Normalization is required so that all the inputs are at a comparable range.  The general approach to data normalization is part of the data exploration which include:

• Data exploration (sometime called data clean-up) 
• Data transformation
• Outlier detection and removal
• Data normalization
• Data analysis
• Validation of results

I like to look at your question in the field of feature extraction/selection.

As a result, you are able to analysis and process your data to prepare them as inputs for NN.

For this, you can look for methods for feature extraction. On the other hand, you can project your data to another space which is more robust to noise and outliers or you may use dimension reduction methods like PCA and LDA to find more important elements and discard poor and non-valuable elements.

It's clear that no exact solution exist for your problem. It is a good idea to use different transfer functions, number of hidden layers, training function, normalization methods and etc to find best architecture and parameters of NN for your application.

The pre-processing stage (encoding) is the process that transforms the input data into a suitable form, which the network can identify and use. Essentially, every input in the network will associate with a bit-vector of output .It aims to increase the number of significant locations and decrease the usage of memory storage. Then the error function (Bishop 1995) or the cost function is used to measure the distance between the targets and the outputs of the network. Then the weights of the network are updated in the direction that makes the error function minimum. However, there is several error functions can be used for training neural networks such as Sum-of-squares error, City-block, Cross-entropy (single & multiple), Kohonen, Cross-entropy, etc.

So, it is not depend on using 0, 1 notation or the bipolar notation -1 for 0 , 1 for 1. It is depend on the behavior of data-sets pair (input, target).

Although the input of neural network can be in any range, there is a saturation effect so that the unit is only sensitive to inputs within a fairly limited range. Numeric values have to be scaled into a range which is appropriate for the network. Typically, raw variable values are scaled linearly. In some circumstances, nonlinear scaling may be appropriate (for example, if you know that a variable is exponentially distributed, you might take the logarithm.)

Normalization is required so that all the inputs are at a comparable range.  The general approach to data normalization is part of the data exploration which include:

• Data exploration (sometime called data clean-up) 
• Data transformation
• Outlier detection and removal
• Data normalization
• Data analysis
• Validation of results

ANN - Processing Units - ANN consists of a pool of simple processing units that communicate by sending signals to each other over a large number of weighted connections. Each unit performs a relatively basic task; receive input from neighbors or external sources and utilize that information to compute an output signal that is propagated to other units in the ANN.

http://www.dhtusa.com/media/NeuralNetworkIntro.pdf

The objectives of data preprocessing include size reduction of the input space, smoother relationships, data normalization, noise reduction, and feature extraction.

I recommend reading chapter 15 of the book: Fuzzy neural intelligent systems: Mathematical foundation and the applications in engineering (see attached PDF).

Data normalization can provide a better modeling and avoid numerical problems. Several algorithms can be used to normalize the data:

1) Min-Max Normalization, min-max normalization is a linear scaling algorithm. It transforms the original input range into a new data range (typically 0-1).

2) Zscore Normalization, in Zscore normalization, the input variable data is converted into zero mean and unit variance.

3) Sigmoidal Normalization, sigmoidal normalization is a nonlinear transformation. It transforms the input data into the range -1 to 1, using a sigmoid function

The outliers of the data points usually have large values. In order to represent those large outlier data, the sigmoidal normalization is an appropriate approach.

Li, H., Chen, C. P., & Huang, H. P. (2000). Fuzzy neural intelligent systems: Mathematical foundation and the applications in engineering. CRC Press.

http://www.sciencedirect.com/science/article/pii/S2212017313003137

Using highly variable data for training ANN models may induce "overrating".  So for practical reasons as well as making all the inputs at a comparable range normalizing or standardizing the inputs in ANNs can make training faster and reduce the chances of getting stuck in local optima. If you use Matlab you can use either mapminmax or mapstd as follows: 

[yn, ps] = mapminmax(ymin, ymax)

MinMax or mapminmax algorithm: y = (x-xmean)*(ystd/xstd) + ymean;

[yn, ps] = mapstd(ymean, ystd)

Z-Score or mapstd algorithm: y = (x-xmean)*(ystd/xstd) + ymean;

where yn is normalized data.

http://www.mathworks.com/help/nnet/ref/mapminmax.html

http://www.mathworks.com/help/nnet/ref/mapstd.html

One common problem with MLP is saturation  When the weighted sum is big, the output of the tanh (or sigmoid) saturates, and the gradient tends towards 0. Hence, we should initialize the parameters such that the average weighted sum is in the linear part of the transfer function. 

Normalization  for regression and classification problems

Normalization of  Inputs:

• normalized with zero mean and unit variance,

Normalizatioh of Targets:

• For regression problems: normalized with zero mean and unit variance,
• For  classification problems:
• output transfer function is tanh: 0.6 and -0.6
• output transfer function is sigmoid: 0.8 and 0.2

See Prof. Leon Bottou’s thesis for details:

Léon Bottou: Une Approche théorique de l'Apprentissage Connexionniste: Applications à la Reconnaissance de la Parole, Ph.D. thesis, Université de Paris XI, Orsay, France, 1991.

http://leon.bottou.org/papers

Normalization of input and target data is a key for properly training the artificial neural network.