On Wed, Jun 29, 2022 at 11:11 PM Boris Kazachenko <cogno...@gmail.com> wrote:
> On Wednesday, June 29, 2022, at 10:29 AM, Rob Freeman wrote: > > 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. > > > Isn't that what backprop does anyway? > They may use a pre-learned relational principle in a sense. Pre-training, you say? But they then revert to back-prop again? That's fine. I guess back-prop would go on to learn hierarchy, just as it learned the grouped predictions at any "pre-training" level. But that is not quite the application as a foundational principle I was talking about. If you actually, actively, substitute things which share predictions, that's a little more foundational than just using initial groupings as a basis for more back-prop. You could do either. I'm suggesting that if you end up getting a different grammar for each sentence, the second way, just actively substituting things which share predictions, and not doing more back-prop on initial groupings, is a more efficient way to do it. Because going the active substitution way just generates groupings for each sentence as you go. Back-prop is trying to optimize over the whole data-set. If there are not actually that many global optimizations, if there are actually an infinite number of chaotically expanding, global optimization attractors, it becomes horribly inefficient. It's a hypothesis anyway. I don't know if anyone has looked at any hierarchy generated by a transformer, as in the paper I linked, and checked to see if it is a different hierarchy for each sentence. Personally, the fact these things are just churning out more and more, billions, of parameters, is a hint to me. But I don't know if anyone has checked. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T5d6fde768988cb74-Md02add195b2316d212728693 Delivery options: https://agi.topicbox.com/groups/agi/subscription