opencog-dev team mailing list archive

[Andre Luiz de Senna <senna@xxxxxxxxxxxxx>]

Hi Trent.

Trent Waddington wrote:
> I was working out what my blowout is for using AtomSpace nodes and
> links to implement a neural network instead of using arrays of doubles
> and I figured out:
>
> Atom = type (4) + atomTable_ptr (4) + incoming (4) + outgoing (4*n)
>                 + indicies (4) + targetTypeIndex (4) + predicateIndexInfo (4)
>                 + atttentionValue (4) + truthValue (4) + flags (1)
>                 + vtable ptr
>          = 45 bytes minimum plus the size of the truth value, which is
> 8 for simple and 21 or more for indefinite and another 4 for
> composite.
> Nodes = Atom + name (4+len)
> Link  = Atom + trail (4) + outgoing (4*n)   <- uhhh, repeated?
>
> I notice that there is an outgoing vector in both Link and Atom.. is
> that intentional?

Humm... Link was not supposed to have it.

> Also can Nodes have outgoing edges? If so, what's that mean?

In early versions of Novamente, Nodes could be directly associated with other
Atoms via "outgoing" rather than via regular Links. Although this was the case
when we wrote the code I'm not sure this is the case nowadays. I suppose Ben
or Cassio may have something to say here.


> For a network I was looking at that had 3676 nodes and 1588000 links,
> I worked out that blowout was about 6 or 7 to 1.
>
> The algorithms would be slower too, as there's now traversal that has
> to be done, instead of straight array accesses.
>
> So I wonder if some kind of conceit is necessary.  Like a "layer node"
> and a "totally connected link" or something.

If your NN algorithm don't have benefits from possible interation
with other Novamente entities (PLN etc) I don't see why you
should use the AtomTable node hierarchy to represent your
networks.

I'm sorry if i missed some important discussion regarding this
design. If that's the case please point me to the proper threads.

Best regards.

Andre Senna