Description

This is a class representation of a non-linear PCA neural network. The nlpcaNet class is not meant for user-level usage.

Details

Creating Objects

new("nlpcaNet", net=[the network structure], hierarchic=[hierarchic design], fct=[the functions at each layer], fkt=[the functions used for forward propagation], weightDecay=[incremental decrease of weight changes over iterations (between 0 and 1)], featureSorting=[sort features or not], dataDist=[represents the present values], inverse=[net is inverse mode or not], fCount=[amount of times features were sorted], componentLayer=[which layer is the 'bottleneck' (principal components)], erro=[the used error function], gradient=[the used gradient method], weights=[the present weights], maxIter=[the amount of iterations that was done], scalingFactor=[the scale of the original matrix])

Slots

net

"matrix", matrix showing the representation of the neural network, e.g. (2,4,6) for a network with two features, a hidden layer and six output neurons (original variables).

hierarchic

"list", the hierarchic design of the network, holds 'idx' (), 'var' () and layer (which layer is the principal component layer).

fct

"character", a vector naming the functions that will be applied on each layer. "linr" is linear (i.e.) standard matrix products and "tanh" means that the arcus tangens is applied on the result of the matrix product (for non-linearity).

fkt

"character", same as fct but the functions used during back propagation.

weightDecay

"numeric", the value that is used to incrementally decrease the weight changes to ensure convergence.

featureSorting

"logical", indicates if features will be sorted or not. This is used to make the NLPCA assume properties closer to those of standard PCA were the first component is more important for reconstructing the data than the second component.

dataDist

"matrix", a matrix of ones and zeroes indicating which values will add to the errror.

inverse

"logical", network is inverse mode (currently only inverse is supported) or not. Eg. the case when we have truly missing values and wish to impute them.

fCount

"integer", Counter for the amount of times features were really sorted.

componentLayer

"numeric", the index of 'net' that is the component layer.

error

"function", the used error function. Currently only one is provided errorHierarchic.

gradient

"function", the used gradient function. Currently only one is provided derrorHierarchic

weights

"list", A list holding managements of the weights. The list has two functions, weights$current() and weights$set() which access a matrix in the local environment of this object.

maxIter

"integer", the amount of iterations used to train this network.

scalingFactor

"numeric", training the network is best made with 'small' values so the original data is scaled down to a suitable range by division with this number.

Methods

vector2matrices

Returns the weights in a matrix representation.

Author(s)

Henning Redestig

See Also

nlpca