Hi Jeff,

Thanks for the clarifying points.

Apologies if certain words trigger an emotional response, but unfortunately
they're all we have. I deliberately used the word "artefact" in my comments
to describe the swarming, I've found that "hack" is both a better
description (inside my head) and a terrible one (inside everyone else's)
due to the connotation that it somehow doesn't measure up to some unspoken
"standard".

The process of modelling some natural phenomenon cannot proceed from zero
to 100% in a flash, and we must use hacks appropriately to bridge the many
gaps in understanding and modelling which necessarily exist. Hacks are not
only necessary to all progress, they are the highest form of creativity as
they embody the "most for the least" in a highly focussed way.

Galileo, lacking the timepieces to time dropping objects accurately, hacked
the problem by rolling them down a ramp, thus slowing the falling time down
so he could make more accurate observations.

One of the reasons why your work is already bearing fruit is that you've
been prepared to attach artificial structures around the CLA in order to
give it life and observe its performance. As long as these structures do
the same sort of thing as the natural structures, you can make progress.
This is similar to using mock objects in test-driven development - it's
absolutely key to focussing on the work you need to do.

The classifier is a good example. Obviously we don't have such a beast in
our brains, but we do have multiple layers, hierarchy, the thalamus,
attention, and all that goes with those. The classifier is a best first
placeholder for all that, and it allows us to get something useful out of
the CLA.

That being said, I stated that the swarming process is a genetic algorithm
in the evolutionary sense, in that it begins with some models and evolves
them based on fitness metrics. All our sensory processing has been
hardwired by evolution to some extent, and swarming is a design process
which emulates this evolution in an engineering-practical timescale.
Perhaps "genetic design process" is a better description.

I used the example of the phoneme cull because (according to Pinker) it's a
scheduled event, a phase shift in that part of the brain. This is
equivalent to a change in model due to a new round of swarming, following a
sufficient amount of data. It appears that this abrupt event (and many
more) free up so much connectivity that it enables the next big leap in
linguistic abilities. It's akin to the metamorphosis found in moths and
butterflies, allowing new organs of thought to emerge from a previous
generation.

Stewart,

The advantage of using MySQL is that the co-operating workers have a joint
noticeboard. MySQL is specifically designed to handle issues like race
conditions and multiple collaborating agents. It's also ubiquitous and
easily installed, it's practically an OS function.

Regards,

Fergal Byrne



On Mon, Sep 9, 2013 at 8:02 PM, stewart mackenzie <[email protected]>wrote:

> Hi Ron,
>
> Great work on your hypersearch implementation! I recently went through
> the code before I saw the mailing list announcement and saw your
> code... "could it be!?... so soon!?"
>
> The video clearly gets your point across. I had to rerun the bit where
> the converging points converge on the global optimal, I enjoyed it so
> much!
>
> Kudos!
>
> A few questions:
> - MySQL adds an extra dependency, would it not have been better to
> just make use multiple flat files or keep it in memory instead?
> - Why not take it a step further and directly expose the CLA via an
> API and have hypersearch constantly adjust the parameters directly? So
> I would imagine an always running hypersearch would scale back the
> processes once the global optimum has been found, but still keep a few
> hanging about double checking. Periodically based on certain data, it
> could research the space to ensure the found global optimum is still
> optimal.
>
> Kind regards
> Stewart
>
>
>
>
>
>
> On Sun, Sep 8, 2013 at 11:26 PM, Matthew Taylor <[email protected]> wrote:
> > Swarming in NuPIC, with Grok engineer Ron Marrianetti:
> >
> > http://youtu.be/xYPKjKQ4YZ0
> >
> > Enjoy!
> >
> > Matt
> >
> > _______________________________________________
> > 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
>



-- 

Fergal Byrne

ExamSupport/StudyHub [email protected] http://www.examsupport.ie
Dublin in Bits [email protected] http://www.inbits.com +353 83
4214179
Formerly of Adnet [email protected] http://www.adnet.ie
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to