CODEWITHSUNDEEP
arraysalgorithmmachine-learningneural-networkprediction
P A S T R Y
P A S T R Y

Reputation: 351

Radial basis network character recognition

I want to develop a simple character recognition program by implementing a given neural network kind; a simple command line-type is enough.

The radial basis function neural network was assigned to me and I already studied the weight training, input-to-hidden-to-output procedures but I am still doubtful of in implementing it. My references are (1) and (2).

A simple one-dimensional array of a 10 by 10 binary object (that represents a character) is the input. For example, the array below

input = array( 
    0,0,0,1,1,1,1,0,0,0,
    0,0,1,0,0,0,0,1,0,0,
    0,1,0,0,0,0,0,0,1,0,
    1,0,0,0,0,0,0,0,0,1,
    1,1,1,1,1,1,1,1,1,1,
    1,0,0,0,0,0,0,0,0,1,
    1,0,0,0,0,0,0,0,0,1,
    1,0,0,0,0,0,0,0,0,1,
    1,0,0,0,0,0,0,0,0,1,
    1,0,0,0,0,0,0,0,0,1 )

is the representation of the character "A":

0   0   0   1   1   1   1   0   0   0
0   0   1   0   0   0   0   1   0   0
0   1   0   0   0   0   0   0   1   0
1   0   0   0   0   0   0   0   0   1
1   1   1   1   1   1   1   1   1   1
1   0   0   0   0   0   0   0   0   1
1   0   0   0   0   0   0   0   0   1
1   0   0   0   0   0   0   0   0   1
1   0   0   0   0   0   0   0   0   1
1   0   0   0   0   0   0   0   0   1

I plan to take the total weight of the input and compare it to the training set as in the saved 1-D arrays of the other characters of the alphabet and the one with the closest is the prediction.

The problem is I tend to understand algorithms better if presented in a CLRS-manner or similar type as opposed to mathematical formula. I find it hard to understand the explanations in those two papers (which I find the easiest to read among others here in the Google search).

Can someone point me to a friendly algorithm for a RBNFF that takes in an array and produces an output of weights? If not, a paper that explains this in Layman's manner would be appreciated.

Upvotes: 1

Views: 569

Answers (1)

SlimJim
SlimJim

Reputation: 2272

Training

For what I could find there is no "one right way" to train them.

The simplest training I could find was by a composition of two algorithms

  1. (Clustering) Taking the left part (input weights and RBFs) of the network and doing unsupervised clustering. There is a few things you can try out hard/soft and the number of clusters/RBFs. Each cluster is a representation of a single RBF with the weights connecting to it. How you go from having clusters to get rbf and rbf weights depends on what clustering you are using. (I can extend this if it's unclear)

  2. (Neural Network) The solving the left out part of the original RBFNN from the last step by using the output from the clustering as input to an ordinary single layer neural network.

Probably easier to find these more primitive algorithms easily explained

EDIT

found some "pseudo"-code with explanations that might explain it all better (written in C#)

http://msdn.microsoft.com/en-us/magazine/dn532201.aspx

(Supposedly) working python code

https://github.com/andrewdyates/Radial-Basis-Function-Neural-Network

Upvotes: 1

Related Questions