RMF - Some major new bug has been introduced somewhere. The code seemed to be working perfectly the other day, but now it no longer is. It no longer even gets the Saffran or Aslin data correct. Any idea what's up? But technically, it looks good and I'm glad we have it on our Website now.
CJMA -
I think the simulator is working fine. If you look at the bottom right Bigram Error over Time graph you will see that word internal bigrams are learned better than boundary-cross bigrams or non-word bigrams
The probem is that the simulator is using a different test criteria to matlab. I'd argue that the simulator one is possibly more appropriate.
In matlab you always pass the hidden representation forwards to next step and output just the single delta for final item.
e.g You test the word ABC by getting the delta for
input1 = Hidden(AB)
input2 = C
Let's call this Delta(Hidden(AB), C)
In the simulator, we take the mean of the all the errors encountered in the word and calculate them according to whether the internal hidden representations were below the criterion
So for ABC, we have two different expressions for early and later in learning
Early on when Delta(A,B) > CRITERION, our error is
0.5 * ( Delta(A, B) + Delta(B , C))
Later when Delta(A,B) < CRITERION,
0.5 * ( Delta(A, B) + Delta(Hidden(AB) , C))
When AB has only just been learned Delta(Hid(AB), C) is quite large. So the part-word delta's can end up being bigger than the word deltas. Though (from the Giroux and Rey simulation result) we know that eventually the latter is smaller.
I guess the question is. how do we show all of this in a transparent graphical fashion in the simulator?