Good question and +1 for having some code! Keep in mind when dealing in floats, there'll funkiness. Eg. in Python a = numpy.array([x * 0.1 for x in range(9)]) $\implies$ array([ 0.1, 0.2, ..., 0.9]) but a[6] * 6 $\ne$ 4.2, instead results in 4.2000000000000002... Side note, there's fun stuff here like MathJax to do perverse stuff like... $\color{#CD8C00}{\fbox{ resultGenerator }{\color{#00A}{\xleftarrow[]{\color{#000}{\text{function}_{\left(key\_word\_args\right)}}}}\over{\xrightarrow[\color{#000}{\text{returned value}}]{}}}}\color{#00A}{\fbox{somethingQuestionable}}$

Most welcome @Mario! Please edit the edits I've made, specifically the second formatted block, I think it'll help with getting readers up-to-speed with the things your asking. Also are you training for correlation within columns, between, columns, across all columns, etc...? In other words, having a sample of what is being fed to your NN is great, knowing how it feeds too will likely be even better for getting solid suggestions. And side note, I recognized that there where different questions, it was more to validate that feeling of déjà vu that other readers may have had.

This is nearly identical to a question I ran across less than 24 hours ago... yeah the questions be different, but the set-up's the same... it would probably help answerers if either of ya provided more information, eg. what's the raw input look like? Doesn't have to be copied, just give people a better idea of your problem space with something similar. Specifically for scaling heard rumors in other questions that this may not necessarily be required... but that might be hear-say.