Simon Haykin. Neural Networks. A Comprehensive Foundation. Prentice Hall, New Jersey, USA, second edition, 1999.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....USA [payman,gray] jpl.nasa.gov ) 1 1 # = NC NC NN NN ants (1) where NN is the number of nodes in the network and NC is the number of colonies. The value of NC minimizing the number of ants is NN NC = From the available techniques in the literature, we choose k means clustering [14], where the distance function is the Euclidean distance between the nodes. Clustering is performed by a central controlling entity that is aware of the geographical locations of all the nodes. B. Discovering Routes After the network colonies have been formed, two types of agents, called ants, ....

S. Haykin, Neural Networks, A comprehensive foundation, Prentice Hall, 1994.

....face veri cation has been addressed by di erent researchers and with di erent methods. The aim of this section is not to propose a new model for face veri cation, but to present the model used to evaluate the new feature set. Our face veri cation method is based on Multi Layer Perceptrons (MLPs) [2, 8]. For each client, an MLP is trained to classify an input to be either the given client or not. The input of the MLP is a feature vector corresponding to the face image with or without its skin color. The output of the MLP is either 1 (if the input corresponds to a client) or 1 (if the input ....

S. Haykin. Neural Networks, a Comprehensive Foundation, second edition. Prentice Hall, 1999.

....can compute the class frequencies of mobile cells. Using these histograms as data to a second level SOM we get a SOM of histograms. The topology of the new SOM is 2D rectangular grid. Grid of size 8 x 8 nodes has been used as in 4.2. The BMU search of the map is based on Kullback Leibler distance [4]. The KullbackLeibler distance or relative entropy between two probability distributions pX (x) and q X (x) is defined by D p q = # x#X pX (x) log( pX (x) q X (x) 4) where the sum is over all states of the system (i.e. the alphabet X of the discrete random variable X) The ....

S. Haykin. Neural Networks, a Comprehensive Foundation. Macmillan, 1999.

....task. 2.2 MLPs for Hand Posture Recognition A Multi Layer Perceptron (MLP) is a particular architecture of Arti cial Neural Networks. Arti cial Neural Networks are learning machines used in many classi cation problems. A good introduction to machine learning algorithms can be found in [1, 6]. We will assume that we have access to a training dataset of l pairs (x i ; y i ) where x i is a vector containing the pattern, while y i is the class of the corresponding pattern often coded respectively as 1 and 1. An MLP is composed of layers of non linear but di erentiable parametric ....

S. Haykin. Neural Networks, a Comprehensive Foundation, second edition. Prentice Hall, 1999.

....experimental results from another campaign. For the campaign AD3, we sent emails to randomly selected customers. With the remaining customers not in the random set, we selected customers by three different targeting methods: collaborative fil tering (CF) 5] artificial neural network (ANN) [6], and CGP LS. Collaborative filtering is a proven stan dard for personalized recommendations [7] 10] A rep resentative company using collaborative filtering is NetPerceptions, Inc. Artificial neural networks have been used to solve a variety of problems in optimization, pattern recognition, ....

....using collaborative filtering is NetPerceptions, Inc. Artificial neural networks have been used to solve a variety of problems in optimization, pattern recognition, prediction, function approximation, etc. An ANN learns from its environment through an interactive process of adjusting its weights [6][9] Table 5 shows the results. CGP LS showed a 4.78 response rate while ANN and CF showed 4.44 and 4.00 , respectively. This represents a 42.7 improvement over the random targeting. Table 4: Results of Our Approaches (1) Result of AD1 Training data Test data Trials S E(A) E(U) R R o (A) ....

S. Haykin. Neural Networks, A Comprehensive Foundation. Prentice Hall, 1975.

....of frequency vectors or class histograms. The topology of the new SOM is 2D rectangular grid. Grid of size 9 x 9 nodes has been used for microcell scenario and grid of size 8 x 8 nodes has been used for macrocell scenario. The BMU search and the training of the map use Kullback Leibler distance [5]. The Kullback Leibler distance or relative entropy between two probability distributions pX (x) and q X (x) is defined by D pjjq = x2X pX (x) log( pX (x) q X (x) 2) where the sum is over all states of the system (i.e. the alphabet X of the discrete random variable X) In Fig. 4 the ....

S. Haykin. Neural Networks, a Comprehensive Foundation. Macmillan, 1999.

....figure 1 with sensor radius r a , e.g. r a = 3. 3.2 The Neuro Controller (NC) A neuro controller gives the possible directions for the agent to move through the world. As seen in figure 2, the neuro controller consists of a multiple layer perceptron as proposed by [11] with a winner takes all ([12]) layer at the output. Hence exactly one output unit responds to the input. The input and hidden units transfer their weighted inputs by a logistic function. The inputs for the neuro controller are polar coordinates indicating the position of a randomly chosen agent a j within r a of a i , with ....

Simon Haykin. Neural Networks. A Comprehensive Foundation. MacMillan, 1994.

....and symbolic reasoning [2] The continuous input output pairing permits the use of traditional supervised gradient methods [1] On the other side, the recurrent connections prevent the computation of the exact error gradient for infinite se quences. Back propagation through time (BPTT, [3]) is a gradient training technique that can compute the error gradient for finite sequences only. Some experiments have used BPTT splitting the input sequence in subsequences both with SRNs [1] and RAAMs [4] as in [5, 6, 2, 1] Reducing the input stream in sub sequences requires a certain ....

....back propagation through time (BPTT) truncated BPTT with ap proximation for small weight change (aBPTT n ) and standard BPTT n (BPTT n ) 2. A reactive ANN (R) without recurrent connections, has been trained for performance comparison. For a reference on the different gradient methods see [13, 14, 3]. 2.2. Environmental settings Each input bit is represented by a floating number, a high bit is represented by 0.5 and a low bit by 0.5. In the output layer a positive value represents a high bit. White noise 9 [0.05, 0.05] is added to the inputs to increase stability. The nodes use a hyperbolic ....

S. Haykin, Neural Networks. A Comprehensive Foundation, Prentice-hall, New Jersey, second edi-

....highly parallel distributed architecture and the ability to learn based on limited data and generalize to different inputs outputs makes ANN a powerful computing tool. In neural networks, the relations are not explicitly given, but are coded in a network and its weights. Their main advantages are [11, 10]: Nonlinearity. All physical mechanisms are nonlinear. The mathematical inter connection of the ANN structure may provide nonlinear characteristics to the function or process. Input Output Mapping. The identification paradigm used is based on nonparametric statistical inference, where no ....

....way, where feedback is introduced internally to the inputs of antecedent neurons, or in a feedforward way, where information flows only in one direction, as shown in figure 2. A dynamic network can be constructed by using a static feedforward network combined with an external feedback connection [10, 11]. A multi layer neural network, consisting of one input layer, one output layer and an appropriate number of hidden layers, may be used either as a static or dynamic approximator. It was proven by Cybenko [5] that one hidden layer is sufficient to achieve the desired accuracy, given a sufficient ....

[Article contains additional citation context not shown here]

Simon Haykin. Neural Networks, A Comprehensive Foundation. Prentice Hall, New Jersey, USA, second edition, 1999.

....by the human brain, as we are able to generate higher order concepts from neural substrates) Besides the opacity of a trained ANN, there are other open research issues concerning the ANN design phase. For the construction of ANN architectures for real world problems only rules of thumb exist [2], and the experience of a human designer is still the most important factor for an appropriate choice. However, due to the exponential nature of this design problem evolutionary design of ANN architectures [3, 4] has been found to be a valuable alternative and usually improves the overall ....

Simon Haykin. Neural Networks. A Comprehensive Foundation. MacMillan, 1994.

....the network weights through a gradient descent method following an error correction strategy. In a MFNN this operation corresponds to the minimization of network s learning error. In the presented experiments training is performed with the on line version of the momentum back propagation algorithm [11]. The incorporation of momentum represents a minor modification to the weight update, yet it may have some beneficial effects on the learning behavior of the algorithm [11] After training, the MFNN is able to discriminate between normal and abnormal texture regions by forming hyperplane decision ....

....error. In the presented experiments training is performed with the on line version of the momentum back propagation algorithm [11] The incorporation of momentum represents a minor modification to the weight update, yet it may have some beneficial effects on the learning behavior of the algorithm [11]. After training, the MFNN is able to discriminate between normal and abnormal texture regions by forming hyperplane decision boundaries in the pattern space. Optimal generalized learning can be achieved by a MFNN structure that has the minimal number of neurons necessary to map the exemplar ....

S. Haykin, Neural Networks.' A Comprehensive Foundation, 2 nd ed., Prentice Hall, New Jersey, 1996.

....areas in images by examining the discrimination abilities of their K level wavelet coefficients based features. Besides neural network classifiers and the K Level 2 D wavelet transform, the tools utilized in such an analysis are vector quantization and Principal Component related analysis [4] of the vectors quantizing the K Level wavelet domain of an image window. The problem at hand can be clearly viewed as image segmentation one, where the image should be segmented in defective and non defective areas only unlike its conventional consideration. Concerning the classical segmentation ....

.... domain [6,5] These multidimensional features, coming from the application of the K Level 2 D DWT, are, in the sequel, processed using vector quantization and PCA methodology, which offer the accurate tools for describing transformed image characteristics and especially complex second order ones [4]. More specifically, PCA of the autocorrelation matrices analysis is well known to provide second order information about pixel intensifies, while Vector Quantization algorithms provide the means for efficient vector space encoding. Two are the main stages of the suggested system. Namely, ....

[Article contains additional citation context not shown here]

Haykin, S.: Neural Networks, A comprehensive foundation, Prentice Hall, Second edition, 1999.

....of appropriate sized weight matrices A; B; C; D [8] Delta Gamma y t Gamma y Delta 2 min A;B;C;D (3) For an overview of algorithmic solution techniques see [6] We solve the system identification task of Eq. 3 by finite unfolding in time using shared weights. For details see [3, 8]. Fig. 1 depicts the resulting neural network solution of Eq. 3. z t 2 D C D C D t 1 s z t 1 t s z t s t 2 s t 3 t 3 y t 2 y s t 1 t 1 y A y t 2 d t 2 u y t 1 y B u t u t 1 A A A B Id Id Id B Figure 1. Error correction neural network (ECNN) using unfolding in time and ....

Haykin S.: Neural Networks. A Comprehensive Foundation., 2 nd ed., Macmillan, N. Y. 1998.

....of the algorithm in a real time control architecture impractical. The use of a Neural Network (NN) to generalize among states and actions reduces the number of values stored in the Q function table to a set of NN weights. The implementation of a feedforward NN with the backpropagation algorithm [6] is known as direct Q learning [3] The Direct Q learning algorithm has no convergence proofs and turned out to be unstable when we tried to learn a behavior. The instability was caused by the lack of weight updating in the whole state action space. The optimal Q function was only learnt in the ....

S. Haykin, Neural Networks, a comprehensive foundation. Prentice Hall, 2 nd ed., 1999.

No context found.

Simon Haykin. Neural Networks. A Comprehensive Foundation. Prentice Hall, New Jersey, USA, second edition, 1999.

No context found.

Haykin, S.: Neural Networks. A Comprehensive Foundation. Second edn. Prentice Hall, New Jersey, USA (1999)

No context found.

S. Haykin. Nueral networks, a comprehensive foundation. Macmillan, 1994.

No context found.

S. Haykin. Neural Networks, A Comprehensive Foundation. Macmillan, New York, 1994. 24

No context found.

HayKin,S. Neural Networks. A comprehensive foundation. Prentice Hall , Ontario, Canada, 1999.

No context found.

S. Haykin. Neural Networks, A Comprehensive Foundation. Prentice Hall, 2nd edition, 1999.

No context found.

S. Haykin, Neural Networks, A Comprehensive Foundation, Prentice Hall, 2 edition, 1999.

No context found.

S. Haykin. Neural Networks. A comprehensive Foundation. Prentice Hall, second edition, 1998.

No context found.

Haykin S.: Neural Networks, a Comprehensive Foundation. second edition, Prentice-Hall, Englewood Cli#s (1999).

No context found.

S. Haykin, Neural Networks, a Comprehensive Foundation, second edition, Prentice Hall, 1999.

No context found.

S. Haykin. Neural Networks, a Comprehensive Foundation, second edition. Prentice Hall, 1999.

First 50 documents  Next 50

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