CODEWITHSUNDEEP
neural-networknormalizationdenormalization
Charles-Ugo Brouillard
Charles-Ugo Brouillard

Reputation: 221

Neural network input normalization by mean / standart deviation : how to do it?

I am consufed on how to normalize the inputs / outputs for a regression neural network using (Gaussian normalization ? ) mean & standart deviation normalization technique :

Most importantly, I Normalize from which data ?

Let me explain :

let's say i have these training data on a 2 input neurons, 2 hidden neurons , 1 output neuron:

[input1 : 10][input2: 5]
[input1:  30][input2: 255]

do i normalize by column(neuron), or from all the inputs data ? Is the mean for input neuron 1 = 

(10+30)/2

or 

(10+30+5+255)/4 ?

Try both with weird result using the typical XOR example (only 1s and 0s in the traning data), where i was actually loosing great accuracy when normalizing.

Upvotes: 0

Views: 422

Answers (1)

csz-carrot
csz-carrot

Reputation: 275

Normalization is to keep each dimension of input data in a certain range so usually it should be done in column. There are several ways for normalization. For example, linear normalization: It's the most common an easiest method and often used when the data is centered. It's counted by (V-Vmin)/(Vmax-V). And Gaussian normalization is counted by (V-Vavg)/Std.

Upvotes: 1

Related Questions