Wasserman, P. D., Advanced Methods in Neural Computing, New York: Van Nostrand Reinhold, 1993.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....value while x a denotes its corresponding attention value. 1 1 a xx xe = 1) Moreover, the values of binary attributes are mapped to 0.1 and 0.9, and the values of ternary attributes are mapped to 0.1, 0.5, and 0.9. Those mappings could speed up the converging of the neural networks [13]. 5. Experiments The described system is tested on a data set comprising 704 cases, among which 528 cases are used as the training set while the rest 176 cases are used as the test set. The data set is provided by the Institute of Electric Science of Shandong Province, P.R.China. All the cases ....

....Shandong Province. The neural networks are trained with SuperSAB algorithm [12] which is one of the fastest variations of Backpropagation. Tollenaere [12] reported that it is 10 100 times faster than standard BP [11] The parameters of SuperSAB are set to the values recommended by Wasserman [13], i.e. the weight increasing factor up is set to 1.05, the weight reducing factor down is set to 0.2, and the upper ground of the maximum step of the k th weight ij k is set to 10. In the experiments, the visualization results are checked by a junior human monitor. If he cannot identify ....

Wasserman P. D.: Advanced Methods in Neural Computing. Van Nostrand Reinhold, New York (1993).

....defined limit properties. Another approach to classification problem could be based on classdependent kernels, which accentuate the regions in TFD [12] where the maximum difference between the classes of the signal occurs, or the use of adaptive neuro fuzzy interference system with clustering [11]. An extension of this work could be the design of a classifier with more than two classes handling different fault mode operations of the inverter drive. 7. CONCLUSIONS In this paper a new method of classification of electric signals was presented, based on the timefrequency representation and ....

Wasserman, P. D., Advanced Methods' in Neural Computing, New York: Van Nostrand Reinhold, 1993.

....and falling into local minimums is often encountered. This has stimulated deeper research into theories and models of ANNs in this phase. In this paper, a new neural leaming model named FANNC, which organically exploits the advantages of both adaptive resonance theory (ART) 3] and field theory [4], is proposed. FANNC is designed to deal with classification tasks. It needs only one pass learning, and achieves not only high predictive accuracy but also fast learning speed. The learning course of FANNC is performed in an incremental style. When new instances are fed, it does not retrain the ....

.... Field Theory Field theory is named from CPM (Coulomb Potential Model) 13] But some researchers had already investigated these kinds of model before Bachmann et al. 14] Those fruits are called field theory methods at present, as well as neural algorithms belonging to this kind developed later [4]. Field theory is a kind of relaxation model, i.e. its algorithm minimizes an energy function. In one of its many forms, the test instance is represented as a freely moving positive charge, analogous to a test charge in electrostatic theory. The test charge is allowed to move under the influence ....

[Article contains additional citation context not shown here]

P. D. Wasserman. Advanced Methods in Neural Computing, Van Nostrand Reinhold, New York, 1993.

....from automated rule evaluation. Therefore in most cases the extracted rules are manually evaluated by knowledgeable human experts. 3. Fast Neural Classifier FANNC Fu,qc [31] is a fast adaptive neural classifier that exploits the advantages of both Adaptive Resonance Theory [32] and Field Theory [33]. It performs one pass learning and achieves not only strong generalization ability but also fast learning speed. Besides, Fu,n, lc has incremental learning ability. When new training examples are fed, it does not retrain the entire training set. Instead, it can team the I second layer competition ....

D. Wasserman, Advanced Methods in Neural Computing, NY: Van Nostrend Reinhold, 1993.

....(ART) 1] is an important family of competitive neural learning model. Its memory mode is very similar to that of biological one, and memory capacity can increase while the learning patterns increase. It can perform real time online learning, and can work under nonstationary world. Field Theory [2] is named from CPM (Coulomb Potential Model) 3] It can perform real time one pass supervised learning with fast speed, and no spurious responses will be produced regardless of the number of memories stored in the network. We have proposed a neural network classifier based on ART and Field ....

....connected with all the input units through Gaussian weights. The response centers are respectively set to the input components of current instance, and the response characteristic widths are set to a default value. FANRE introduces the notion of attracting basin, which is proposed in Field Theory [2]. Each second layer unit of FANRE defines an attracting basin by responsecenters and response characteristic widths of Gaussian weights connecting with it. Thus, FANRE constructs its first attracting basin according to the first instance. And it will add or move basins according to the later ....

Wasserman P D. Advanced Methods in Neural Computing. Van Nostrand Reinhold, New York, 1993.

....Through exploiting these two technologies, we successfully achieve satisfactory experimental results. 2. Fast neural model The neural model we used is called FTART (Field Theory based Adaptive Resonance Theory) 2] which is proposed based on Adaptive Resonance Theory [3] and Field Theory [4]. Figure 1 shows its architecture. The initial network is composed of only input and output units, and hidden units will be adaptively appended during the training phase. Thus, the disadvantage of manually configure hidden units of traditional feed forward neural networks is overcome. The ....

Wasserman P D, Advanced Methods in Neural Computing, Van Nostrand Reinhold Press, New York, 1993.

....equivalent; this does not imply that these systems can be implemented where necessary. On the other hand, the associative machine structure, is a general structure to carry out many of the most common tasks in neurocomputation: pattern recognition and completion, classification, filtering, etc. [4]. In this paper we propose a novel model of information representation, explored in a previous work [2] which is used in an associative machine that operates with 2 D qualities, allowing to receive directly this kind of stimuli. Both the input stimuli and the desired output, are processed without ....

P. Wasserman, Advanced Methods in Neural Computing, VNR, 1993.

....a learning process. The method of learning varies greatly depending on the implementation of the neural network. No learning model is ideal and most have some limitations. For this reason, network training algorithms continue occupy more research hours than any other aspect of neural networks (Wasserman 1993). On a fundamental level 65 though, all learning models must present a group of closely related training problems to the network to commence the learning process. In doing so, this process follows the models of human cognition where individuals are exposed to a number of problems to increase ....

....Among the supervised learning techniques, the error backpropagation training method continues to be the most success technique for training multiple layer neural networks. It is estimated that over 85 percent of the neural network applications employ some form of an error backpropagation training (Wasserman 1993). The widespread implementation is primarily attributed to a theoretically sound technique for weight adjustment. The discovery of backpropagation training by Rumelhart, Hinton, and Williams in 1986 accelerated the growth of activity in the neural network field. Prior to this discovery, only ....

[Article contains additional citation context not shown here]

Wasserman, P. (1993). Advanced Methods in Neural Computing. New York: Van Nostrand Reinhold.

....(or data set) to a standard unit length, usually 1. The length of the vector is computed and each vector component is divide by its length. A vector or data set can be normalised across many different dimensions, and with respect to many different statistical measures such as the mean or variance [Wasserman93]. Furthermore it is necessary to re scale all the inputs (normally between 0 and 1 or between 1 and 1) This should be done so that initially all the input variables have the same importance. During the learning phase the network will deliberately alter the importance of variables, by changing ....

P. D. Wasserman, "Advanced Methods in Neural Computing", New York: Van Nostrand Reinhold, 1993.

....the full RBF architecture parameters: the centers, the radii, and forward weights. This is done after those parameters were found via clustering and a pseudo matrix inversion. Performance comparison on several benchmark data sets is given. Comparison includes a conventional RBF implementation 1 [25], Bishop s EM implementation [14] and Orr s regression trees approach [23] II. A modified clustering algorithm A Radial Basis Functions approximation network (RBF) is composed of a set of kernel functions # located at cluster centers m i in input space with a width r. We use Gaussian kernels, ....

P. D. Wasserman, Advanced Methods in Neural Computing, Van Nostrand Reinhold, New York, 1993.

....these methods. Eventually we will discuss the parameters (virtually none) and the complexity of SCG. 9. 17.1 Conjugate Gradient Methods (CGMs) They are general purpose second order techniques that help minimize goal functions of several variables, with sound theoretical foundations [P 88, Was95] Second order means that these methods make use of the second derivatives of the goal function, while first order techniques like standard backpropagation only use the first derivatives. A second order technique generally finds a better way to a (local) minimum than a first order technique, but ....

Philip D. Wasserman. Advanced Methods in Neural Computing. Van Nostrand Reinhold, 1995. 34

....training used on a 20,000 instance speech recognition task show a roughly linear increase in training time required with an increase in batch size. 1. Introduction Multilayer Perceptrons (MLPs) are often trained using error backpropagation (BP) Bishop, 1995; Rumelhart McClelland, 1986; Wasserman, 1993). MLPs consist of several layers of nodes, including an input layer, one or more hidden layers and an output layer. The activation of the input nodes are set by the environment, and the activation A j of each node j in the hidden and output layers is set according to the equations: A j = f (S j ) ....

Wasserman, Philip D., (1993). Advanced Methods in Neural Computing, New York, NY: Van Nostrand Reinhold.

....function to be successful, including the instance based learning algorithms and the related models mentioned in the introduction. In addition, many neural network models also make use of distance functions, including radial basis function networks (Broomhead Lowe 1988; Renals Rohwer 1989; Wasserman 1993), counterpropagation networks (Hecht Nielsen 1987) ART (Carpenter Grossberg 1987) self organizing maps (Kohonen 1990) and competitive learning (Rumelhart McClelland 1986) Distance functions are also used in many fields besides machine learning and neural networks, including statistics ....

Wasserman, Philip D. 1993. Advanced Methods in Neural Computing, New York, NY: Van Nostrand Reinhold.

....a neural network for systems identification and control has been experimented [17] since the early revival of this computational paradigm in [51] To ground our design approach, we report in table 1 a categorisation of existing neurocontrol algorithms. The table, beside other well known taxonomies [54 ,78, 115, 136, 145, 146], focuses on the way of processing the data and the formal knowledge on the plant, more than on technical properties of the algorithms, such as convergence, stability, learning technique, etc. In this sense we refer to a grey box , where at time t an input u(t) is given to the box and an output ....

P. D. Wasserman, Advanced Methods in Neural Computing", chapter 7, Van Nostrand Reinhold, NY, 1993.

....original. It is therefore better to consider them as computational paradigms that can solve hard problems. 2.1 Brief overview of Neural Networks There is plethora of bibliographic references in the subject of Neural Networks. Nevertheless, the following can be suggested: HKP91] Hay94] and [Was93]. A Neural Network or to be more accurate an Artificial Neural Network can be considered, from an application point of view, as a computational device that overcomes the deficiencies of the serial digital computers 1 . To say that an ANN (Artificial Neural Network) mimics the brain, it would be ....

....2 for the iris data set. The compiler used was: gcc v2.6, with optimiser version O2 turned on. The machine upon which the experiments were performed was an HP 9000 715 50. 1 I am referring to Setosa, Versicolor and Virginica 2 A brief overview of the notion of classifier can be found in [Was93] Iris Genetic Algorithm The SANE method, with binary encoding of the genotype in the chromosomes has been used. It evolves the links and the weights. ffl The algorithm is let to run for 500 generations. It might stop earlier in case it finds a network which learns 100 of the training data. ffl ....

P. D. Wasserman. Advanced Methods in Neural Computing. Van Nostrand Reinhold, 1993.

....only the means of these Gaussian activation functions are adaptively tuned; the scales are set after training of the means using some heuristic. Most of these heuristics make use of distance from the mean to K nearest means [2, 3] But also more advanced algorithms are known to tune the scales [4, 5, 6]. We do not use radial basis function networks, but we propose variants of vector quantization algorithms. 3 Vector Quantization Assume given training patterns x 2 R D and a much smaller number of codebook vectors m i 2 R D . The codebook vectors are given random initial values. Repeatedly ....

P.D. Wasserman. Advanced Methods in Neural Computing. Van Nostrand Reinhold, 1993.

....mean value will damage the network performance and or even render the data meaningless. Hence, it is critical to determine the significance of the characteristics of the data set, before any normalisation procedure can be carried out. On the importance of correct normalisation, Wasserman says [Was93, p.235] Normalisation is a complicated topic. It can take many forms and has many hazards; it is all too easy to inadvertently remove the precise information required for accurate operation. On the other hand, correct normalisation can transform a poorly performing network into one that is ....

....the precise information required for accurate operation. On the other hand, correct normalisation can transform a poorly performing network into one that is nearly error free. Therefore, careful selection of normalisation techniques may pay large dividends in the success of a project. Wasserman [Was93, pp.235 242] proposed 5 types of normalisation methods: 1. Removing the mean 2. Normalising vector magnitudes 3. Local normalisation for multimodal distribution 4. Detrending data 5. Nonlinear normalisation CHAPTER 2. LITERATURE REVIEW 29 1. Removing the mean Significant information of a data ....

[Article contains additional citation context not shown here]

Philip D. Wasserman. Advanced Methods in Neural Computing, chapter 6, 11. Van Nostrand Reinhold, New York, 1993.

No context found.

D.Philip and Wasserman, Advanced Methods in Neural Computing (Van Nostrand Reinhold, New York, 1993).

No context found.

Wasserman, P. D., Advanced Methods in Neural Computing, New York: Van Nostrand Reinhold, 1993.

No context found.

WASSERMAN, P. D. "Advanced Methods in Neural Computing," (Van Nostrand Reinhold, New York, 1993).

No context found.

P. D. Wasserman, Advanced Methods in Neural Computing. New York: Van Nostrand Reinhold, 1993.

No context found.

P. Wasserman, " Advanced Methods in Neural Computing ", Van Nostrand Reinhold, ( 1993 ).

No context found.

P. D. Wasserman. Advanced Methods in Neural Computing. Van Nostrand Reinhold, New York, 1993. Neural Engineering, Chapter 11. ISBN: 0-442-00461-3.

No context found.

Wasserman, P. D., Advanced Methods in Neural Computing, New York: Van Nostrand Reinhold, 1993.

No context found.

P. D. Wasserman, Advanced Methods in Neural Computing, Van Nostrand Reinhold, NY, 1993.

First 50 documents

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC