to solve large-scale problems. However, the accelerated variants of APCD algorithms are rarely studied. In this paper, we extend APCD method to the accelerated algorithm (AAPCD) for nonsmooth and nonconvex problems that satisfies the sufficient descent property, by comparing between the function values at proximal update and a linear extrapolated point using a delay-aware momentum value. To the best of our knowledge, we are the first to provide stochastic and deterministic accelerated extension of APCD algorithms for general nonconvex and nonsmooth problems ensuring that for both bounded delays and unbounded delays every limit point is a critical point. By leveraging Kurdyka-Łojasiewicz property, we will show linear and sublinear convergence rates for the deterministic AAPCD with bounded delays. Numerical results demonstrate the practical efficiency of our algorithm in speed.

Recent advancements in recurrent neural network (RNN) research have demonstrated the superiority of utilizing multiscale structures in learning temporal representations of time series. Currently, most of multiscale RNNs use fixed scales, which do not comply with the nature of dynamical temporal patterns among sequences. In this paper, we propose Adaptively Scaled Recurrent Neural Networks (ASRNN), a simple but efficient way to handle this problem. Instead of using predefined scales, ASRNNs are able to learn and adjust scales based on different temporal contexts, making them more flexible in modeling multiscale patterns. Compared with other multiscale RNNs, ASRNNs are bestowed upon dynamical scaling capabilities with much simpler structures, and are easy to be integrated with various RNN cells. The experiments on multiple sequence modeling tasks indicate ASRNNs can efficiently adapt scales based on different sequence contexts and yield better performances than baselines without dynamical scaling abilities.

As big data processing is being widely adopted by many domains, massive amount of generated data become more reliant on the parallel computing platforms for analysis, wherein Spark is one of the most widely used frameworks. Spark’s abnormal tasks may cause significant performance degradation, and it is extremely challenging to detect and diagnose the root causes. To that end, we propose an innovative tool, named LADRA, for log-based abnormal tasks detection and root-cause analysis using Spark logs. In LADRA, a log parser first converts raw log files into structured data and extracts features. Then, a detection method is proposed to detect where and when abnormal tasks happen. In order to analyze root causes we further extract pre-defined factors based on these features. Finally, we leverage General Regression Neural Network (GRNN) to identify root causes for abnormal tasks. The likelihood of reported root causes are presented to users according to the weighted factors by GRNN. LADRA is an off-line tool that can accurately analyze abnormality without extra monitoring overhead. Four potential root causes, i.e., CPU, memory, network, and disk I/O, are considered. We have tested LADRA atop of three Spark benchmarks by injecting aforementioned root causes. Experimental results show that our proposed approach is more accurate in the root cause analysis than other existing methods.

Variational autoencoders are employed to provide a framework for learning deep latent state representation. Inverse autoregressive flow is a type of normalizing flow that is employed to provide strategies for flexible variational inferences of posteriors over latent variables. The study aimed to prove that the agent can find a solution faster and at a lower cost. The proposed architecture comprises three basic methods, whereby the first one initiates the parameters and other layers of the TensorFlow framework; the second one is the build method that develops a layer using the Kera Library, and the last method, transform, determines the next sequence in the chain and changes the input. The model was then tested on a car racing simulator from OpenAI Gym. It was concluded that the proposed model is fast because it achieved a score of 928 ± 14 over 100 random trials, which is the best in the tested environment.

This paper is concerned with ranking many pre-trained deep neural networks (DNNs), called checkpoints, for the transfer learning to a downstream task. Thanks to the broad use of DNNs, we may easily collect hundreds of checkpoints from various sources. Which of them transfers the best to our downstream task of interest? Striving to answer this question thoroughly, we establish a neural checkpoint ranking benchmark (NeuCRaB) and study some intuitive ranking measures. These measures are generic, applying to the checkpoints of different output types without knowing how the checkpoints are pre-trained on which dataset. They also incur low computation cost, making them practically meaningful. Our results suggest that the linear separability of the features extracted by the checkpoints is a strong indicator of transferability. We also arrive at a new ranking measure, NLEEP, which gives rise to the best performance in the experiments., Accepted to CVPR 2021

Transfer learning has become the de facto practice to reuse a deep neural network (DNN) that is pre-trained with abundant training data in a source task to improve the model training on target tasks with smaller-scale training data. In this paper, we first investigate the correlation between the DNN’s pre-training performance in the source task and their transfer results in the downstream tasks. We find that high performance of a pre-trained model does not necessarily imply high transferability. We then propose a metric, named Frechet Pre-train Distance, to estimate the transferability of a deep neural network. By applying the proposed Frechet Pre-train Distance, we are able to identify the optimal pre-trained checkpoint, and then achieve high transferability on downstream tasks. Finally, we investigate several factors impacting DNN’s transferability including normalization, different networks and learning rates. The results consistently support our conclusions.

Functional gradient materials (FGMs), as a modern group of materials, can provide multiple functions and are able to well mimic the hierarchical and gradient structure of natural systems. Because biomedical implants usually substitute the bone tissues and bone is an organic, natural FGM material, it seems quite reasonable to use the FGM concept in these applications. These FGMs have numerous advantages, including the ability to tailor the desired mechanical and biological response by producing various gradations, such as composition, porosity, and size; mitigating some limitations, such as stress-shielding effects; improving osseointegration; and enhancing electrochemical behavior and wear resistance. Although these are beneficial aspects, there is still a notable lack of comprehensive guidelines and standards. This paper aims to comprehensively review the current scenery of FGM metallic materials in the biomedical field, specifically its dental and orthopedic applications. It also introduces various processing methods, especially additive manufacturing methods that have a substantial impact on FGM production, mentioning its prospects and how FGMs can change the direction of both industry and biomedicine. Any improvement in FGM knowledge and technology can lead to big steps toward its industrialization and most notably for much better implant designs with more biocompatibility and similarity to natural tissues that enhance the quality of life for human beings.

Ultraviolet is a vital spectrum in wireless optical communication. There are a lot of studies have been done in wireless ultraviolet communication. However, Ultraviolet communication networking mainly focuses on self-organizing networks and multi-hop networks. We proposed a new multi-access scheme of star topology network for confidential communications in semi-open spaces on the surface of ships. We verified the feasibility of the scheme through theoretical analysis and system simulation. It is proved that the scheme is suitable for ultraviolet communication and has better performance in a multi-user network. This scheme could be a sub-scheme for the next generation communication network combined with ultraviolet light communication.

With the continuous progress and development in the biomedicine field, metallic biomedical materials have attracted the considerable attention of researchers, but the related procedures need to be further developed. Since the traditional metal implant materials are not highly compatible with the human body, the modern materials with excellent mechanical properties and proper biocompatibility should be developed urgently in order to solve any adverse reactions caused by the long-term implantations. The advent of the high-entropy alloy (HEA) as an innovative and advanced idea emerged to develop the medical implant materials through the specific HEA designs. The properties of these HEA materials can be predicted and regulated. In this paper, the progression and application of titanium-based HEAs, as well as their preparation and biological evaluation methods, are comprehensively reviewed. Additionally, the prospects for the development and use of these alloys in implant applications are put forward.

Deep Neural Network (DNN) trained by the gradient descent method is known to be vulnerable to maliciously perturbed adversarial input, aka. adversarial attack. As one of the countermeasures against adversarial attacks, increasing the model capacity for DNN robustness enhancement was discussed and reported as an effective approach by many recent works. In this work, we show that shrinking the model size through proper weight pruning can even be helpful to improve the DNN robustness under adversarial attack. For obtaining a simultaneously robust and compact DNN model, we propose a multi-objective training method called Robust Sparse Regularization (RSR), through the fusion of various regularization techniques, including channel-wise noise injection, lasso weight penalty, and adversarial training. We conduct extensive experiments to show the effectiveness of RSR against popular white-box (i.e., PGD and FGSM) and black-box attacks. Thanks to RSR, 85 % weight connections of ResNet-18 can be pruned while still achieving 0.68 % and 8.72 % improvement in clean- and perturbed-data accuracy respectively on CIFAR-10 dataset, in comparison to its PGD adversarial training baseline.

Rapidly generated data and the amount magnitude of data analytical jobs pose great pressure to the underlying computing facilities. A distributed multi-cluster computing environment such as a hybrid cloud consequently raises its necessity due to its advantages in adapting geographically distributed and potentially cloud-based computing resources. Different clusters forming such an environment could be heterogeneous and may be resource-elastic as well. From analytical perspective, in accordance with increasing needs on streaming applications and timely analytical demands, many data analytical jobs nowadays are time-critical in terms of their temporal urgency. And the overall workload of the computing environment can be hybrid to contain both time-critical and general applications. These all call for an efficient resource management approach capable to apprehend both computing environment and application features. However, the added up complexity and high dynamics of the system greatly hinder the performance of traditional rule-based approaches. In this work, we propose to utilize deep reinforcement learning for developing elasticity-compatible resource management for a heterogeneous distributed computing environment, aiming for less occurrences of missing temporal deadline while maintaining low average execution time ratio. Along with reinforcement learning we design a deep model employing Long Short-Term Memory (LSTM) structure and partial model sharing for multi-target learning mechanism. The experimental results show that the proposed approach could greatly outperform baselines and serve as a robust resource management for variant workloads.

At present, by establishing complex mathematical model to simulate the ironmaking process in Blast Furnace (BF), the real-time and applicability are not strong in actual production. To solve this problem, an image reconstruction algorithm for BF Burden surface based on Compressive Sensing is proposed. Sparse matrix is constructed by Discrete Cosine Transform (DCT) and random Gaussian matrix is constructed to obtain part of the image information of BF Burden surface directly. Then the original image information is reconstructed by orthogonal matching pursuit algorithm. The final simulation results show that the image reconstruction algorithm proposed in this study works well.

Additive manufacturing (AM) is among the most attractive methods to produce implants, the processes are very swift and it can be precisely controlled to meet patient’s requirement since they can be produced in exact shape, dimension, and even texture of different living tissues. Until now, lots of methods have emerged and used in this field with diverse characteristics. This review aims to comprehensively discuss 3D printing (3DP) technologies to manufacture metallic implants, especially on techniques and procedures. Various technologies based on their main properties are categorized, the effecting parameters are introduced, andthe history of AM technology is briefly analyzed. Subsequently, the utilization of these AM-manufactured components in medicine along with their effectual variables is discussed, and special attention is paid on to the production of porous scaffolds, taking pore size, density, etc., into consideration. Finally, 3DP of the popular metallic systems in medical applications such as titanium, Ti6Al4V, cobalt-chromium alloys, and shape memory alloys are studied. In general, AM manufactured implants need to comply with important requirements such as biocompatibility, suitable mechanical properties (strength and elastic modulus), surface conditions, custom-built designs, fast production, etc. This review aims to introduce the AM technologies in implant applications and find new ways to design more sophisticated methods and compatible implants that mimic the desired tissue functions.

We study how to train a student deep neural network for visual recognition by distilling knowledge from a blackbox teacher model in a data-efficient manner. Progress on this problem can significantly reduce the dependence on large-scale datasets for learning high-performing visual recognition models. There are two major challenges. One is that the number of queries into the teacher model should be minimized to save computational and/or financial costs. The other is that the number of images used for the knowledge distillation should be small; otherwise, it violates our expectation of reducing the dependence on large-scale datasets. To tackle these challenges, we propose an approach that blends mixup and active learning. The former effectively augments the few unlabeled images by a big pool of synthetic images sampled from the convex hull of the original images, and the latter actively chooses from the pool hard examples for the student neural network and query their labels from the teacher model. We validate our approach with extensive experiments., Accepted to CVPR 2020

Object frequency in the real world often follows a power law, leading to a mismatch between datasets with long-tailed class distributions seen by a machine learning model and our expectation of the model to perform well on all classes. We analyze this mismatch from a domain adaptation point of view. First of all, we connect existing class-balanced methods for long-tailed classification to target shift, a well-studied scenario in domain adaptation. The connection reveals that these methods implicitly assume that the training data and test data share the same class-conditioned distribution, which does not hold in general and especially for the tail classes. While a head class could contain abundant and diverse training examples that well represent the expected data at inference time, the tail classes are often short of representative training data. To this end, we propose to augment the classic class-balanced learning by explicitly estimating the differences between the class-conditioned distributions with a meta-learning approach. We validate our approach with six benchmark datasets and three loss functions., Comment: Accepted for publication at CVPR2020

Speaker diarization, which is to find the speech segments of specific speakers, has been widely used in human-centered applications such as video conferences or human-computer interaction systems. In this paper, we propose a self-supervised audio-video synchronization learning method to address the problem of speaker diarization without massive labeling effort. We improve the previous approaches by introducing two new loss functions: the dynamic triplet loss and the multinomial loss. We test them on a real-world human-computer interaction system and the results show our best model yields a remarkable gain of +8% F 1 -scores as well as diarization error rate reduction. Finally, we introduce a new large scale audio-video corpus designed to fill the vacancy of audio-video dataset in Chinese.

We propose a new task towards more practical application for image generation - high-quality image synthesis from salient object layout. This new setting allows users to provide the layout of salient objects only (i.e., foreground bounding boxes and categories), and lets the model complete the drawing with an invented background and a matching foreground. Two main challenges spring from this new task: (i) how to generate fine-grained details and realistic textures without segmentation map input; and (ii) how to create a background and weave it seamlessly into standalone objects. To tackle this, we propose Background Hallucination Generative Adversarial Network (BachGAN), which first selects a set of segmentation maps from a large candidate pool via a background retrieval module, then encodes these candidate layouts via a background fusion module to hallucinate a suitable background for the given objects. By generating the hallucinated background representation dynamically, our model can synthesize high-resolution images with both photo-realistic foreground and integral background. Experiments on Cityscapes and ADE20K datasets demonstrate the advantage of BachGAN over existing methods, measured on both visual fidelity of generated images and visual alignment between output images and input layouts., Accepted to CVPR 2020

Dilated convolution kernels are constrained by their shared dilation, keeping them from being aware of diverse spatial contents at different locations. We address such limitations by formulating the dilation as trainable weights with respect to individual positions. We propose Adaptive Dilation Convolutional Neural Networks (ADCNN), a light-weighted extension that allows convolutional kernels to adjust their dilation value based on different contents at the pixel level. Unlike previous content-adaptive models, ADCNN dynamically infers pixel-wise dilation via modeling feed-forward inter-patterns, which provides a new perspective for developing adaptive network structures other than sampling kernel spaces. Our evaluation results indicate ADCNNs can be easily integrated into various backbone networks and consistently outperform their regular counterparts on various visual tasks.

MapReduce is an effective and widely-used framework for processing large datasets in parallel over a cluster of computers. Data skew, cluster heterogeneity, and network traffic are three issues that significantly affect the performance of MapReduce applications. However, the hash-based partitioner in the native Hadoop does not consider these factors. This paper proposes a new partitioner for Yarn (Hadoop 2.6.0), namely, NPIY, which adopts an innovative parallel sampling method to distribute intermediate data. The paper makes the following major contributions: (1) NPIY mitigates data skew in MapReduce applications; (2) NPIY considers the heterogeneity of computing resources to balance the loads among Reducers; (3) NPIY reduces the network traffic in the shuffle phase by trying to retain intermediate data on those nodes running both map and reduce tasks. Compared with the native Hadoop and other popular strategies, NPIY can reduce execution time by up to 41.66% and 58.68% in homogeneous and heterogeneous clusters, respectively. We further customize NPIY for parallel image processing, and the execution time has been improved by 28.8% compared with the native Hadoop.

Due to stochasticity and uncertainty of malicious Web services over the Internet, it becomes difficult to select reliable services while meeting non-functional requirements in service-oriented systems. To avoid the unreliable real-world process of obtaining services, this paper proposes a novel service selection approach via two-phase decisions for enhancing the reliability of service-oriented systems. In the first-phase decision, we define the problem of finding reliable service candidates as a multiple criteria decision making (MCDM) problem. Then, we construct a decision model to address the problem. In the second-phase decision, we define the problem of selecting services based on non-functional requirements as an optimization problem. Finally, we propose a convex hull based approach for solving the optimization problem. Large-scale and real-world experiments are conducted to show the advantages of the proposed approach. The evaluation results confirm that our approach achieves higher success rate and less computation time to guarantee the reliability when compared to the other state-of-the-art approaches.

Mobile agents environment is a new application paradigm with unique features such as mobility and autonomy. Traditional deadlock detection algorithms in distributed computing systems do not work well in mobile agent systems due to the unique feature property of the mobile agent. Existing deadlock detection and resolution algorithms in mobile agent systems have limitations such as performance inefficiency and duplicate detection/resolution when multiple mobile agents simultaneously detect/resolve the same deadlock. To address these problems, we propose an improved deadlock detection and resolution algorithm that adopts priority-based technique and lazy reaction strategy. The priority-based technique aims to ensure that there is only one instance of deadlock detection and resolution, and it also helps reduce mobile agent movement and data traffic together with the lazy reaction strategy. The liveness and safety properties of the proposed algorithm are proved in this paper. Theoretical analysis and experimental results show that the proposed algorithm provides better performance in terms of agent movement, data traffic, and execution time.

Deadlock detection in a distributed system without shared memory is important to ensure the reliability of the system. It becomes more complex when multiple deadlock detection algorithm instances execute concurrently in the system. In addition, the problem of communication disconnection between computing nodes or processes makes deadlock detection more difficult. Existing centralized algorithms suffer from single point failure of the central controller (due to communication disconnection), and they are performance-inefficient in the case of concurrent execution. In this paper, we extend our previous work (Lu et al. 2016) and propose a fault tolerant deadlock detection algorithm in distributed systems. The extended proposed algorithm can tolerate a certain extent of communication disconnection between computing nodes or processes. A central controller is used to collect requesting conditions, construct a wait-for graph, and detect deadlocks. The proposed algorithm can select a new central controller if the current central leader fails due to communication disconnections. The liveness and safety properties of the proposed algorithm are proved in this paper. Experimental results show that the proposed algorithm provides better performance than most of existing algorithms in terms of message number, data traffic, and execution time. In addition, the proposed algorithm provides additional fault tolerance compared to existing deadlock detection algorithms in the case of communication disconnection.

Matching information across image and text modalities is a fundamental challenge for many applications that involve both vision and natural language processing. The objective is to find efficient similarity metrics to compare the similarity between visual and textual information. Existing approaches mainly match the local visual objects and the sentence words in a shared space with attention mechanisms. The matching performance is still limited because the similarity computation is based on simple comparisons of the matching features, ignoring the characteristics of their distribution in the data. In this paper, we address this limitation with an efficient learning objective that considers the discriminative feature distributions between the visual objects and sentence words. Specifically, we propose a novel Adversarial Discriminative Domain Regularization (ADDR) learning framework, beyond the paradigm metric learning objective, to construct a set of discriminative data domains within each image-text pairs. Our approach can generally improve the learning efficiency and the performance of existing metrics learning frameworks by regulating the distribution of the hidden space between the matching pairs. The experimental results show that this new approach significantly improves the overall performance of several popular cross-modal matching techniques (SCAN, VSRN, BFAN) on the MS-COCO and Flickr30K benchmarks., Comment: 8 pages

We study how to leverage Web images to augment human-curated object detection datasets. Our approach is two-pronged. On the one hand, we retrieve Web images by image-to-image search, which incurs less domain shift from the curated data than other search methods. The Web images are diverse, supplying a wide variety of object poses, appearances, their interactions with the context, etc. On the other hand, we propose a novel learning method motivated by two parallel lines of work that explore unlabeled data for image classification: self-training and self-supervised learning. They fail to improve object detectors in their vanilla forms due to the domain gap between the Web images and curated datasets. To tackle this challenge, we propose a selective net to rectify the supervision signals in Web images. It not only identifies positive bounding boxes but also creates a safe zone for mining hard negative boxes. We report state-of-the-art results on detecting backpacks and chairs from everyday scenes, along with other challenging object classes.

In order to prevent large-scale disconnection accidents of photovoltaic power stations caused by voltage fluctuations and faults in power grids, low voltage ride-through characteristics and other detection tests are needed when photovoltaic power stations are connected to power grids. At present, field test and type test are the main methods of detection, but due to the limitation of detection devices and test conditions, the evaluation of low voltage ride-through characteristics of photovoltaic power stations still remains in the single-machine detection mode, which ignores the interaction between photovoltaic inverters and inverters, inverters and static var generator(SVG) devices. The test results cannot truly reflect the low voltage traversing characteristics of the photovoltaic power stations. Using RT-LAB hardware-in-the-loop simulation platform, real photovoltaic inverters and SVG controllers are introduced to build simulation models of photovoltaic power stations. The differences between single-machine detection of photovoltaic inverters and reactive power compensation devices and low-voltage ride-through characteristics detection of the whole station are compared and studied. The simulation results show that, compared with the stand-alone test, the whole station simulation modeling test method can reflect the voltage rise caused by the reactive power characteristics of photovoltaic inverters and SVG devices after the drop, and more truly reflect the low voltage traversing characteristics of photovoltaic power stations.

One method for three-dimensional (3D) measurement based on 23 binary structured light patterns is proposed for the medical electronic endoscope in the present study. The patterns are designed by using simple logical operations and tools from combinatorial mathematics. First, we project these patterns onto the objects in proper order, and all the modulated patterns are captured by a CMOS camera. Then, a series of preprocessing methods which contain projector-camera calibration, reflection detection and suppression, radial and tangential distortion correction are used to process these modulated patterns. Finally, we decode the patterns and reconstruct the shape of objects by structured light triangulation. An experimental prototype endoscope is built and the results of experiments on real pig stomach show that the proposed method is effective. The precise of depth measurement at the working distance of 80 mm is better than 0.5 mm.

Fake news, doubtful statements and other unreliable content not only differ with regard to the level of misinformation but also with respect to the underlying intents. Prior work on algorithmic truth assessment has mostly pursued binary classifiers—factual versus fake—and disregarded these finer shades of untruth. In manual analyses of questionable content, in contrast, more fine-grained distinctions have been proposed, such as distinguishing between hoaxes, irony and propaganda or the six-way truthfulness ratings by the PolitiFact community. In this paper, we present a principled automated approach to distinguish these different cases while assessing and classifying news articles and claims. Our method is based on a hierarchy of five different kinds of fakeness and systematically explores a variety of signals from social media, capturing both the content and language of posts and the sharing and dissemination among users. The paper provides experimental results on the performance of our fine-grained classifier and a detailed analysis of the underlying features.

User experience (UX) is a key factor that affects software survival time. A rich line of research has studied the relationships between UX and software factors to modify software and improve user satisfaction. However, the existing machine learning models for predicting UX on small data set is not accurate enough, and research with traditional statistical methods only obtained indistinct relations among UX, user characteristics and software factors. With the goal of improving the accuracy of UX model and obtaining sufficient UX relationships, we propose Transfer in Cart (TrCart) algorithm and Transfer Adaboost in Cart (TrAdaBoostCart) algorithm. To verify this approach, we present the UX study on a desktop game and an android game. According to the experimental results, we find that the TrAdaBoostCart has better accuracy and interpretable results. Hence, the proposed approach provides important guidelines for the design process of mobile applications.

As deep neural networks (DNNs) have become increasingly important and popular, the robustness of DNNs is the key to the safety of both the Internet and physical world. Unfortunately, some recent studies show that adversarial examples, which are hard to be distinguished from real examples, can easily fool DNNs and manipulate their predictions. Upon observing that adversarial examples are mostly generated by gradient-based methods, in this paper, we first propose to use a simple yet very effective non-differentiable hybrid model that combines DNNs and random forests, rather than hide gradients from attackers, to defend against the attacks. Our experiments show that our model can successfully and completely defend the white-box attacks, has a lower transferability, and is quite resistant to three representative types of black-box attacks; while at the same time, our model achieves similar classification accuracy as the original DNNs. Finally, we investigate and suggest a criterion to define where to grow random forests in DNNs.

Social business moves beyond linear, process-driven organizations to create new, dynamic, networked businesses that focus on customer value. Enterprise network (EN) is used to support social business by maximizing current and future opportunities and facilitate network-enabled processes, which can lead to value co-creation. EN is a multi-level hypergraph model with enterprises, employees, products and other related entities. In this paper the authors refine the EN model and present the foundation of EN to support social businesses. Then they introduce a case study on China automobile supply network (CASN). For the similarity with social networks, they verify power-law and small world theories in EN with statistical results on this data set. These theories are fitful in EN, but some new characteristics exist. The structure of EN consists of star-shaped clusters and the authors extract ego networks taking suppliers and manufacturers as the ego respectively. With the structure and distribution features of EN, they present the enterprise business similarity analysis method based on common-neighbors. And they also introduce the tentative work to detect Dunbar circles in EN. To analyze the data in a more intuitional and effective way, the authors use some data visualization tools to process the data in EN.

The real-time crash risk analyses were proposed to establish the relationships between crash occurrence probability and pre-crash traffic operational conditions. Given its great application potentials that link with Active Traffic Management System (ATMS) for proactive safety management, it has become an important research area. Currently, researchers mainly developed the real-time crash risk analysis models with traffic flow descriptive statistics employed as explanatory variables and with re-sampled balanced dataset, which hold the limitations of insufficiently capturing the temporal-spatial traffic flow characteristics and failing to provide classification capabilities when deal with the imbalanced datasets. In this study, a Convolutional Neural Network (CNN) modelling approach with refined loss functions has been first time introduced to the real-time crash risk analyses. The primary objectives of the proposed CNN models are: (1) utilizing the tensor-based data structure to explore the multi-dimensional, temporal-spatial correlated pre-crash operational features; and (2) optimizing the loss functions to overcome the low classification accuracy issue brought by the imbalanced data. Data from the Shanghai urban expressway system were utilized for the empirical analysis. And a total of three types of loss functions, including traditional binary cross entropy, the α-weighted cross entropy and the focal loss, were introduced and being tested with varying ratios of crash and non-crash datasets. The modeling results show that the CNN model has better classification performance compared to the traditional Multi-layer Perceptrons (MLP) model with the tensor-based structure data. Besides, the developed CNN model with focal loss function has substantial classification enhancement under the imbalanced datasets. Finally, the distributions of predicting probabilities for balanced and imbalanced datasets were plotted to understand the effects of the imbalanced dataset and revealed how the proposed CNN model with focal loss function improves the model performance.