Hi,
Sorry for late reply,

To: Ramesh Ganesan ,

I dint really understand what you explained. But I do think we can get 
predictive states out of the weight matrix.
When a cell in a column is in predictive state, we can get the whole predictive 
state as the output by interpreting the column
which has that predictive cell. There wont just be weight matrixs in this 
model. there would also we a matrix representing all cells.
So every cell would have a current value , which if is above certain threshold 
its in predictive state, and above another threshold means its firing. But I am 
going to make a video explaining this algorithm and Should I post it on youtube 
and send a link here. 
That way I can be more clear in explaining my thoughts and you can point out my 
mistakes precisely.

To: Gil,

In a recent video uploaded on youtube by Rahul Agarwal, he states that temporal 
pooler's job is to select which cell in an activated column
is active, whether its one cell or whether the whole column bursts. He says the 
Classifier is the one that does multi step predictions.
It gives predictions certain steps ahead.
May be you are right, I am not addressing the classifier in particular, but I 
am thinking of the prediction algorithm that goes in grok.
As i mentioned above, it is possible to interpret the prediction out of the 
cells that are in predictive state. Prediction is going to be a
pattern like an input pattern.

To: Scott Purdy,

Current Classifier makes predictions n steps ahead. And when it does so, we 
dont get predictions from 1 to n-1 steps.
Separate task specific tables are required for those predictions. Plus it needs 
lot of memory. I dont know how much memory would be 
required to store weights of every cell but i guess it would be a fixed memory 
allocation. It wont vary with inputs.



Thankyou so much for replying :) ! This work is very exciting and your replies 
have excited me further.
ASEEM HEGSHETYE



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

Reply via email to