Ensemble echo network with deep architecture for time-series modeling

Neural Computing and Applications - Echo state network belongs to a kind of recurrent neural networks that have been extensively employed to model time-series datasets. The function of reservoir in...     

J-LDFR: joint low-level and deep neural network feature representations for pedestrian gender classification

Neural Computing and Applications - Appearance-based gender classification is one of the key areas in pedestrian analysis, and it has many useful applications such as visual surveillance, predict...     

Long-term time-series pollution forecast using statistical and deep learning methods

Neural Computing and Applications - Tackling air pollution has become of utmost importance since the last few decades. Different statistical as well as deep learning methods have been proposed till...     

Visualising hierarchical associations

Recent association-mining research has led to the development of techniques that allow the accommodation of concept hierarchies within the mining process. This extension results in the discovery of rules which associate not only groups of items but which are also influenced by the hierarchies within which an item may reside. Given this, there then arises a need for techniques whereby such hierarchical associations can be presented to the user. Current association rule visualisation techniques are limited, as they do not effectively incorporate or enable the visualisation of hierarchical semantics. This paper presents a review of current hierarchical and association visualisation techniques and introduces a novel technique for visualising hierarchical association rules.     

多步积累奖励的双重时序Q网络算法

车辆行驶控制决策是无人驾驶的核心技术,现有基于深度强化学习的无人驾驶控制决策算法存在处理数据效率低、无法有效提取状态间时序特征等问题.因此本文提出了一种基于多步积累奖励的双重时序Q网络算法.首先,设计了一种多步积累奖励方法,该方法对未来多步即时奖励的累加和进行均值化,与当前即时奖励共同作用于智能体的控制策略,并在奖励函...     

GSAM: A deep neural network model for extracting computational representations of Chinese addresses fused with geospatial feature

Addresses are one of the most important geographical reference systems in natural languages. In China, due to the relatively backward address planning, there are a large number of non-standard addresses. This kind of unstructured text makes the management and application of Chinese addresses much more difficult. However, by extracting the computational representations of addresses, it can be structured and its related applications can be extended more conveniently. Therefore, this paper utilizes a deep neural language model from natural language processing (NLP) to automatically extract computational representations through an unsupervised address language model (ALM), which is trained in an unsupervised way and is suitable for a large-scale address corpus. We propose a solution to fuse addresses and geospatial features and construct a geospatial-semantic address model (GSAM) that supports a variety of downstream tasks. Our proposed GSAM constructing process consists of three phases. First, we build an ALM using bidirectional encoder representations from Transformers (BERT) to learn the addresses' semantic representations. Then, the fusion clustering results of the semantic and geospatial information are obtained by a high-dimensional clustering algorithm. Finally, we construct the GSAM based on the fused clustering results using novel fine-tuning techniques. Furthermore, we apply the extracted computational representation from GSAM to the address location prediction task. The experimental results indicate that the target task accuracy of the ALM is 90.79%, and the result of semantic geospatial fusion clustering strongly correlates with fine-grained urban neighbourhood area division. The GSAM can accurately identify clustering labels and the values of evaluation metrics are all above 0.96. We also demonstrate that our model outperforms purely ALM-based and word2vec-based models by address location prediction task.     

Learning social representations with deep autoencoder for recommender system

World Wide Web - With the development of online social media, it attracts increasingly attentions to utilize social information for recommender systems. Based on the intuition that users are...     

Enhancing attributed network embedding via enriched attribute representations

Applied Intelligence - Attributed network embedding enables to generate low-dimensional representations of network objects by leveraging both network structure and attribute data. However, how to...     

Artificial neural network representations for hierarchical preference structures

In this paper, we introduce two artificial neural network formulations that can be used to assess the preference ratings from the pairwise comparison matrices of the Analytic Hierarchy Process. First, we introduce a modified Hopfield network that can determine the vector of preference ratings associated with a positive reciprocal comparison matrix. The dynamics of this network are mathematically equivalent to the power method, a widely used numerical method for computing the principal eigenvectors of square matrices. However, this Hopfield network representation is incapable of generalizing the preference patterns, and consequently is not suitable for approximating the preference ratings if the pairwise comparison judgments are imprecise. Second, we present a feed-forward neural network formulation that does have the ability to accurately approximate the preference ratings. We use a simulation experiment to verify the robustness of the feed-forward neural network formulation with respect to imprecise pairwise judgments. From the results of this experiment, we conclude that the feed-forward neural network formulation appears to be a powerful tool for analyzing discrete alternative multicriteria decision problems with imprecise or fuzzy ratio-scale preference judgments.     

Fuzzy neural network with general parameter adaptation for modelingof nonlinear time-series

By taking advantage of fuzzy systems and neural networks, a fuzzy-neural network with a general parameter (GP) learning algorithm and heuristic model structure determination is proposed in this paper. Our network model is based on the Gaussian radial basis function network (RBFN). We use the flexible GP approach both for initializing the off-line training algorithm and fine-tuning the nonlinear model efficiently in online operation. A modification of the robust unbiasedness criterion using distorter (UCD) is utilized for selecting the structural parameters of this adaptive model. The UCD approach provides the desired modeling accuracy and avoids the risk of over-fitting. In order to illustrate the operation of the proposed modeling scheme, it is experimentally applied to a fault detection application.     

改进的弹性网模型在深度神经网络中的应用

由于具有较高的模型复杂度，深层神经网络容易产生过拟合问题，为了减少该问题对网络性能的不利影响，提出一种基于改进的弹性网模型的深度学习优化方法。首先，考虑到变量之间的相关性，对弹性网模型中的L1范数的不同变量进行自适应加权，从而得到L2范数与自适应加权的L1范数的线性组合。其次，将改进的弹性网络模型与深度学习的优化模型相结合，给出在这种新正则项约束下求解神经网络参数的过程。然后，推导出改进的弹性网模型在神经网络优化中具有群组选择能力和Oracle性质，进而从理论上保证该模型是一种更加鲁棒的正则化方法。最后，在多个回归问题和分类问题的实验中，相对于L1、L2和弹性网正则项，该方法的回归测试误差可分别平均降低87.09、88.54和47.02，分类测试准确度可分别平均提高3.98、2.92和3.58个百分点。由此，在理论和实验两方面验证了改进的弹性网模型可以有效地增强深层神经网络的泛化能力，提升优化算法的性能，解决深度学习的过拟合问题。     

Learning multi-level and multi-scale deep representations for privacy image classification

Multimedia Tools and Applications - Privacy image classification can help people detect privacy images when people share images. In this paper, we propose a novel method using multi-level and...     

深浅层表示融合的半监督视频目标分割

为了解决半监督视频目标分割任务中,分割精度与分割速度难以兼顾以及无法对视频中与前景相似的背景目标做出有效区分的问题,提出一种基于深浅层特征融合的半监督视频目标分割算法。首先,利用预先生成的粗糙掩膜对图像特征进行处理,以获取更鲁棒的特征;然后,通过注意力模型提取深层语义信息;最后,将深层语义信息与浅层位置信息进行融合,从而得到更加精确的分割结果。在多个流行的数据集上进行了实验,实验结果表明：在分割运行速度基本不变的情况下,所提算法在DAVIS 2016数据集上的雅卡尔（J）指标相较于学习快速鲁棒目标模型的视频目标分割（FRTM）算法提高了1.8个百分点,综合评价指标为J和F得分的均值J&F相较于FRTM提高了2.3个百分点;同时,在DAVIS 2017数据集上,所提算法的J指标比FRTM提升了1.2个百分点,综合评价指标J&F比FRTM提升了1.1个百分点。以上结果充分说明所提算法能够在保持较快分割速度的情况下实现更高的分割精度,并且能够有效区别相似的前景与背景目标,具有较强的鲁棒性。可见所提算法在平衡速度与精度以及有效区分前景背景方面的优越性能。     

基于梯度的深度网络剪枝算法

深度神经网络模型通常存在大量冗余的权重参数,计算深度网络模型需要占用大量的计算资源和存储空间,导致深度网络模型难以部署在一些边缘设备和嵌入式设备上。针对这一问题,提出了一种基于梯度的深度网络剪枝(GDP)算法。GDP算法核心思想是以梯度作为评判权值重要性的依据。首先,通过自适应的方法找出阈值进行权值参数的筛选;然后,剔除那些小于阈值的梯度所对应的权值;最后,重新训练剪枝后的深度网络模型来恢复网络精度。实验结果表明:在CIFAR-10数据集上,GDP算法在精度仅下降0.14个百分点的情况下,计算量减少了35.3个百分点;与当前流行的PFEC算法相比,GDP算法使网络模型精度提高了0.13个百分点,计算量下降了1.1个百分点,具有更优越的深度网络压缩与加速性能。     

再现3D对象

多亏有了Google上的一种免费工具,3D建模变得不再是一个痛苦而笨拙的过程。Jerome Turner介绍SketchUp的主要功能。     

A local linear radial basis function neural network for financial time-series forecasting

In this paper a Local Linear Radial Basis Function Neural Network (LLRBFN) is presented. The difference between the proposed neural network and the conventional Radial Basis Function Neural Network (RBFN) is connection weights between the hidden layer and the output layer which are replaced by a local linear model in the LLRBFN. A modified Particle Swarm Optimization (PSO) with hunter particles is introduced for training the LLRBFN. The proposed methods have been applied for prediction of financial time-series and the result shows the feasibility and effectiveness.     

Identification of time-series signals using a dynamic neural network with GA-based training

We propose a dynamic neural network (DNN) that realizes a dynamic property and has a network structure with the properties of inertia, viscosity, and stiffness without time-delayed input elements, and a training algorithm based on a genetic algorithm (GA). In a previous study, we proposed a modified training algorithm for the DNN based on the error back-propagation method. However, in the previous method it was necessary to determine the values of the DNN property parameters by trial and error. In the newly proposed DNN, the GA is designed to train not only the connecting weights but also the property parameters of the DNN. Simulation results show that the DNN trained by the GA obtains good performance for time-series patterns generated from an unknown system, and provides a higher performance than the conventional neural network. This work was presented in part at the 10th International Symposium on Artificial Life and Robotics, 0ita, Japan, February 4–6, 2005     

Learning deep representations for semantic image parsing: a comprehensive overview

Semantic image parsing, which refers to the process of decomposing images into semantic regions and constructing the structure representation of the input, has recently aroused widespread interest in the field of computer vision. The recent application of deep representation learning has driven this field into a new stage of development. In this paper, we summarize three aspects of the progress of research on semantic image parsing, i.e., category-level semantic segmentation, instance-level semantic segmentation, and beyond segmentation. Specifically, we first review the general frameworks for each task and introduce the relevant variants. The advantages and limitations of each method are also discussed. Moreover, we present a comprehensive comparison of different benchmark datasets and evaluation metrics. Finally, we explore the future trends and challenges of semantic image parsing.     

基于深度双Q网络的多用户蜂窝网络功率分配算法研究

针对现有蜂窝网络功率分配算法存在泛化能力弱、效率低等问题进行了研究,提出基于深度双Q网络(deep double Q network,DDQN)的功率分配算法。采用双神经网络结构,解决强化学习过程中易出现的维度灾难及值函数过估计问题;对状态信息进行设计并输入神经网络,输出智能体的动作行为,并设计奖赏函数反馈给神经网络,使智能体可以有效地自主学习,多次迭代得到最优的功率分配策略。仿真结果表明,所提的模型可获得的平均速率为1.89,平均运行时间为0.0013 s,在不同用户密度及小区数量下均可达到最高的平均速率,验证了算法的有效性,为蜂窝网络资源分配问题提供了新的思路。