Difference between 2D input and multiple input with recurrent neural networks for time series

Hi,

Note: Question edited in order to focus on the subject.

I'm using neural networks with 5 input time series of 3000 samples, to model 1 output time serie of 3000 samples. To do so, I used code generated thanks to the Neural Network toolbox for Time Series, and adapted it by using layrecnet. I use a 5x3000 matrix and a 1x3000 matrix to generate X (1x3000 cells of 5x1 double) and T (1x3000 cells of 1x1 double) with tonndata function followed by preparets before training. My network looks like this:

In network properties, it has only 1 input, which is a 2D dimensional input (since I have 5x1 double in each cell). I found this topic which explains how to use multiple inputs for a feed forward network. I then generated this 5 inputs network:

The only way I've been able to use this of network is by transforming by input_data into a 5x3000 cells and target into a 1x3000 cells. Which seems to be working fine, but training seems different from the previous one, with more frequent exit based on Mu threshold.

>> In my scenario, what are the pratical differences between the two pictured RNNS ?

I understand that second solutions allows for more freedom for each input (maybe delays, or the possibility to feed inputs to different layers), but is this useful in my scenario ?

Thank you in advance.

[Greg Heath]

We can help you better if you illustrate your problem using MATLAB data

help layrecnet

doc layrecnet

Greg

[Nicolas M.]

Thank you for your answer. However, neither help layrecnet nor doc layrecnet provide informations for the use and interests of multiple inputs.

What kind of data do you request ? Example of inputs/outputs, output of the network ? Datasets or just curves ?

Apologies if I misunderstand, but how data would tell if there is practical difference between the two network structures (in the present case where all my inputs go to the same first layer) ? I may have given too much details and gone misleading, but basically my question can be resumed as what would be the difference between having 1 input of size 5 and 5 inputs of size 1, if they all go to the same layer. Is there any difference between the way the two networks work (regarding the weights for example ?), and where could I find more info on this ?

(A bit like this one: https://fr.mathworks.com/matlabcentral/answers/296677-what-is-the-difference-between-net-numinputs-and-net-input-i-size)

I edited the quesiton in order to give clear focus on the subject.

Thank you in advance.

Answer 1

[Greg Heath]

Written MATLAB should treat the cases the same.

Unfortunately I don't have time to prove it

Greg

[Nicolas M.]

Thank you for your answer.

I'll try to update the question in case I find detailed documentation on the subject, or if someone else has confirmation/demonstration.