I think I gave a biologically possible implementation of this reconstruction; it probably does this better than classic back-prop models. I also think it is more easy to understand through the idea of mirroring between a (re)constructing and deconstructing neural network.

In the HTM model biologically this reconstruction feedback would go to cortical neural level 6.

Bert

op 06-03-14 00:08, Ian Danforth schreef:

Using the same weights going up and coming down isn't biological but reconstruction in general is fundamental to biological intelligence.

Ian

On Mar 5, 2014 10:20 AM, "Subutai Ahmad" <[email protected] <mailto:[email protected]>> wrote:


    Yes, that is what the old reconstruction algorithm tried to do. I
    looked around and found an old writeup of it which I've attached
    here.  The language in it is outdated, but perhaps still helpful?
     Section 3 (top down compute) is the most relevant.

    Historically, we stopped using it in nupic because the classifier
    based approach worked much better in the OPF for prediction.

    Also, this algorithm is not at all biologically realistic.
    However, it may still be quite useful!

    --Subutai


    On Wed, Mar 5, 2014 at 9:49 AM, Marek Otahal <[email protected]
    <mailto:[email protected]>> wrote:

        If by the 'old reconstruction' we mean that each block (SP,
        TP, encoder) can process in reverse direction (top-down)? then
        YES for me. I'd have lots usages.
        I'd be willing to mentor this task in SoN.


        On Thu, Feb 27, 2014 at 6:49 PM, Chetan Surpur
        <[email protected] <mailto:[email protected]>> wrote:

            In order to use CEPT SDRs with both the spatial pooler and
            the temporal pooler (rather than just the TP directly), we
            will need some form of reconstruction. This will allow us
            to bring back the output of the TP into the original CEPT
            SDR space.

            This reasoning applies to any predetermined SDRs that you
            want to feed into a region and get a prediction for.

            On Feb 27, 2014 9:37 AM, "Matthew Taylor"
            <[email protected] <mailto:[email protected]>> wrote:

                Is the old reconstruction algorithm a good idea and
                should we bring
                back a working implementation in NuPIC? This is a
                meaty topic that
                Subutai and Ian have started on Github[1][2]. We'd
                like to have this
                conversation here on the mailing list before we create
                further tickets
                and take action.

                Please chime in if you have an opinion.

                [1]
                https://github.com/numenta/nupic/pull/421#issuecomment-30059730
                [2] https://github.com/numenta/nupic/issues/683

<<attachment: BertF.vcf>>

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

Reply via email to