Hyperparameter on-line learning of stochastic resonance based threshold networks

Aiming at training the feed-forward threshold neural network consisting of nondifferentiable activation functions, the approach of noise injection forms a stochastic resonance based threshold network that can be optimized by various gradient-based optimizers. The introduction of injected noise extends the noise level into the parameter space of the designed threshold network, but leads to a highly non-convex optimization landscape of the loss function. Thus, the hyperparameter on-line learning procedure with respective to network weights and noise levels becomes of challenge. It is shown that the Adam optimizer, as an adaptive variant of stochastic gradient descent, manifests its superior learning ability in training the stochastic resonance based threshold network effectively. Experimental results demonstrate the significant improvement of performance of the designed threshold network trained by the Adam optimizer for function approximation and image classification.     

考虑多种损坏构成特征的沥青路面预防性养护决策方法

由于沥青路面损坏构成的多样性, 相同的路面状况指数(pavement condition index, PCI)可能代表不同的损坏组合. 当多种损坏并存且损坏程度接近时, 用PCI和主导损坏(最严重、扣分最多的路面损坏)难以得到具有针对性的养护对策. 因此, 通过对PCI的深入分析, 明确了主导损坏代表性不足的路段, 以现行预防性养护决策方法为基础, 补充了一种考虑损坏构成特征、更具针对性的决策方法. 以上海城市道路近5年的检测、养护数据为分析基础, 首先利用有序聚类算法将路段按PCI水平分组, 分析了不同阶段路面损坏构成和差异水平; 然后, 针对多种损坏并存且损坏差异不显著的路段, 根据预防性养护的实施效果筛选了能够反映正确预防性养护经验的有效养护路段; 最后, 基于有效养护路段建立并对比分析了2个基于BP(back propagation)神经网络的养护决策模型. 结果表明: 当PCI水平介于优良(84.4~93.0分)时, 不同损坏程度接近, 主导损坏代表性不足; 考虑多种损坏构成特征的BP神经网络模型表现出更高的决策精度, 测试集决策正确率达86.20%, 优于仅考虑主导损坏的模型(58.50%). BP神经网络与传统决策树法结合能够优化沥青路面决策过程, 提高养护对策选取的针对性.     

植入式硅神经微电极的发展

 神经科学和神经工程研究需要研究大脑神经元的电活动情况，以了解大脑产生、传输和处理信息的机制。植入式神经微电极作为一种传感器件，是时间分辨率最高的神经电活动传感手段之一。介绍了国内外几种主要的植入式硅基神经微电极的结构特点、制备方法和性能特点。分析表明，未来通过不断结构优化和改性修饰，特别是在高通量的神经记录方面，通过与同样基于硅材料的电路的集成，硅神经微电极能够进一步提高生物相容性，解决大规模的电极通道体内外传输与连接问题，实现对神经元的在体大规模长时间记录。     

Enhancing signal-to-noise ratio and resolution in low-field NMR relaxation measurements using post-acquisition digital filters

The traditional way to enhance signal-to-noise ratio (SNR) of nuclear magnetic resonance (NMR) signals is to increase the number of scans. However, this procedure increases the measuring time that can be prohibitive for some applications. Therefore, we have tested the use of several post-acquisition digital filters to enhance SNR up to one order of magnitude in time domain NMR (TD-NMR) relaxation measurements. The procedures were studied using continuous wave free precession (CWFP-T₁) signals, acquired with very low flip angles that contain six times more noise than the Carr–Purcell–Meiboom–Gill (CPMG) signal of the same sample and experimental time. Linear (LI) and logarithmic (LO) data compression, low-pass infinity impulse response (LP), Savitzky–Golay (SG), and wavelet transform (WA) post-acquisition filters enhanced the SNR of the CWFP-T₁ signals by at least six times. The best filters were LO, SG, and WA that have high enhancement in SNR without significant distortions in the ILT relaxation distribution data. Therefore, it was demonstrated that these post-acquisition digital filters could be a useful way to denoise CWFP-T₁, as well as CPMG noisy signals, and consequently reducing the experimental time. It was also demonstrated that filtered CWFP-T₁ method has the potential to be a rapid and nondestructive method to measure fat content in beef and certainly in other meat samples.     

梯度网络平稳点集合的Hopfield稳定性与吸引域

给出梯度神经网络模型的电路实现 ,证明了网络平稳点集合的全局 Hopfield稳定性。对各个平稳点吸引域的几何估计得到了相关结果 ,这为网络初始输入的数量与方向判定提供了基本的量度参数准则。     

Research on Tracing Evaluation System in Virtual Enterprise Based on Neural Network

The tracing evaluation index system was designed to be used in virtual enterprise and established neural network trace evaluation model. As a result, it was more simple and nicer than the traditional method, so it had wider application foreground.     

基于神经网络智能天线的一种波达方向估计方法

该文提出利用径向基函数神经网络来估计无线通信中智能天线的信号到达方向．为了减少输入，利用信号相关阵的对称性质，仅考虑相关阵中的上三角或下三角的部分元素作为网络的输入量，计算量较小．计算机模拟结果表明，该方案是有效的．     

焊接过程的RBF神经网络建模

介绍了RBF神经网络的性能和算法结构 ,建立了RBF神经网络在船舶焊接过程中用于焊接变形预测分析的模型 ,并探讨了其应用和发展趋势     

Hybrid SVM／HMM Method for Face Recognition

A face recognition system based on Support Vector Machine(SVM) and Hidden Markov Model (HMM) has been proposed. The powerful discriminative ability of SVM is combined with the temporal modeling ability of HMM. The output of SVM is moderated to be probability output, which replaces the Mixture of Gauss (MOG) in HMM. Wavelet transformation is used to extract observation vector, which reduces the data dimension and improves the robustness.The hybrid system is compared with pure HMM face recognition method based on ORL face database and Yale face database. Experiments results show that the hybrid method has better performance.     

广义正定矩阵的分析与应用

分析了广义正定矩阵,导出了矩阵广义正定的若干新判定准则,并以应用实例研究了其在神经网络中的应用.