1
$\begingroup$

I want to train the network based on two sets of data. For example, I want the network to predict the humidity based on past humidity trends AND past temperature trends. In this case, how should I organize the input layer? I would think that for just one time-series, I would use a regular sliding window on the series. With two series, do I just present both windows (from the humidity and temperature series) to the input layer that is twice the size of the window? if not, how else can I configure the input layer so that it doesn't confuse between the two sets of training data?

Do I just let the network sort this out by itself or is there a preferred method of presenting two (or more) sets of training data to the network? Thanks.

Improve this question
$\endgroup$
4
  • $\begingroup$ Could an outer product layer be useful? It seems like it would get every variable in one vector combined with every variable in another. $\endgroup$
    – chris
    Commented Dec 18, 2016 at 0:48
  • $\begingroup$ Could you elaborate on this outer product layer? I'm just finding presenting two series as one a little counter-intuitive and am wondering if I'm missing something obvious here. What do you have in mind - some form of pre-processing? $\endgroup$
    – Daniel Wee
    Commented Dec 18, 2016 at 2:08
  • $\begingroup$ Layer is a silly word I guess. It's just the outer product of two vectors. Really it seems like you need an RNN of some kind. Maybe you could preprocess the two to combine them into a single sequence though. $\endgroup$
    – chris
    Commented Dec 18, 2016 at 2:13
  • $\begingroup$ I intend to feed this into an LSTM type NN. As for submitting the vector product - won't that destroy some of the data features that might be useful for the NN? $\endgroup$
    – Daniel Wee
    Commented Dec 18, 2016 at 2:19

1 Answer 1

Reset to default
1
$\begingroup$

The whole idea of NN is that it learns by itself to give the appropriate weights and biases to the input neurons, so yes - just give your network the 2 series and it should return the appropriate weight and bias for each neuron in the input.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thanks beniev. I suspected as much but wondered if there were some optimisations that could be done eg: should I present the series in concatenated fashion or in interleaved fashion? $\endgroup$
    – Daniel Wee
    Commented Dec 18, 2016 at 0:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.