Experimentation on end-to-end performance aware algorithms in the federated environment of the heterogeneous PlanetLab and NITOS testbeds

The constantly increasing diversity of the infrastructure used to deliver Internet services to the end user has created a demand for experimental network facilities featuring heterogeneous resources. Therefore, federation of existing network testbeds has been identified as a key goal in the testbed community, leading to a recent activity burst in this research field. In this paper, we present a federation scheme that was built during the Onelab 2 EU project. This scheme federates the NITOS wireless testbed with the wired PlanetLab Europe testbed, allowing researchers to access and use heterogeneous experimental facilities under an integrated environment. The usefulness of the resulting federated facility is demonstrated through the testing of an implemented end-to-end delay aware association scheme proposed for wireless mesh networks. We present extensive experiments under both wired congestion and wireless channel contention conditions that demonstrate the effectiveness of the proposed approach in realistic settings. The experiments are also reproduced in a well-established network simulator and a comparative study between the results obtained in the realistic and simulated environments is presented. Both the architectural building blocks that enable the federation of the testbeds and the execution of the experiment on combined resources, as well as the important insights obtained from the experimental results are described and analyzed, pointing out the importance of integrated experimental facilities for the design and development of the Future Internet.     

Voice conversion using General Regression Neural Network

The objective of voice conversion system is to formulate the mapping function which can transform the source speaker characteristics to that of the target speaker. In this paper, we propose the General Regression Neural Network (GRNN) based model for voice conversion. It is a single pass learning network that makes the training procedure fast and comparatively less time consuming. The proposed system uses the shape of the vocal tract, the shape of the glottal pulse (excitation signal) and long term prosodic features to carry out the voice conversion task. In this paper, the shape of the vocal tract and the shape of source excitation of a particular speaker are represented using Line Spectral Frequencies (LSFs) and Linear Prediction (LP) residual respectively. GRNN is used to obtain the mapping function between the source and target speakers. The direct transformation of the time domain residual using Artificial Neural Network (ANN) causes phase change and generates artifacts in consecutive frames. In order to alleviate it, wavelet packet decomposed coefficients are used to characterize the excitation of the speech signal. The long term prosodic parameters namely, pitch contour (intonation) and the energy profile of the test signal are also modified in relation to that of the target (desired) speaker using the baseline method. The relative performances of the proposed model are compared to voice conversion system based on the state of the art RBF and GMM models using objective and subjective evaluation measures. The evaluation measures show that the proposed GRNN based voice conversion system performs slightly better than the state of the art models.     

Deep learning based pipeline with multichannel inputs for patent classification

This work introduces a deep learning pipeline for automatic patent classification with multichannel inputs based on LSTM and word vector embeddings. Sophisticated text mining methods are used to extract the most important segments from patent texts, and a domain-specific pre-trained word embeddings model for the patent domain is developed; it was trained on a very large dataset of more than five million patents. The deep learning pipeline is using multiple parallel LSTM networks that read the source patent document using different input dimensions namely embeddings of different segments of patent texts, and sparse linear input of different metadata. Classifying patents into corresponding technical fields is selected as a use case. In this use case, a series of patent classification experiments are conducted on different patent datasets, and the experimental results indicate that using the segments of patent texts as well as the metadata as multichannel inputs for a deep neural network model, achieves better performance than one input channel.     

Affective Experience Measurement in Outdoor Environment via Machine Visual Cognition

Outdoor environments with quality landscapes can benefit people’s physical and mental health. Real-time assessment on individuals’ environmental affective experience can improve the scientism in measuring the quality of outdoor environments. Existing measurement methods are often insufficient for the cases of a larger site area or sample size. The machine visual cognition of Artificial Intelligence can realize the recognition of facial expressions and the changes in video images, which supports high-precision and long-cycle measurements on individuals’ affective experience in outdoor environments. Taking an urban community square as the study site, this research simultaneously collects participants’ facial data from video images and their electrodermal activity data, wherein Convolutional Neural Network algorithm model is trained with a deep learning algorithm, i.e. codec–SVM optimized model, whose reliability is tested through an additional experiment. The research reveals that: 1) The accuracy rate of the main and additional experiments in measuring individuals’ affective experience is 82.01% and 65.08%, respectively; 2) The additional experiment verifies the application potential of the codec–SVM optimized model; And 3) the model works more effective for outdoor scenarios with varying usage behaviors and open views. Therefore, machine visual cognition can be used for emotion measurement in a larger site area or sample size and contributes to the effectiveness of landscape optimization efforts, especially as an instrumental tool to study the affective experience of the ones who have communication or reading disability. The findings also demonstrate the model’s great potential in building Smart Cities with refined public services.     

基于双重金字塔网络的视频目标分割方法

针对复杂视频场景中难以分割特定目标的问题，提出一种基于双重金字塔网络（DPN）的视频目标分割方法。首先，通过调制网络的单向传递让分割模型适应特定目标的外观。具体而言，从给定目标的视觉和空间信息中学习一种调制器，并通过调制器调节分割网络的中间层以适应特定目标的外观变化。然后，通过基于不同区域的上下文聚合的方法，在分割网络的最后一层中聚合全局上下文信息。最后，通过横向连接的自左而右结构，在所有尺度中构建高阶语义特征图。所提出的视频目标分割方法是一个可以端到端训练的分割网络。大量实验结果表明，所提方法在DAVIS2016数据集上的性能与较先进的使用在线微调的方法相比，可达到相竞争的结果，且在DAVIS2017数据集上性能较优。     

盲去模糊的多尺度编解码深度卷积网络

针对拍摄场景中物体运动不一致所带来的非均匀模糊，为提高复杂运动场景中去模糊的效果，提出一种多尺度编解码深度卷积网络。该网络采用"从粗到细"的多尺度级联结构，在模糊核未知条件下，实现盲去模糊；其中，在该网络的编解码模块中，提出一种快速多尺度残差块，使用两个感受野不同的分支增强网络对多尺度特征的适应能力；此外，在编解码之间增加跳跃连接，丰富解码端信息。与2018年国际计算机视觉与模式识别会议（CVPR）上提出的多尺度循环网络相比，峰值信噪比（PSNR）高出0.06 dB；与2017年CVPR上提出的深度多尺度卷积网络相比，峰值信噪比和平均结构相似性（MSSIM）分别提高了1.4%和3.2%。实验结果表明，该网络能快速去除图像模糊，恢复出图像原有的边缘结构和纹理细节。     

Neural network modelling on temperature coefficient of surface tension and its usage in melting point prediction of nanosized metal particles

Temperature coefficient of surface tension is a very important parameter to calculate phase diagrams of nanoparticle metal systems. In this paper, neural network calculation was for the first time used to evaluate the temperature coefficient. It shows that the constructed neural network can predict the temperature coefficient values for 37 metals, with the deviation from the averaged experimental measurements smaller than 25%. Furthermore, the neural network predictions were compared with the calculated values by using an empirical equation and it shows a better performance.     

Using Neural Networks to predict HFACS unsafe acts from the pre-conditions of unsafe acts

Abstract

Human Factors Analysis and Classification System (HFACS) is based upon Reason’s organizational model of human error which suggests that there is a ‘one to many’ mapping of condition tokens (HFACS level 2 psychological precursors) to unsafe act tokens (HFACS level 1 error and violations). Using accident data derived from 523 military aircraft accidents, the relationship between HFACS level 2 preconditions and level 1 unsafe acts was modelled using an artificial neural network (NN). This allowed an empirical model to be developed congruent with the underlying theory of HFACS. The NN solution produced an average overall classification rate of ca. 74% for all unsafe acts from information derived from their level 2 preconditions. However, the correct classification rate was superior for decision- and skill-based errors, than for perceptual errors and violations.

Practitioner Summary: A model to predict unsafe acts (HFACS level 1) from their preconditions (HFACS level 2) was developed from the analysis of 523 military aircraft accidents using an artificial NN. The results could correctly predict approximately 74% of errors.     

基于改进SSD的高效目标检测方法

为改善一阶段目标检测算法检测精度较差的缺陷，提出一种基于SSD的高效多目标定位检测算法FSD。该算法主要从两个方面对一阶段目标检测算法进行改进：设计了一个更高效的密集残差网络，即R-DenseNet，通过采用一种更窄的密集网络结构形式，在保持特征提取容量的同时降低了计算复杂度，从而提高了算法的检测和收敛性能；改进了损失函数，通过抑制易分样本在损失函数中的权重，提高算法的鲁棒性，改善了目标检测中样本失衡的现象。采用Tensorflow深度学习框架部署网络，并在搭载Nvidia Titan X的Ubuntu上开展实验，实验表明FSD在COCO和PASCAL VOC这两个目标检测数据集上上都取得了最高的检测精度，其中FSD300D的检测精度相比SSD300有3.7%提升，检测相率比SSD有10.87%提升。     

GPR target detection using a neural network classifier designed by a multi-objective genetic algorithm

Ground Penetrating Radar (GPR) is an electromagnetic sensing technology employed for localization of underground utilities, pipes, and other types of objects. The radargrams typically obtained have a high dimensionality, containing a number of signatures with hyperbolic pattern shapes, and can be processed to retrieve information about the target’s locations, depths and material type of underground soil. The classical Hough Transform approach used to reconstruct these hyperbola shapes is computationally expensive, given the large dimensionality of the radargrams. In literature, several approaches propose to first approximate the location of hyperbolas to small segments through a classification stage, before applying the Hough transform over these segments. However, the published classifiers designed for this task present a relatively complex architecture.Aiming at an improved target localization, we propose an alternative classification methodology. The goal is to classify windows of GPR radargrams into two classes (with or without target) using a neural network radial basis function (RBF), designed via a multi-objective genetic algorithm (MOGA). To capture samples’ fine details, high order statistic cumulant features (HOS) were used. Feature selection was performed by MOGA, with an optional prior reduction using a mutual information (MIFS) approach. The obtained results demonstrate improvement of the classification performance when compared with other models designed with the same data and are among the best results available in the literature, albeit the large reduction in classifier complexity.