Another property of neurons is that they tend to
compress or auto adjust, so when a consistent
input is seen, the output will at first be quite active,
and decay to a baseline as it becomes less sensitive
to the current non fluctuating input. This has the effect
of making what sounded very loud at first, not as loud
later. In the case of vision, putting on rose colored
glasses, gives the viewer a very pink color shifted
image, but after an hour, the white point has been
adjusted so the redness goes almost unnoticed.

This is generally a really good thing, otherwise our
attention would constantly be pulled from one thing
to another by regularly occurring inputs that don't
really need the full focus of our attention, such as
feeling the cloths we wear or the reverberations of
sound inside a room.

I think this property would be good to implement
at least at the encoder level, but perhaps throughout
the CLA. An obvious issue with this to me is that given
a consistent stimuli to a sensor for a long period of
time it would at first generate an SDR that looks different
than what is generate later on, possibly making the
input unrecognized. The benefit, I believe, is it would
then be focusing on the faster changing input, which is
probably more important for the CLA to generate
meaningful and useful predictions.




Patrick





On Aug 2, 2013, at 5:00 PM, Scott Purdy wrote:

> I had promised this a long time ago but here is a version of an adaptive 
> encoder that isn't quite as brittle as the current adaptive encoder:
> 
> https://github.com/numenta/nupic/pull/155
> 
> It is loosely based on the idea of arithmetic coding and adapts to arbitrary 
> value ranges, while typically not changing representations by more than a bit 
> when a new value comes in.  The main downside is that it requires a moving 
> window of values seen and the size of this is arbitrary.  When I tested it I 
> did not find that it worked very well.  Obviously if you can fix the 
> encodings like the regular scalar encoder then you don't have to worry about 
> losing learning when the representations adapt.  But here it is in case you 
> are curious.
> 


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

Reply via email to