Dear Fergal,

In my view, this CLA model is located between implementation of the
functions of memory and recognition. The brain performs other functions
such as planning, reasoning. Probably relevant external programs will be
added to the model  later. The brain generates the text for explanation his
actions or for conscious reasoning. Various concepts of the language (text)
have a semantic hierarchy that can be placed within one field of cortex.
This hierarchy does not require a structural model hierarchy as recognition.

In my view, this CLA model can store and reproduce (predict) the time
sequence due to the temporal order and due to the time scale between the
levels of the hierarchy. First means that the time classifier will look for
the best match (maximum correlation) for more than one time cycle. Second
means that you do lossy compression information when going from the lower
layer to the upper layer. Of course, the loss of information is a schematic
taking into account a lower storage layer.

I think, more clear division of tasks and concepts will help the progress.

Regards,
Ivan

On Date: Mon, 9 Sep 2013 22:40:16 +0100
From: Fergal Byrne <[email protected]>
To: "NuPIC general mailing list." <[email protected]>
Subject: Re: [nupic-dev] Inter-layer plumbing
Message-ID:
        <cad2q5yek9b7spzhy-vbrgmv++h3q+8aaedvevqogzayqqpn...@mail.gmail.com>
Content-Type: text/plain; charset="windows-1252"

Hi Chetan,

At the risk of using either the word "artefact" or "hack," we need to
recognise that we're still working on the single-layer of a single region
which is trying to predict a single value a single number of steps ahead.
We haven't figured out universal ways of feeding this guy its data, and
we're avoiding for now the whole question of how the different layers work
and how to connect CLA's in a hierarchy.

So for now, we have encoders, classifiers, and swarming to help us put
together the missing infrastructure around our tiny 1mm squared layer of a
region. I'm really interested in the idea of a neurologically justified
design involving adaptive encoders which simultaneously incorporate
encoding parameters, classification and swarming in a single structure.
This design will likely have an analogue with inter-layer and inter-region
plumbing.

If that's correct, it'll hopefully give us the initial machinery for
building hierarchies. There are all sorts of computations humans do which
require a hierarchy to even begin to figure out.

Regards,

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

Reply via email to