On Wed, Jun 29, 2022 at 2:19 PM John Rose <johnr...@polyplexic.com> wrote:

> ...Bob Coecke’s spidering and togetherness goes along with how I think
> about these things. The spidering though is a simplicity, a visual
> dimension reduction itself for symbolic communication coincidentally like a
> re-grammaring of representation. But I like it a lot, it's great, the
> ZX-calculus, etc..
>

I like Coecke's work. It's putting a finger on this structural
indeterminacy which I think has been preventing us from finding adequate
representations for cognition, since forever.

But Coecke is starting top down. That makes the formal representation
implications clear, but complicates the computational problem.

I think it is more powerful to start from the learning procedure
perspective. Then these things are generated organically, and you don't
have to go to special maths to represent them.

Given the learning procedures, there is no reason to sweat bricks to group
general representation formally, and then sweat bricks using quantum
computing to collapse it down to specific cases again. You can just
generate the one you want, when you want it.

I hypothesize that transformers are generating structures which might well
be grouped formally in the ways Coecke does. It's just hidden because we
don't pay attention to the internal structures at all. So transformers
might actually be a representation for Coecke's formalisms.

The problem with transformers in that case, is only that they are trying to
collapse all possible observations beforehand.

You might use the same learning procedure that the transformers use, but do
it only at need when presented with a particular problem.

Well, you wouldn't use exactly the learning procedures transformers use.
You wouldn't use back-prop on a dot product to learn the relational
principle implicitly from the task. You would start with the relational
principle those dot products learn, by which I mean grouping things
according to shared predictions, make it instead a foundational
principle, and then just generate groupings with them.

------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T5d6fde768988cb74-Mcba94316fd2639f8b7d5d45a
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to