That was my idea: the encoding it is a complex task to it in HW. I think
the TP and SP are doable (especially if you use non-volatile memory
devices).

My approach was to use the examples of the temporal pooler as start point
(nupic/examples/tp/). Do you think that it could easier to start from the
"real thing" and just bypass internally the encoding? I'm a bit lost here
:-)


2013/8/6 Scott Purdy <[email protected]>

> One thing we have talked about but not implemented is a pass-through
> encoder.  It would let you create an SDR in any way you want and pass it in
> to the CLA model's compute method as the field value.  That would probably
> be the best way to approach your problem if you want to implement it with
> the CLA.
>
>
> On Tue, Aug 6, 2013 at 1:45 PM, Fergal Byrne 
> <[email protected]>wrote:
>
>>
>>  Hi Valentin,
>>
>>  The inputs to a region are best if they are an SDR (large number of
>> bits, maybe 500-2,000, with a sparse on-bit population of say 2%) or a
>> sensory representation of say 128 bits with 20-30 bits on at any time. It
>> is important in both cases that you have overlapping bits semantically
>> significant (so you can subsample the on-bits and still approximate the
>> data).
>>
>>  Less than 4% is fine for input data as long as the 4% is semantically
>> good (ie that if the underlying meaning of the data changes semantically by
>> 5-10%, only 5-10% of the bits change).
>>
>>  Regards
>>
>>  Fergal Byrne
>>
>> _______________________________________________
>> nupic mailing list
>> [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
>
>


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

Reply via email to