Platt, J. (1991). A resource-allocating network for function interpolation. Neural Computation, 3(2), 213--225.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....is remarkable. B. Unrealizable Case In the previous experiment, the learning target function f belongs to SN . Here, we perform a simulation for an unrealizable case that f does not belong to SN . Let us consider the chaotic series created by the Mackey Glass delay di#erence equation [10]: g(t 1) # (1 b)g(t) a g(t 1 g(t 10 for t # 1, 0.3 for 0 #, 24) where a = 0.2, b = 0.1, and # = 17. Let t=1 be h t = g(t # 1) 25) We are given M degraded sample values ym = h r(m) # m , 26) where r(m) is an integer such that 1 600. r(m) ....

J. Platt. A resource-allocating network for function interpolation. Neural Computation, 3(2):213--225, 1991.

....Taylor expansion to an error function [8] 12] Parallel to regularization based methods, cluster driven methods control network complexity by sequentially adding or removing clusters. A sequential network growth algorithm provides a systematic way to add clusters responding to new data features [13]. The convergence rate of algorithm is enhanced by applying the extended Kalman filter (EKF) algorithm to adjust network parameters [14] Further generalization improvement is obtained as strategies of cluster pruning are combined with network growth criteria [15] The above methods indirectly ....

J. Platt, "A resource-allocating network for function interpolation," Neural Comput., vol. 3, pp. 213--225, 1991.

....is critical and is usually achieved using the whole data set or a clustering technique. The convergence of gradient descent algorithms is guaranteed only in the case of o# line learning. Empirically, an on line version appear to converge reasonably well. Dynamic versions were presented by Platt [67] and, combined with an on line clustering algorithm by Fritske [32] As noted in the introduction, most of the medical applications considers RBFNs as ANNs. The melanoma diagnosis scenario provides a natural application for ANNs (see references in [10] 12 5 RBFNs as Regularization Networks In ....

....learning frameworks, adopt the technique of initializing the centres with the samples of the training set and performing almost no other computation during the training phase. Hence, it is straightforward to add a new function when a new sample is available. Unfortunately, as noted by Platt [67] referring to Parzen Windows and k NN, the two mentioned algorithms present a drawback. The resulting RBFN grows linearly with the number of the samples. Platt also proposed a neural architecture, called Resource Allocation Networks, which combines an on line gradient descent with a method for ....

J. Platt. A resource-allocating network for function interpolation. Neural Computation, 3:213--225, 1991.

....prediction, exceed conventional methods by orders of magnitude in accuracy [Lapedes and Farber, 1987] 1.1 ANN prediction Ensuing the pioneering work of [Lapedes and Farber, 1987] substantial research has been carried out to utilize ANNs for the prediction of nonlinear and chaotic time series. Platt (1991) already observed the importance of the ANN architecture for accurate prediction and proposed a growing scheme using neurons with Gaussian activation function. In this procedure a neuron is added, whenever a new training pattern causes poor prediction of the network, hereby adapting the network ....

....of the ANN architecture for accurate prediction and proposed a growing scheme using neurons with Gaussian activation function. In this procedure a neuron is added, whenever a new training pattern causes poor prediction of the network, hereby adapting the network topology to the specific problem [Platt, 1991]. A network growing scheme based on [Moody and Utans, 1994] is used by Cholewo and Zurada (1997) for the construction of instances of time delay and recurrent ANNs for time series prediction [Cholewo and Zurada, 1997] Here, neurons are added sequentially up to a preset limit in order to improve ....

[Article contains additional citation context not shown here]

Platt, J. (1991). A Resource--Allocating Network for Function Interpolation. Neural Computation, 3(2):213--225.

....asymmetrically weighted similarity metric (LASM) approach for data compression [9] is shown to have superior performance compared to conventional k NN classification based methods. Similar concepts of data reduction and locally varying models based on neural networks are discussed in [10] 11] [12]. The classification based condensation methods are, how ever, specific to (i.e. dependent on) the classification tasks and the models (e.g. k NN, perceptron) used. Data condensation of more generic nature is performed by classical vector quantization methods [13] using a set of codebook ....

J. Platt, "A Resource-Allocating Network for Function Interpolation, " Neural Computation, vol. 3, pp. 213-255, 1991.

....rule extraction and training methods. 18] introduced the GenTower algorithm which, inspired by the Tower algorithm, adds small subnetworks to the neural network during training. Another group of algorithms which start with a small network and successively add new units has been introduced by [31]. Building on the same idea are Fritzke s growing cell structures [12] where units are added by evaluating local statistical measures gathered during previous adaption steps and the network dimensions are preserved. In growing neural gas [11] the network topology is generated incrementally by ....

J.C. Platt. A resource-allocating network for function interpolation. Neural Computation, 3(2):213--225, 1991.

.... (3) dynamically create new modules have open structure; 4) memorise information that can be used at a later stage; 5) interact continuously with the environment in a life long learning mode; 6) deal with knowledge (e.g. rules) as well as with data; 7) adequately represent space and time [2,5,35,36,38,61,66,71]. Developing a computational model called evolving fuzzy neural network (EFuNN) that meets the seven requirements above is the objective of the current paper. Several methods and systems have been developed so far that meet some of the criteria above and that have influenced the development of ....

....above is the objective of the current paper. Several methods and systems have been developed so far that meet some of the criteria above and that have influenced the development of EFuNNs. These are methods and systems for: adaptive learning [4,5,7,8,14,30,46,47,48] incremental learning [6,7,8,9,19,53,58,61,71]; lifelong learning [69,35,36,82] on line learning [17,21,22,28,31,35,36,42,44,61,66,67,69] constructivist structural learning [15,19,11,14,9] that is supported by biological facts [14,62,73,77,82] selectivist structural learning [26,29,49,56,59,64,50,32] hybrid constructivist selectivist ....

[Article contains additional citation context not shown here]

Platt, J., "A resource allocating network for function interpolation, Neural Computation, 3, 213-225 (1991)

.... (3) dynamically create new modules have open structure; 4) memorise information that can be used at a later stage; 5) interact continuously with the environment in a life long learning mode; 6) deal with knowledge (e.g. rules) as well as with data; 7) adequately represent space and time [2,5,35,36,38,61,66,71]. Developing a computational model called evolving fuzzy neural network (EFuNN) that meets the seven requirements above is the objective of the current paper. Several methods and systems have been developed so far that meet some of the criteria above and that have influenced the development of ....

....fuzzy neural network (EFuNN) that meets the seven requirements above is the objective of the current paper. Several methods and systems have been developed so far that meet some of the criteria above and that have influenced the development of EFuNNs. These are methods and systems for: adaptive [6,7,8,9,19,53,58,61,71]; learning [4,5,7,8,14,30,46,47,48] incremental lifelong learning [69,35,36,82] on line [17,21,22,28,31,35,36,42,44,61,66,67,69] constructivist structural [ 15,19,11,14,9] that is supported by biological facts [ 14,62,73,77, 82] selectivist structural learning [26,29,49,56,59,64,50,32] ....

[Article contains additional citation context not shown here]

Platt, J., "A resource allocating network for function interpolation, Neural Computation, 3,213-225 (1991)

....of creating fuzzy inference rules and evolving a fuzzy system. For the similar purpose, the ECM can also be fuzzified and applied to the EFuNN, Evolving Fuzzy Neural Network [3] Although the ECM suits on line, dynamic systems it can also be applied to off line tasks, e.g. the Kmeans clustering [4] can get initial value made by an ECM more effective than made by using of a random method. In the off line cases, a constrained optimisation is applied to the ECM, which makes a pre defined objection function, based on a distance measure, to reach a minimum value subject to the given ....

Platt, J., "A Resource Allocating Network for Function Interpolation", Neural Comp., 3, 213225, 1991.

....of creating fuzzy inference rules and evolving a fuzzy system. For the similar purpose, the ECM can also be fuzzified and applied to the EFuNN, Evolving Fuzzy Neural Network [3] Although the ECM suits on line, dynamic systems it can also be applied to off line tasks, e.g. the Kmeans clustering [4] can get initial value made by an ECM more effective than made by using of a random method. In the off line cases, a constrained optimisation is applied to the ECM, which makes a pre defined objection function, based on a distance measure, to reach a minimum value subject to the given ....

Platt, J., "A Resource Allocating Network for Function Interpolation", Neural Comp., 3, 213- 225, 1991.

....require sophisticated methods and tools for building on line, adaptive intelligent systems. Such systems should be able to grow as they operate, to update their knowledge and refine the model through interaction with the environment, that is the systems should have an ability of on line learning [1, 14, 11]. On line learning is concerned with learning data as the system operates (usually in a real time) and the data might exist only for a short time. Here we propose a model called dynamic evolving neural fuzzy inference system (DENFIS) DENFIS is similar to EFuNN (evolving fuzzy neural systems ....

....clustering method [4] is also presented in this section. Section 3 introduces the DENFIS model, and in section 4, DENFIS model is applied to Mackay Glass time series [3, 5] prediction problem. Results are compared with the results obtained with the use of resource allocation network (RAN) [14], evolving fuzzy neural networks (EFuNNs) and evolving self organising maps (ESOM) 6] Conclusions and directions for further research are presented in the final section. The analysis of results indicates clearly the advantages of DENFIS when used especially for on line learning applications. ....

[Article contains additional citation context not shown here]

Platt, J., "A Resource Allocating Network for Function Interpolation", Neural Comp., 3, 213-225, 1991.

....networks and low error patterns must not be extrapolated to more general situations. Krusche[8] has pointed out that the huge receptive field of the weighted sum units is responsible for interference in neural networks. Units with a limited receptive field are increasingly being used [9] 10] [11]. Locally tuned units that use radial basis functions (RBF) are the common choice. This can be a valid solution, but an important drawback of RBF units is that they need many more examples than weighted sum units to generalize well, especially in high dimensional input spaces. The problem comes ....

J. Platt, "A Resource-Allocating Network for Function Interpolation", Neural Computation, Vol. 3, No. 2, 1989.

....mode corresponds to a component) and only approximate the time frequency regions they occupy, it is important to improve the approximation of the number of the modes. Many of the techniques designed to adapt the number of Gaussians comprising a mixture model are based on pruning [19] or growing [17, 16] the model, but both approaches generate models wherein the number of Gaussians does not truly reflect the complexity of the underlying distribution. 3.2 Determining the model order: Functional merging An alternative technique for adapting the number of Gaussians is functional merging [21] This ....

J.C. Platt. A resource allocating network for function interpolation. Neural Computation, 3:213--225, 1991.

....is critical and is usually achieved using the whole data set or a clustering technique. The convergence of gradient descent algorithms is guaranteed only in the case of off line learning. Empirically, an on line version show to converge reasonably well. Dynamic versions were presented by [Platt, 1991] and, combined with a on line clustering algorithm by [Fritzke, 1994b] 34 2.5 Statistical Approach The architecture of the RBFNs presents a strong similarity with the regression techniques, based on non parametric estimation of an unknown density function [Scott, 1992] and with the ....

....was introduced. The dynamic algorithms modify the number of basis functions of the network, integrating the actions occurring in the initialization and in the refining phases in an incremental learning algorithm. The first work on this direction is the Resource Allocation Networks proposed by Platt [Platt, 1991]. Other works that introduced structural changes [Fritzke, 1994a, Fritzke, 1995] or structural selection criteria [Kadirkamanathan and Niranjan, 1993, Orr, 1995] were subsequently proposed. In this chapter, the algorithms mentioned above, will be briefly reviewed. More attention is devoted to two ....

[Article contains additional citation context not shown here]

Platt, J. (1991). A resource-allocating network for function interpolation. Neural Computation, 3:213--225.

....evolving system. Although the method presented is unsuccessful, some useful information about the properties of the ECoS model is still derived from the work. 1 Introduction There are several methods that have been introduced to optimise the parameters of neural networks (NN) such as MLP, RBF [5] fuzzy neural networks and others [7] These NN have more or less a fixed structure (except the ones that have their number of hidden nodes optimised as a parameter) deal with off line, batch mode learning, apply many iterations for learning and optimisation. These systems despite their ....

John Platt. A resource-allocating network for function interpolation. Neural Computation, 3(2):213--225, 1991.

....evolving system. Although the method presented is unsuccessful, some useful information about the properties of the ECoS model is still derived from the work. 1 Introduction There are several methods that have been introduced to optimise the parameters of neural networks (NN) such as MLP, RBF [5] fuzzy neural networks and others [7] These NN have more or less a fixed structure (except the ones that have their number of hidden nodes optimised as a parameter) deal with off line, batch mode learning, apply many iterations for learning and optimisation. These systems despite their ....

John Platt. A resource-allocating network for function interpolation. Neural Computation, 3(2):213--225, 1991.

No context found.

Platt, J. (1991). A resource-allocating network for function interpolation. Neural Computation, 3(2), 213--225.

No context found.

Platt, J. (1991). A resource-allocating network for function interpolation. Neural Computation, 3 (2), 213--225.

No context found.

Platt J, A resource-allocating network for function interpolation. Neural Comp. 3 (1991) 213-225

No context found.

J. Platt. A resource-allocating network for function interpolation. Neural Computation, 3:213--225, 1991.

No context found.

J. Platt, "A resource-allocating network for function interpolation," Neural Comput., vol. 3, pp. 213--225, 1991.

No context found.

J. Platt, " A resource - allocating network for function interpolation, " Neural Computation ", 3 ( 2 ), pp. 213 - 225, ( 1991 ). _________________________________________________________________ 7.53

No context found.

Platt, J., "A Resource Allocating Network for Function Interpolation", Neural Comp., 3, 213225, 1991.

No context found.

J. Platt, "A resource-allocating network for function interpolation," Neural Computation, vol. 3, no. 2, pp. 213--225, 1991.

No context found.

Platt, J. (1991). A resource-allocating network for function interpolation. Neural Computation, 3(2), 213--225.

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