How does the input to a neural network change for a given input delay?

I have an input array [a,b] from a time series and specify the input delay T. This is being sent to a NARXnet for training. How does my original array change for that delay? And is there any way to see the inputs to the network either before or after training? I would like to see how my original inputs were modified for training the network.

Answer 1

[Greg Heath]

It is not exactly clear what you mean.

When you have an input 1:N with a timedelay net that has a delay d, the first d inputs do not result in an output. Instead, they are routed to a delay buffer. Once the buffer has stored d inputs, it starts yielding an output.

Then, for every new input data point, an output data point is created from the current d input points that were in the delay buffer..

As a result, Your output will have values over times d+1:N

Hope this helps.

Greg

[ss32]

Right, I understand that. Until d inputs have filled the buffer, nothing is output, because the output y(t) is a function of {x(t),x(t-1),...,x(t-d)}. What I want to know is how is Matlab doing that? Is it iteratively filling the buffer and shifting the values for each new x(t)? Or does it create an entirely new array with all of the values? I ask this because I am trying to validate results I have from a NARXnet using Tensorflow and I need to be sure that my inputs are the same across both networks.

[Greg Heath]

Try contacting MATLAB directly.