This is most likely nonsense.... but:
Build and train a net.
Create a new input column consisting of the predictions (single ouput
net) of the old net.
Train and get the results from the new net.
I was thinking that if you are trying to predict the sin(x) and you
have 1 input column, you get a "1 interval" lag. So you add the
ubiquitous "lag" column.
Interesting, this lag column is exactly what the output of the old net
is. Is this a coincidence or did this feedback give the net what it
Does this new column tell the new net what its error is going to be, so
I'm sure this has been done before. Is there a name for it?