Jx-NNT : Neural Network Toolbox

---

"Toward Talent Scientist: Sharing and Learning Together" --- Jingwei Too

---

Introduction

• This toolbox contains six type of neural networks

  • Artificial neural network ( ANN )
  • Feed Forward Neural Network ( FFNN )
  • Cascade Forward Neural Network ( CFNN )
  • Recurrent Neural Network ( RNN )
  • Generalized Regression Neural Network ( GRNN )
  • Probabilistic Neural Network ( PNN )

• The Main file shows the examples of how to use these neural network programs with the benchmark dataset

Usage

The main function jnn is used to perform the neural network. You may switch the algorithm by simply changes the 'ffnn' to other abbreviations

• If you wish to use feed forward neural network ( FFNN ) then you may write

NN = jnn('ffnn',feat,label,opts);

• If you want to use recurrent neural network ( RNN ) then you may write

NN = jnn('rnn',feat,label,opts); 

Input

• feat : feature vector matrix ( Instance x Features )
• label : label matrix ( Instance x 1 )
• opts : parameter settings
  • tf : choose either hold-out / k-fold
  • ho : ratio of testing data in hold-out validation
  • kfold : number of folds in k-fold cross-validation

Output

• NN : Neural Network model ( It contains several results )
  • acc : classification accuracy
  • con : confusion matrix
  • t : computational time (s)

How to choose the validation scheme?

There are two types of validation strategies listed as follows:

• Hold-out validation

opts.tf    = 1;
opts.ho    = 0.3;   % 30% data for testing 

• K-fold cross-validation

opts.tf    = 2
opts.kfold = 10;    % 10-fold cross-validation

Example 1 : Feed Forward Neural Network ( FFNN ) with hold-out validation

% Benchmark dataset 
load iris.mat; 

% Perform neural network 
opts.tf        = 1;
opts.ho        = 0.3;
opts.H         = 10;
opts.Maxepochs = 50;
NN = jnn('ffnn',feat,label,opts); 

% Accuracy
accuracy = NN.acc;
% Confusion matrix
confmat  = NN.con; 

Example 2 : Multi-layer Neural Network ( MNN ) with k-fold cross-validation

% Benchmark dataset 
load iris.mat; 

% Perform neural network 
opts.tf        = 2;
opts.kfold     = 10;
opts.H         = [10, 10, 10];
opts.Maxepochs = 50;
NN = jnn('nn',feat,label,opts); 

% Accuracy
accuracy = NN.acc;
% Confusion matrix
confmat  = NN.con; 

Requirement

• MATLAB 2014 or above
• Statistics and Machine Learning Toolbox

List of available neural network methods

• Use the opts to set the specific parameters
• The NN, FFNN, CFNN, and RNN have two extra parameters
  • H : hidden layer sizes ( up to three layers )
  • Maxepochs : maximum number of epochs
• The GRNN and PNN have one extra parameter
  • nSpread : number of spreads

No.	Abbreviation	Name	Extra Parameter(s)
06	'nn'	Neural Network	opts.H = [10, 10];
05	'ffnn'	Feed Forward Neural Network	opts.Maxepochs = 50;
04	'cfnn'	Cascade Forward Neural Network
03	'rnn'	Recurrent Neural Network
-----	--------------	----------------------------------------	----------------------
02	'grnn'	Generalized Regression Neural Network	opts.nSpread = 1;
01	'pnn'	Probabilistic Neural Network	opts.nSpread = 0.1;