Hi Allan, 

This merely depends on the data itself. However, ANN in some platforms normalizes the data by default between [-1, 1].

HTH.

Samer

Dear Dr Brimicombe,

From my personal experience, if the values of the inputs are numerically too scattered i.e. in my case I was modeling an electric circuit and there were capacitances in microfarads and frequency in gigahertz I had to use scaling which necessarily may not be between 0 & 1 max and min values can be different. But then you have to apply scaling and descaling formula at input and output  respectively. This is readily available in literature, explained also in the book my Simon Haykin. Else normalizing is not required. Moreover, if your output layer uses limiting function normalizing may cause inaccuracies. Hope it did help. Regards 

Lets take an example to explain the necessity of Normalisation,

Consider that you and one of your students are 6ft and 4.5ft tall respectively. Assume that in a class room your student is standing a bench of 2ft height and you are standing on ground adjacent to him, for a person who is walks past the class room door gets a glimpse inside and visually perceives that your student is taller than you. This is called "Benching Effect".

The severely effects the performance of the ANN when one of your input to the neural network varies in the range 0-100 and other varies from 0-0.01. 

Hence to avoid this, "Normalisation" of data is required.

 Depending on the input, normalization is required. it can scale the inputs in a uniform way. from my experience the NN gives a faster training whenever the inputs are normalized.

All depends on the dat, structure and type of neural network. You can scale the real number to the range [0,1]. However, if you normilize your inputs to the {0,1} it is possible to reach the faster training and reduce the chances of getting stuck in local optima.

Dear Allan,

The interval of input variables is important due to the method of generating initial values of hidden neurons weights and biases. When this method does not take into account different input intervals, it is safer to normalize data. See link below.

Output variable should be normalized when you use sigmoid in output neuron. This limits the output to (0, 1) or (-1, 1).

Regards

Chapter Extreme Learning Machine as a Function Approximator: Initial...

Normalising a Dataset helps to bring the attributes of the Data within a common scale, say [0 1] or [-1 1] depending on the normalisation technique used. Normalisation is good when there is a huge variation in the scale of the values of a dataset. But in some cases, for example when the values are binary, it is irrelevant to normalise because the values would remain unchanged.

Hi Allan,

I've checked (in one of my papers) that applied method of standarization of input data influences the result achieved. But it was based on simulated data.

Now I now - because I've checked it - that this is true, based on real data. I've applied non standardized data too, achieving lower accuracy of predictions from ANN MLP type.

That is one of reasons: predictions without standardization are less accurate.

Best Regards

Hubert

It is definitely not necessary to normalize the data in order to limit it to an interval, whether [0,1] or [-1, +1] (the so-called Min-Max normalization), and it may even have very detrimental effects. What is definitely necessary is to normalize the data so that different inputs are of the same order of magnitude (as discussed in some of the above replies); the reasonable thing to do (which is universal practice in signal processing, by the way) is to susbtract the mean from each input and divide the result by the standard deviation of the input; by so doing, all normalized inputs have zero mean, and standard deviation equal to one (this is also known as z-score normalization). Same for the output.

Dear Gerard,

I've trated standardization method as one of paratemeters of ANN. The learning agorythm, number of hidden layers, nuber of neurons in the hidden layer, fitting function etc. I've searched for the most accurate set  of predictions. There was 7 types of standardization checked, z-score too. It has occured that typical i.e. linear maximum was the best (in my case). The case is not included inthe paper, I mentioned previously, but in my phd thesis. It is not published yet. I will make an article after getting my phd (probably Sept 2017)

Best regards

Hubert

Normalizing the input will avoid " overflow" . I have tried z-score, robust z-score normalization methods, which have given satisfactory results.

Amir is right. Sorry for forgetting this point in my previous reply; I should have added that using a logistic function in the hidden layer does not make any sense, as all examples that will drive the ouput of the hidden units to zero will not have any influence on the network output, because 0 multiplied by any number is equal to 0. Always use the tanh function in the hidden layer, unless prior knowledge on the task suggests otherwise (see e.g. bump modeling for ECG or EEG signals). The logistic function is useful only in an output neuron, when one wants the neural network to provide an estimation of the posterior probabilities of the classes.

All these questions were discussed at length in papers and theses published in the mid-1980s and in the early 1990s, and the answers are in good textbooks. As the saying goes, "six months in front of the computer may save you one day in the library"...

It is better to normalize data in order to have the same range of values for each of the inputs to the ANN model. This guarantees stable convergence of network weights and biases.