[nengo-user] SPA Vocabulary and Associative Memory: same fact, different semantic pointers?

[Aditya Gilra]

Hi,
A*B is supposed to be ~orthogonal to A and B, so that answers my previous question. Sorry was a bit hasty in dashing off the previous question. 
Cheers,Aditya.


     On Wednesday, 14 October 2015 1:33 PM, Aditya Gilra <aditya_gilra at yahoo.com> wrote:
   
 

 Hi,
I had not specified the vocabulary for the model (spa.SPA()) and for the buffers, and was comparing the memory output to the default vocabulary of the model.After specifying the same vocabulary consistently, both Fact0 and A*B give the same high similarity post cue-ing. So that solves the previous issue. Sorry for the traffic.
Now however, I'm surprised that A and B have negative dot products (similarity) with A*B. I'd expect the dot product to be at least greater than zero. Or am I missing something?
My code is updated at:https://gist.github.com/adityagilra/c3077515722662499ff6

Thanks,Aditya.
     On Tuesday, 13 October 2015 1:23 PM, Aditya Gilra <aditya_gilra at yahoo.com> wrote:
   
 

 Hi,
I created a semantic pointer 'Fact0'='A*B' in the spa.Vocabulary facts by:
facts.add('Fact0',facts.parse('A*B'))
Then, I created an spa.AssociativeMemory with above vocabulary.
Now when I query this associative memory with 'A*B' versus 'Fact0', I get very different responses!I'd expect the underlying neural activations to be the same for the two and the associative memory should return the same Fact0 match for both.
Why this difference?
My code is at:
https://gist.github.com/adityagilra/c3077515722662499ff6

Thanks,Aditya.

 
   

 
   
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://artsservices.uwaterloo.ca/pipermail/nengo-user/attachments/20151014/5e87c886/attachment-0002.html>