NEURAL NETWORK-SIGMOID FUNCTION

I am working on using Artificial Neural Network for prediction.

My expected_Targets have values between 0 and 1.

The trained Neural_Target has sometimes -0.234 and 1.567 . 

Therefore, I want the trained Neural_Target to be in the range of 0 and 1. That is, any number less than 0.5 should be 0 and numbers above 0.5 should be 1.

I know I am to use "logsig" but I could not figure it well using MATLAB documentation.

Thanks in advance for your help

NEURAL NETWORK-SIGMOID FUNCTION. Available from: https://www.researchgate.net/post/NEURAL_NETWORK-SIGMOID_FUNCTION [accessed Jul 26, 2017].

Answer 1

[Greg Heath]

Apparently

 1. You are using the term prediction when you 
    mean classification

 2. You are using the term targets when
    you mean outputs. Targets are desired outputs.

You did not state

 3. The network design function. Patternnet 
   is the appropriate function for classification.

 4. Any non-default patternnet input values. Use 
   the help and doc commands for the 
   description of patternnet and it's default 
   inputs.

 5. The hidden and output layer transfer functions 
   are obtained from 

 >> hidlaytransfcn = net.layers{1}.transferFcn

   hidlaytransfcn =  'tansig'

 >> outlaytransfcn = net.layers{2}.transferFcn

   outlaytransfcn =  'tansig'

which yields outputs in the open range (-1,1)

This is because patternnet (and other training functions) AUTOMATICALLY transform inputs and targets to the closed interval [-1, 1]. Then reverses the transformation upon output.

Therefore, if you use patternnet, all you have to do is input your 0/1 targets and let patternnet do its job!

Hope this helps.

Thank you for formally accepting my answer

Greg

[Abu Muaadh AlDrWiy]

@Greg Heath.

Thank you for the answer.

I am very new to Neural Network, please kindly answer in the simplest form.

1. I am using feedforward net, so, I followed the documentation on feedforward net in this link.

https://se.mathworks.com/help/nnet/ref/feedforwardnet.html

With this code:

[x,t] = simplefit_dataset; net = feedforwardnet(10); net = train(net,x,t); view(net) y = net(x); perf = perform(net,y,t) -----------------------------------------------------------

2. In my own case, I have 23 inputs and 1 output.

The nature of the output is that the values are 0 and 1 only.

The trained output(neural output) gives values like 0.00007, 1.23e-10, 1.23.

Therefore, I need to make prediction with the trained network using : pred = net [a;b;....z]

The predicted output also yield = 1.234, -0.233456, 0.23e-12.

My Issues:

I want the value to be just 0 and 1.

Thank you for your valuable time and look forward to hear from you.

Regards, Alabi

[Greg Heath]

I am sorry I did not make my self more clear.

PATTERNNET is a special case of FEEDFORWARDNET that is designed for classification of 0/1 targets.

See my QUICKIE code for starters

http://www.mathworks.com/matlabcentral/newsreader/view_thread/348883#955252

Then searching the NEWSGROUP and ANSWERS using

 greg patternnet

should help more.

Hope this helps.

If it does,

Thank you for formally accepting my original answer

Greg

[Abu Muaadh AlDrWiy]

@Greg Heath. Thanks for your time. I have checked through the link, I did not quite get it well.

I have explained my dataset and the procedure in the small msword (1page) attached. I will be grateful if you can help to make suggestions on what to do.

Thanks in advance

Regards Alabi

Answer 2

[Abu Muaadh AlDrWiy]

@ Greg Heath.

Thanks for your answer.

Actually, I think I need to use "logsig" because I need the output between 0 and 1.

When I tried the suggestion you made, the following was obtained.

Yet, the outputs were not the expected 0 and 1. Thanks in Advance.

Answer 3

[Greg Heath]

I just posted a followup to QUICKIES

https://www.mathworks.com/matlabcentral/newsreader/view_thread/349202

The zeros and ones are the result of

      vec2ind ( y ) ~= vec2ind( t )

Hope this helps.

Thank you for formally accepting my answer

Greg