Neural Network Input

1. If  a data set shows that several/different inputs are mapping to same output, then we may use Artificial Neural Network Model, to model this issue.

[Note: If input and Output has one-to-one relationship we can model them using Rules. If input and output have one-to-many relationship you can use Fuzzy Logic ].

 

2. ANN input values must be expressed as very small quantities.

 

X = x1, x2, x3, x4, …, xn

For these purpose we divide all the inputs components by a big number like 100 or normalize the input vector.

3. Most of the image related data can be represented or prepared as an input for ANN by digitizing them in to 1 and 0 .

 

Image = { 0 1 0 0 0 0 1 0 0 1 1 0 0 0 0 0 }

 

4. When we have inputs with zeros (0) as a component, the The computation of Net value, ignore that component whatever the value of weights for that component.

 

Some inputs with all components zero ‘0’ may have non-zero output.

X = ( 0  0  0 )  =>>  d (output) = ( 1   1  )

In this case, for whatever weights, you couldn’t achieve the desired output. There for we have the following solutions.

  •  To address this issue we can add extra non-zero component (e.g: -1 ), to each input vector. This unit is called a Bias.

As a matter of policy, Bias are added to all the neurons in the network.

Note: During Net calculations, neurons in the intermediary layers may also may generate their output as zero, which become the input for next layer. Therefore better to add Bias for all the Neurons.

 

  • Use of Bi-polar activation function.

 

 

 

5. During a training session you never get the inputs repeatedly from the same class. Instead randomize over different classes. This help to develop a higher generalization.

If you train one class completely and train another class obviously error will increase at the transition point.

 

In order to randomize over classes, we must have a way to classify data.

Neural Network itself can be used to classify data.

In this case we can train the data set using un-supervised training by considering single layer Neural Network.

E.g. : –  If we have a three (3) neuron in one layer during un-supervised training, we can identify which Input generate maximum output on which neuron. On this we have 3 different classes identified.

In this sense Neural Network training session, may be supported by another small neural network which work as the pre-processor for data classification.

.

.

.

Related Articles :

1. Choosing a propper Neural Network Architecture

2. Neural Network Training Issues     

3. AForge.NET Framework for Artificial Neural Networks development

 

 

 

Leave a Comment

You must be logged in to post a comment.