Feedback and Input Delays in Neural Networks

Pretty simple question. I am training a NARX net.

Here is my delay information:

Delay=14;

inputDelays=[0,7:Delay];

feedbackDelays = 7:Delay;

My physical interpretation of this is that the training net uses the current input data (0), as well as the input data corresponding from 7 timesteps to 14 timesteps ago (in addition the target data generated from this input data also from 7-14 timesteps ago) to generate the current target data.

Can someone clarify if this is correct?

Also for the above delays what is the correct interpreation of the Error-correlation plots. The net doesn't use for example the 1-6 lags and yet it displays them in the correlation plots. What is it plotting?

Answer 1

[Raunak Gupta]

Hi,

The inputDelays signifies the data that is needed for current time series that depends on the array of values entered in inputDelays (The values signify the data with some ‘i’ timesteps ago for a delay value of ‘i’). So, the target data is generated with same strategy.

As for the Error Correlation Plots these are plotted to find autocorrelation of errors across varying lags and cross correlation of inputs to errors across varying lags. Even though the lags of 1-6 are not used but while creating data with preparets we also store the data with various lags. Correlation plots help identifying goodness of the trained model. You may look for ploterrcorr , plotinerrcorr and plotresponse for more information.