Normalizing data for neural networks

Hi,

I've read that it is good practice to normalize data before training a neural network.

There are different ways of normalizing data.

Does the data have to me normalized between 0 and 1? or can it be done using the standardize function - which won't necessarily give you numbers between 0 and 1 and could give you negative numbers.

Many thanks

Answer 1

[Chandra Kurniawan]

Hi,

I've heard that the artificial neural network training data must be normalized before the training process.

I have a code that can normalize your data into spesific range that you want.

p = [4 4 3 3 4;            
     2 1 2 1 1;
     2 2 2 4 2];

a = min(p(:));
b = max(p(:));
ra = 0.9;
rb = 0.1;
pa = (((ra-rb) * (p - a)) / (b - a)) + rb;

Let say you want to normalize p into 0.1 to 0.9.

p is your data.

ra is 0.9 and rb is 0.1.

Then your normalized data is pa

[Greg Heath]

Demos in the comp.a.neural-nets FAQ indicate that better precision is obtained when the input data is relatively balanced about 0 AND TANSIG (instead of LOGSIG) activation functions are used in hidden layers.

Hope this helps.

Greg

[Kaushal Raval]

i want to find out my output in original form how can i find it ? please help me

[Greg Heath]

If you use the standard programs e.g., FITNET, PATTERNNET, TIMEDELAYNET, NARNET & NARXNET,

All of the normalization and de-normalization is done automatically (==>DONWORRIBOUTIT).

All you have to do is run the example programs in, e.g.,

 help fitnet
 doc fitnet

If you need additional sample data

 help nndatasets
 doc nndatasets

For more detailed examples search in the NEWSGROUP and ANSWERS. For example

 NEWSGROUP              2014-15     all-time
 tutorial                  58         2575
 tutorial neural           16          127
 tutorial neural greg      15           58

Hope this helps.

Greg

[murat tuna]

Sorry, where is NEWSGROUP?

Answer 2

[Greg Heath]

The best combination to use for a MLP (e.g., NEWFF) with one or more hidden layers is

1. TANSIG hidden layer activation functions

2. EITHER standardization (zero-mean/unit-variance: doc MAPSTD)

   OR [ -1 1 ] normalization ( [min,max] => [ -1, 1 ] ): doc MAPMINMAX)

Convincing demonstrations are available in the comp.ai.neural-nets FAQ.

For classification among c classes, using columns of the c-dimensional unit matrix eye(c) as targets guarantees that the outputs can be interpreted as valid approximatations to input conditional posterior probabilities. For that reason, the commonly used normalization to [0.1 0.9] is not recommended.

WARNING: NEWFF automatically uses the MINMAX normalization as a default. Standardization must be explicitly specified.

Hope this helps.

Greg

[John]

Hi Greg,

Thank you for your detailed response.

I'm only new matlab so to be honest I don't really understand some of it. However I've googled the terms and I think your advising me if I was building my own network?

I intend to use the nprtool GUI to classify data, which is a 2 layed feed forward network. This means that I will have to standardize the inputs before I import them into the tool.

What method would you recommend I that I use to standardize these? The standardize function or between 0 and 1?

Many thanks

John

[owr]

I dont have access to the Neural Network Toolbox anymore, but if I recall correctly you should be able to generate code from the nprtool GUI (last tab maybe?). You can use this code to do your work without the GUI, customize it as need be, and also learn from it to gain a deeper understanding.

What I think Greg is referring to above is the fact that the function "newff" (a quick function to initialize a network) uses the built in normalization (see toolbox function mapminmax). If you want to change this, you'll have to make some custom changes. I dont recall if the nprtool uses newff - this can be verified by generating and viewing the code.

This is all from memory as I dont have access to the toolbox anymore - so take my comments as general guidelines, not as absolute.

Good luck.

[John]

Thank you

[Greg Heath]

Standardization means zero-mean/unit-variance.

My preferences:

1. TANSIG in hidden layers

2. Standardize reals and mixtures of reals and binary.

3. {-1,1} for binary and reals that have bounds imposed by math or physics.

Hope this helps.

Greg

Answer 3

[Greg Heath]

In general, if you decide to standardize or normalize, each ROW is treated SEPARATELY.

If you do this, either use MAPSTD, MAPMNMX, or the following:

[I N] = size(p)

%STANDARDIZATION

meanp = repmat(mean(p,2),1,N);

stdp = repmat(std(p,0,2),1,N);

pstd = (p-meanp)./stdp ;

%NORMALIZATION

minp = repmat(min(p,[],2),1,N);

maxp = repmat(max(p,[],2),1,N);

pn = minpn +(maxpn-minpn).*(p-minp)./(maxp-pmin);

Hope this helps

Greg

[John]

Many thanks

[fehmi zarzoum]

hi, Undefined function or variable 'pmin'.

[Greg Heath]

Yeah, should be minp.

[electronx engr]

plz can u help me in this that after training with normalized data, how can I get the network (using gensim command) that works on unnormalized input, since I have created and trained the network using normalized input and output?

Answer 4

[Sarillee]

y=(x-min(x))/(max(x)-min(x))

try this...

x is input....

y is the output...

[Greg Heath]

Not valid for matrix inputs

Answer 5

[Imran Babar]

mu_input=mean(trainingInput); std_input=std(trainingInput); trainingInput=(trainingInput(:,:)-mu_input(:,1))/std_input(:,1);

I hope this will serve your purpose

[Greg Heath]

Not valid for matrix inputs

[Abul Fujail]

in case of matrix data, the min and max value corresponds to a column or the whole dataset. E.g. i have 5 input columns of data, in this case whether i should choose min/max for each column and do the normalization or min/max form all over the column and calculate.