Neural network input ordering, does it matter?

Question

I have seen similar questions, but most involve preprocessing. Do the order in which the input are ordered have any importance. For example, say I have three features with three examples each such as:

[0, 0, 0, .5, .5, .5, 1, 1, 1]

Would it make any difference if they were ordered like this:

[0, .5, 1, 0, .5, 1, 0, .5, 1]

i.e does the order of the inputs have any effect at all?

Thanks

score 2 · Answer 1 · answered Apr 13 '16 at 00:05

2

First your input should be 2D: (n_examples, n_feat) so in your case:

[[0,.5,1],[0,.5,1],[0,.5,1]]

then the ordering of features does not matter as long as it is consistent through the whole training, classification process. So you could as well use:

[[.5,1,0],[.5,1,0],[.5,1,0]]

as long as you keep putting the first feature last everywhere else.

answered Apr 13 '16 at 00:05

Julien

13,986
5
29
53

I may be wrong but I think most libraries only accept 1d arrays or tuples for inputs. – shell Apr 13 '16 at 00:07
Which library do you use? – Julien Apr 13 '16 at 00:08
ok don't know that one. maybe add it as tag and mention it again in your post. – Julien Apr 13 '16 at 00:10
but anyway, however the library handles batches (running several examples at once) the order of features should be irrelevant as long as it is consistent – Julien Apr 13 '16 at 00:12

score 0 · Answer 2 · answered Apr 13 '16 at 05:44

0

They are different tensors, of course matters.

answered Apr 13 '16 at 05:44

knh170

2,960
1
11
17

Can you explain a little bit more? – shell Apr 13 '16 at 06:07
More what? [1,2,3] is different from what [3,2,1] expresses, it's that simple. – knh170 Apr 13 '16 at 06:09
Given enough training examples in the input, wouldn't it learn the same patterns from [1, 2, 3] as [3, 2, 1]? – shell Apr 13 '16 at 06:10
No. Unless you mean [shuffle](https://pythonprogramming.net/shuffling-data-learning/) – knh170 Apr 13 '16 at 06:12

Neural network input ordering, does it matter?

2 Answers2