Aseem,

I think you might be confusing the tasks of the Temporal Pooler and the 
Classifier. The Temporal Pooler's job is to store the permanence values between 
cells and determine the predictive state of the cells in the next time step.

The Classifier is trying to solve a different problem: How do we map a given 
set of a cells to an actual set of input bits? The temporal pooler is capable 
of making predictions about which cells are going to be activated next, but it 
has no idea of how these cells correlate to input bits. Because at the end of 
the day, we care about predicting the actual input, not cells. Cells are only a 
means to an end.

There's a history lesson also hiding in here. In the past, Numenta tried to 
tackle this problem of mapping cells back to input bits by using a technique 
called "reconstruction". Reconstruction, also referred to in the code as "Top 
Down Computing" entails feeding in a set of cells into the temporal pooler and 
asking the temporal pooler for the set of columns for which the predictive 
cells reside. These columns, in turn, would be fed into the spatial pooler, who 
would be ask to produce the most likely set of input bits that could have 
activated these set of columns. For those of you familiar with deep learning 
and restricted boltzman machines, the idea is pretty similar. This approach 
didn't work out very well in practice so Numenta switched to using the CLA 
classifier. Personally, however, I think reconstruction would be a very 
interesting avenue to explore.

Gil.

From: Scott Purdy <[email protected]<mailto:[email protected]>>
Reply-To: "NuPIC general mailing list." 
<[email protected]<mailto:[email protected]>>
Date: Friday, August 9, 2013 10:54 AM
To: "NuPIC general mailing list." 
<[email protected]<mailto:[email protected]>>
Subject: Re: [nupic-dev] CLA classifer

Aseem, what limitations of the current classifier are you trying to address?


On Fri, Aug 9, 2013 at 12:34 AM, Ramesh Ganesan 
<[email protected]<mailto:[email protected]>> wrote:
Good thoughts Aseem Hegshetye :)

I think, storing weight matrix itself wouldn't enough to calculate 2nd, 3rd 
etc., prediction steps. we have to find out activation of cells using the 
predicted state values and do the inhibition. The process might be like

pattern -> get predicted state cells -> next phase without learning using 
predicted value of that predicted cells (2nd step)-> get predicted state cells 
-> next phase without learning using predicted value of that predicted cells 
(3rd step) -> ...

is it make sense?

Thanks
Ramesh Ganesan.


_______________________________________________
nupic mailing list
[email protected]<mailto:[email protected]>
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org


_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to