From reading the problem statement (i.e. " This is essentially the same as X , but shifted one step to the left (to the past)") and looking at X and Y I wonder, why are we using Xt to “predict” Xt-1?
I suppose it doesn’t matter since we’re not trying to forecast anything, but just synthethize music, but I still wonder:
- what’s the advantage of doing it this way?
- why didn’t we do this for the synthethizing dinosaur names assignment (assignment 2)?
Hi tyatabe,
During training, the correct value of y at a particular timestep is the same as the input x of the next timestep - as this is how the song is built up - tone after tone, sequentially.
This means that, for instance, X[0,1,:] (the second value of the first training example) is the same as Y[0,0,:] (the first output value of the first training example). In other words, the value of Y is the same as the value of X shifted one step to the left.
I hope this clarifies.
Thank you reinoudbosch. I think this helps. Still need to ponder on this a bit more though 