0

I'm working on a neural network to predict if an event is a neutrino or not. Each input x is a matrix : (3, 5484), three different measures, 5484 sensors.

So I should normalize the input for the network to work properly but I have doubts about it, let's explain :

My dataset is : (x_1,....,x_N). Let's focus on just one feature and think x is a vector of size (5484). Then each x has a mean : mean_x, and a std :std_x But the point of normalizing input is to rescale it without losing information (global normalization, same normalizing factors for each input) right ?

So, I'm not sure how I should rescale it. Should I flatten (x_1, .., x_N) to a 1-D vector : (x_1, ..., x_5484*N) and compute its mean and standard deviation ?

Or should I treat independently each of the 5484 input channels ?

I'm kind of lost.

Gericault
  • 219
  • 1
  • 3
  • 8

2 Answers2

0

It depends on how you treat the sensor data.

If you think the 5484 sensors represent different features, you should not flatten them and should treat the 3 * 5484 as image channel. Otherwise if the sensors is similar, you should flatten them.

But intuitively, I think first method is better.

danche
  • 1,775
  • 15
  • 22
  • Thanks both of you, Indeed sensors are supposed to represent more or less the same thing, but I treated each 3*5484 channels separately. However, data's are pretty sparse (say for a channel, over the 100k examples, maybe 5-10k are nonzeros), and very diverse (big variance compared to mean), so I'm afraid this preprocessing is gonna kill lots of data to zeros. Any opinion ? – Gericault Jun 27 '17 at 16:40
  • The preprocess is just scales the data and keeps the distribution. Just try it. – danche Jun 27 '17 at 17:08
0

If 3 sensors are independent from each other than you have 5484*3 features. For each one of the feature, you should flatten and standardize data.

For example your all-over data set might be a matrix of m * (5484*3) where m is the number of samples. Every row in the matrix is a sample.

you can use scipy http://scikit-learn.org/stable/modules/preprocessing.html

canbax
  • 3,432
  • 1
  • 27
  • 44