Fast generating algorithm for a general three-layer perceptron

Zöllner, R.; Schmitz, H. J.; Wünsch, Friedrich; Krey, U.

doi:10.1016/s0893-6080(05)80138-8

Cited by 21 publications

(11 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Sequential Window Learning method [25] uses window transfer functions for which weights are obtained from solution of a system of algebraic equations. Target Switch algorithm [33] use traditional perceptron learning. Fig.…”

Section: Boolean Functionsmentioning

confidence: 99%

Constructive Neural Network Algorithms That Solve Highly Non-separable Problems

Grochowski

Duch

2009

Constructive Neural Networks

View full text Add to dashboard Cite

Learning from data with complex non-local relations and multimodal class distribution for widely used classification algorithms is still very hard. Even if accurate solution is found the resulting model may be too complex for a given data and will not generalize well. New types of learning algorithms are needed to extend capabilities of standard machine learning systems. Projection pursuit methods can avoid "curse of dimensionality" by discovering interesting structures in lowdimensional subspace. This paper introduces constructive neural architectures based on projection pursuit techniques that are able to discover simplest models of data with inherent highly complex logical structures.

show abstract

Section: Boolean Functionsmentioning

confidence: 99%

Constructive Neural Network Algorithms That Solve Highly Non-separable Problems

Grochowski

Duch

2009

Constructive Neural Networks

View full text Add to dashboard Cite

show abstract

“…One type is algorithm oriented. Many polynomial-time network construction/training algorithms [3], [14], [16], [17], [20] were developed only to solve approximate versions of the NCP because they do not take into account and cannot necessarily produce the exact minimum size.…”

Section: Introductionmentioning

confidence: 99%

A mathematical solution to a network construction problem

Takahashi¹

2000

IEEE Trans. Circuits Syst. I

View full text Add to dashboard Cite

One of the major open issues in neural networks includes a network construction problem (NCP) to find a procedure, polynomial time if possible, that produces a minimal structure (minimum size, threshold and weight) of a multilayer threshold feed-forward network where its output must not exceed any given admissible distortion from a sample. The NCP includes a subproblem, a network training problem (NTP), where the size is prespecified. Approximate versions of the NCP/NTP have been solved with iterative algorithms for the network construction/training. This paper provides a mathematically rigorous solution to the NCP using rate distortion theory from information theory and linear algebra. This solution is used to develop a mathematical procedure that specifically constructs a minimal structure from the sample. The procedure attains the exact minimum though its computational time is, at worst, nonpolynomial. The paper also constructs a polynomial-time shortcut for approximate minimum sizes, which is a promising alternative to current algorithms.

show abstract

“…To limit the e ective size of a network in order to avoid over tting one can either use additive or subtractive methods or regularization. Additive often called constructive methods start out from a small network and then insert additional units also known as nodes or neurons and connections also known as weights or links until the network can represent the required function Ash, 1989, Fahlman and Lebiere, 1990, Frean, 1990, Gallant, 1986, Hanson, 1989, M ezar and Nadal, 1989, Simon, 1993, Wang et al, 1994, Wynne-Jones, 1991, Zollner et al, 1992. Subtractive often called pruning methods start out from a large network and remove super uous parts until the network can just still represent the required function Le Cun et al, 1989, Finno et al, 1993, Hassibi and Stork, 1993, Levin et al, 1994.…”

Section: Introductionmentioning

confidence: 99%

Investigation of the CasCor Family of Learning Algorithms

Prechelt

1997

Neural Networks

View full text Add to dashboard Cite

Six learning algorithms are investigated and compared empirically. All of them are based on variants of the candidate training idea of the Cascade Correlation method. The comparison was performed using 42 di erent datasets from the Proben1 benchmark collection. The results indicate: 1 for these problems it is slightly better not to cascade the hidden units, 2 error minimization candidate training is better that covariance maximization for regression problems but may be a little worse for classi cation problems, 3 for most learning tasks, considering validation set errors during the selection of the best candidate will not lead to improved networks, but for a few tasks it will.Section | Computational Analysis.

show abstract

Fast generating algorithm for a general three-layer perceptron

Cited by 21 publications

References 7 publications

Constructive Neural Network Algorithms That Solve Highly Non-separable Problems

Constructive Neural Network Algorithms That Solve Highly Non-separable Problems

A mathematical solution to a network construction problem

Investigation of the CasCor Family of Learning Algorithms

Contact Info

Product

Resources

About