SoftmaxWeightTransformation
Related Doc:
package optim
Prevents underflows and overflows when optimizing the weights
Prevents underflows and overflows when optimizing the weights
Because of the functions involved in mapping points from and to the weight simplex sometimes over/underflows can occur and make some weights go to zero or one. To avoid this, at each iteration this function add an offset to the vector without affecting the resulting simplex points, or bound them if necessary, in this case affecting the resulting weights.
(possibly) arbitrary vector that is going to be mapped to a valid weight vector
rescaled/translated vector
Take a valid weight vector of a mixture model and map it to another in which unconstrained gradient ascent can be performed.
Take a valid weight vector of a mixture model and map it to another in which unconstrained gradient ascent can be performed. see https://en.wikipedia.org/wiki/Simplex
mixture weights
Take an arbitrary vector and map it to the weight simplex see https://en.wikipedia.org/wiki/Simplex
Take an arbitrary vector and map it to the weight simplex see https://en.wikipedia.org/wiki/Simplex
valid mixture weight vector
Softmax mapping to fit the weights vector
The precise mapping is w_i => log(w_i/w_last) and is an implementation of the procedure described here