最优线性联想记忆网络的γ谱解析

利用最优线性联想记忆神经网络,对γ能谱进行了定性识别与定量分析,成功克服了传统解谱方法对操作人员要求高、运算速度慢、不能准确识别有重峰的复杂γ能谱等问题.采用全谱输入法,利用整个能谱的信息,降低了对探测器能量分辨率的要求,避免了寻峰、能量刻度与效率刻度,准确识别了单核素能谱与几种核素的混合能谱,从而成为一种行之有效的解谱手段,为高性能便携式探测器解谱软件的开发提供了依据.     

Multilayer associative memory and its hybrid optical implementation

We propose a multilayer associative memory with a winner-take-all operation on the inner product between an input and stored exemplars. The winner-take-all operation is performed by a unit-step operation with an adaptive-threshold strategy. We show that the multilayer-associative-memory unit-step operation with an adaptive-threshold strategy has a high noise immunity and a large storage capacity, and it is also capable of extending to a gray-level associative memory with a phase-representation technique. A hybrid optical implementation with a proof-of-concept experiment is also provided.     

Convergence of higher-order two-state neural networks with modified updating

The Hopfield network is a standard tool for maximizing aquadratic objective function over the discrete set {− 1,1} ⁿ . It is well-known that if a Hopfield network is operated in anasynchronous mode, then the state vector of the network converges to a local maximum of the objective function; if the network is operated in asynchronous mode, then the state vector either converges to a local maximum, or else goes into a limit cycle of length two. In this paper, we examine the behaviour ofhigher-order neural networks, that is, networks used for maximizing objective functions that are not necessarily quadratic. It is shown that one can assume, without loss of generality, that the objective function to be maximized ismultilinear. Three methods are given for updating the state vector of the neural network, called the asynchronous, the best neighbour and the gradient rules, respectively. For Hopfield networks with a quadratic objective function, the asynchronous rule proposed here reduces to the standard asynchronous updating, while the gradient rule reduces to synchronous updating; the best neighbour rule does not appear to have been considered previously. It is shown that both the asynchronous updating rule and the best neighbour rule converge to a local maximum of the objective function within a finite number of time steps. Moreover, under certain conditions, under the best neighbour rule, each global maximum has a nonzero radius of direct attraction; in general, this may not be true of the asynchronous rule. However, the behaviour of the gradient updating rule is not well-understood. For this purpose, amodified gradient updating rule is presented, that incorporates bothtemporal as well as spatial correlations among the neurons. For the modified updating rule, it is shown that, after a finite number of time steps, the network state vector goes into a limit cycle of lengthm, wherem is the degree of the objective function. Ifm = 2, i.e., for quadratic objective functions, the modified updating rule reduces to the synchronous updating rule for Hopfield networks. Hence the results presented here are “true” generalizations of previously known results.     

Unipolar terminal-attractor-based neural associative memory with adaptive threshold and perfect convergence

A perfectly convergent unipolar neural associative-memory system based on nonlinear dynamical terminal attractors is presented. With adaptive setting of the threshold values for the dynamic iteration for the unipolar binary neuron states with terminal attractors, perfect convergence is achieved. This achievement and correct retrieval are demonstrated by computer simulation. The simulations are completed (1) by exhaustive tests with all of the possible combinations of stored and test vectors in small-scale networks and (2) by Monte Carlo simulations with randomly generated stored and test vectors in large-scale networks with an M/N ratio of 4 (M is the number of stored vectors; N is the number of neurons ≤ 256). An experiment with exclusive-oR logic operations with liquid-crystal-television spatial light modulators is used to show the feasibility of an optoelectronic implementation of the model. The behavior of terminal attractors in basins of energy space is illustrated by examples.     

Modeling of materials with fading memory using neural networks

A neural network‐based concept for the solution of a fractional differential equation is presented in this paper. Fractional differential equations are used to model the behavior of rheological materials that exhibit special load (stress) history characteristics (e.g. fading memory). The new concept focuses on rheological materials that exhibit Newtonian‐like displacement behavior when undergoing (time varying) creep loads. For this purpose, a partial recurrent artificial neural network is developed. The network supersedes the storage of the entire load (stress) history in contrast to the exact solution of the fractional differential equation, where access to all previous load (stress) increments is required to determine the new displacement (strain) increment. The network is trained using data obtained from six different creep simulations. These creep simulations have been conducted by means of the exact solution of the fractional differential equation, which is also included in the paper. Furthermore, the network architecture as well as a complete set of network parameters is given. A validation of the network has been carried out and its outcome is discussed in the paper. To illustrate the particular way the network works, all relevant algorithms (e.g. scaling of the input data, data processing, transformation of the output signal, etc.) are provided to the reader in this paper. Copyright © 2008 John Wiley & Sons, Ltd.     

Passivity analysis of complex dynamical networks with multiple time-varying delays

We investigate input passivity and output passivity for a generalized complex network with non-linear, time-varying, non-symmetric and delayed coupling. By constructing some suitable Lyapunov functionals, several sufficient conditions ensuring input passivity and output passivity are derived for complex dynamical networks. Finally, two numerical examples are given to show the effectiveness of the obtained results.     

Data analysis of stellar specklegrams with neural networks

An artificial neural network is applied to analysis of specklegrams of binary stars. Parameters of a binary star, the angular separation and the position angle, are estimated from the specklegrams by use of neural networks for each parameter. It is shown that a neural network is useful to analyze stellar specklegrams of binary stars.     

Bogdanov–Takens bifurcation in a neutral BAM neural networks model with delays

In this study, the authors first discuss the existence of Bogdanov–Takens and triple zero singularity of a five neurons neutral bidirectional associative memory neural networks model with two delays. Then, by utilising the centre manifold reduction and choosing suitable bifurcation parameters, the second‐order and the third‐order normal forms of the Bogdanov–Takens bifurcation for the system are obtained. Finally, the obtained normal form and numerical simulations show some interesting phenomena such as the existence of a stable fixed point, a pair of stable non‐trivial equilibria, a stable limit cycles, heteroclinic orbits, homoclinic orbits, coexistence of two stable non‐trivial equilibria and a stable limit cycles in the neighbourhood of the Bogdanov–Takens bifurcation critical point.Inspec keywords: neurophysiology, neural nets, bifurcation, delays, critical pointsOther keywords: Bogdanov‐Takens bifurcation critical point, neutral BAM neural networks, bidirectional associative memory, delays, triple zero singularity, neurons, centre manifold reduction, bifurcation parameters, second‐order normal forms, third‐order normal forms, numerical simulations, stable fixed point, stable nontrivial equilibria, stable limit cycles, heteroclinic orbits, homoclinic orbits     

Convergence and stability analysis of the deformable discrete element method

This work investigates numerical properties of the algorithm of the discrete element method (DEM) employing deformable circular disks presented in the authors' earlier publication. The new formulation called the deformable DEM (DDEM) enhances the standard DEM (SDEM) by introducing an additional (global) deformation mode caused by the stresses in the particles induced by the contact forces. An accurate computation of the contact forces would require an iterative solution of the implicit relationship between the contact forces and particle displacements. In order to preserve efficiency of the DEM, the new formulation has been adapted to the explicit time integration. It has been shown that the explicit DDEM algorithm is conditionally stable and there are two restrictions on its stability. Except for the limitation of the time step as in the SDEM, the stability in the DDEM is governed by the convergence criterion of the iterative solution of the contact forces. The convergence and stability limits have been determined analytically and numerically for selected regular and irregular configurations. It has also been found out that the critical time step in DDEM remains unchanged with respect to the SDEM.     

Convergence of the discrete dipole approximation. I. Theoretical analysis

We perform a rigorous theoretical convergence analysis of the discrete dipole approximation (DDA). We prove that errors in any measured quantity are bounded by a sum of a linear term and a quadratic term in the size of a dipole d when the latter is in the range of DDA applicability. Moreover, the linear term is significantly smaller for cubically than for noncubically shaped scatterers. Therefore, for small d, errors for cubically shaped particles are much smaller than for noncubically shaped ones. The relative importance of the linear term decreases with increasing size; hence convergence of DDA for large enough scatterers is quadratic in the common range of d. Extensive numerical simulations are carried out for a wide range of d. Finally, we discuss a number of new developments in DDA and their consequences for convergence.     

An analysis of the equilibria of neural networks with linear interconnections

In this paper, we analyse the equilibria of neural networks which consist of a set of sigmoid nonlinearities with linear interconnections,without assuming that the interconnections are symmetric or that there are no self-interactions. By eliminating these assumptions, we are able to study the effects of imperfect implementation on the behaviour of Hopfield networks. If one views the neural network as evolving on the openn-dimensional hypercubeH = (0, 1) ⁿ , we have the following conclusions as the neural characteristics become steeper and steeper: (i) There is at most one equilibrium in any compact subset ofH, and under mild assumptions this equilibrium is unstable. In fact, the dimension of the stable manifold of this equilibrium is the same as the number of eigenvalues of the interconnection matrix with negative real parts. (ii) There might be some equilibria in the faces ofH, and under mild conditions these are always unstable. Moreover, it is easy to compute the dimension of the stable manifold of each such equilibrium. (iii) A systematic procedure is given for determining which corners of the hypercubeH contain equilibria, and it is shown that all equilibria in the corners ofH are asymptotically stable.     

前馈神经网络混合训练算法及其应用

介绍了单隐层前馈神经网络的混合训练算法(HFM)和正则化混合训练算法(RHFM),然后将该算法应用于UCI数据库上的实际回归例子中,并将其与BP、NNRW以及FM算法进行了比较.仿真实验表明,HFM算法的收敛速度优于其它几种算法,RHFM算法有较好的泛化性能,而NNRW算法在训练时间上占优,尽管如此,HFM算法在时间上还是大大优于BP算法.说明,混合训练算法是一种有效的算法.     

Implementation of associative memory using grating structures

An associative-memory model and its optical implementation with grating structures are presented. The transmission function of each pixel of the content-addressable memory is calculated by use of scalar diffraction theory. The filter of the calculated transmission function can be fabricated with computer-generated holography and a multiexposure holographic technique. The proposed approach is found useful in terms of storage and the simple thresholding at the number of on-state pixels in the input.     

Variance decomposition-based sensitivity analysis via neural networks

This paper illustrates a method for efficiently performing multiparametric sensitivity analyses of the reliability model of a given system. These analyses are of great importance for the identification of critical components in highly hazardous plants, such as the nuclear or chemical ones, thus providing significant insights for their risk-based design and management. The technique used to quantify the importance of a component parameter with respect to the system model is based on a classical decomposition of the variance. When the model of the system is realistically complicated (e.g. by aging, stand-by, maintenance, etc.), its analytical evaluation soon becomes impractical and one is better off resorting to Monte Carlo simulation techniques which, however, could be computationally burdensome. Therefore, since the variance decomposition method requires a large number of system evaluations, each one to be performed by Monte Carlo, the need arises for possibly substituting the Monte Carlo simulation model with a fast, approximated, algorithm. Here we investigate an approach which makes use of neural networks appropriately trained on the results of a Monte Carlo system reliability/availability evaluation to quickly provide with reasonable approximation, the values of the quantities of interest for the sensitivity analyses. The work was a joint effort between the Department of Nuclear Engineering of the Polytechnic of Milan, Italy, and the Institute for Systems, Informatics and Safety, Nuclear Safety Unit of the Joint Research Centre in Ispra, Italy which sponsored the project.     

Exact model reduction with delays: closed-form distributions and extensions to fully bi-directional monomolecular reactions

In order to systematically understand the qualitative and quantitative behaviour of chemical reaction networks, scientists must derive and analyse associated mathematical models. However, biochemical systems are often very large, with reactions occurring at multiple time scales, as evidenced by signalling pathways and gene expression kinetics. Owing to the associated computational costs, it is then many times impractical, if not impossible, to solve or simulate these systems with an appropriate level of detail. By consequence, there is a growing interest in developing techniques for the simplification or reduction of complex biochemical systems. Here, we extend our recently presented methodology on exact reduction of linear chains of reactions with delay distributions in two ways. First, we report that it is now possible to deal with fully bi-directional monomolecular systems, including degradations, synthesis and generalized bypass reactions. Second, we provide all derivations of associated delays in analytical, closed form. Both advances have a major impact on further reducing computational costs, while still retaining full accuracy. Thus, we expect our new methodology to respond to current simulation needs in pharmaceutical, chemical and biological research.     

基于混合神经网络的人类基因组启动子识别研究

提出了基于混合神经网络的人类基因组启动子识别的新方法：PromPredictor。该方法通过对转录起始位点（TSS）信息，调控区、编码区组成成分特征信息及CpG岛相关信息的综合来预测人类基因组启动子。对人类4、21、22号染色体启动子的预测结果为：敏感性达到了64．47％，特异性达到了82．2％。与其它三个算法相比，PmmPredictor具有更高的敏感性和特异性。研究中所用到的数据集合及用MATLAB编写的程序代码都可以从www．whtelecom．com／Prompredictor．htm下载得到。     

Neural networks to model dynamic systems with time delays

An algorithm is proposed to identify a neural network model that represents a nonlinear dynamic system with a multivariate time delay response. The algorithm consists of two major parts. The first one identifies the time delay vector for a given neural network structure. This task is accomplished by using an exhaustive integer enumeration algorithm that minimizes a statistical parameter to assess the performance of the neural network model. The second part uses a cross-validation strategy to identify the best neural network model. Since the structure that models a nonlinear system is usually unknown, the identification strategy consists of selecting several neural network structures and identifying the best time delay vector for each network. The modeling process starts with the simplest structure and progressively the complexity of the network is increased to end up with a complex structure. Finally, the network that offers the simplest structure with the best network performance is the one that exhibits the appropriate neural network structure with the corresponding optimal time delay vector. The Monte Carlo simulation technique was used to test the performance of the algorithm under the presence of linear and nonlinear relationships among several variables of dynamic systems and with a different time delay applied to each input variable. The introduced algorithm is used to detect a chemical reaction delay among enriched amyl acetate, acetic acid, water, and the pH of erythromycin sail. An appropriate neural network model was designed to model the pH of the erythromycin during a continuous extraction process. To the best of the authors knowledge the proposed algorithm is the only one currently available to identify time delay interactions in the multivariate input output variables of a system. The major drawback of the introduced algorithm is that it becomes very slow as the number of system inputs increases. This algorithm works efficiently in a system that involves five inputs or less.     

High-quality frame interpolation in computer generated holographic movies using coherent neural networks with a hybrid learning method

Computer generated holograms (CGHs) are widely used in optical tweezers, which will be employed in various research fields. We previously proposed an efficient generation method of CGH movies based on frame interpolation using coherent neural networks (CNNs) to reduce the high calculation cost of three-dimensional CGHs. At the same time, however, we also found that the quality observed in the interpolated CGH images needed to be improved even further so that the method could be accepted for general use. We report a successful error reduction in interpolated images by developing a new learning method of CNNs. We reduce the error by combining locally connected correlation learning and steepest descent learning in a sequential manner.     

Cost-tolerance analysis model based on a neural networks method

The purpose of tolerance design in product components is to produce a product with the least manufacturing cost possible, while meeting all functional requirements of the product. The product designer and process planner must fully understand the process accuracy and manufacturing cost of all kinds of manufacturing process to perform a good process plan job. Usually, the cost-tolerance model is constructed by a linear or non-linear regression analysis based on the data of the cost-tolerance experiment and to derive the correlation curve between the two. Though these correlation curves can show the relationship between manufacturing cost and tolerance, a fitting error is inevitable. In particular, there is considerable discrepancy in terms of the non-experimental data. A cost-tolerance analysis model based on a neural networks method is proposed. The cost-tolerance experimental data are used to set the training sets to establish a cost-tolerance network. Three representation modes of the cost-tolerance relationship are presented. First, the cost-tolerance relationship is derived from the grid points setting by the required tolerance accuracy. Second, a reasonable manufacturing cost of an unknown cost-tolerance experimental pair can be derived by the simulation of a cost-tolerance network. Third, an inference model based on a network's output is proposed to express the scope of the cost variation of various tolerances by means of a cost band. Comparison is also made with the high-order polynomial power function and exponential function cost-tolerance curves adopted by Yeo et al . Analytical results prove that the application of the cost-tolerance analysis model based on neural networks yields better performance in controlling the average fitting error than all conventional fitting models. The representation model using a cost band can identify precisely the possible cost variation range and reduce the chances of error in the tolerance design and cost estimation. It can thus provide important references for tolerance designers and process planners.     

The role of impressions on users’ tactile interaction with product materials: An analysis of associative concept networks

Depth impressions are an inner associative layer of humans’ expressed impressions. To analyze tactile interaction, it is essential to examine what users feel and imagine and how they create depth impressions by touching and looking at different product materials. On the basis of tactile interactions, this study aims to capture and analyze users’ depth impressions of materials. This research also proposes an ‘impressionably’ new tactile material for design from the viewpoint of depth impressions. To capture depth impressions, we investigated users’ tactile interactions in an experiment. The experiment used samples of six common natural and artificial materials, along with the proposed new micro-print-based material. A concept network-based method was employed in two stages to analyze the experimentally obtained verbalized protocols and to identify any depth impressions. This method allowed us to capture and analyze the depth impressions behind the surface impressions. This research found that the feel of materials’ tactile naturalness and users’ habituation to the tested samples are related to their depth impressions and the complexity of their concept networks. The depth impressions and concept network of the proposed micro-print material are distinct and beyond those for existing natural or artificial materials. These findings will provide the basis for employing new analysis tools and facilitate the development of impressionably better tactile materials for design.