random viedo in my feed that reminded me I need to post here:
https://www.youtube.com/watch?v=gvjCu7zszbQ

Come to think of it, I have a growing pile of unfinished drafts that I
had been composing for this list, all new material, but where I was
either unsure of my premise or didn't have enough material to present
the idea the way I like to, or just lazy... =\

[[ brief rundown of other posts in my drafts folder: P-zombies and
facing the problem of consciousness, an attempt to look at the AGI
problem from the perspective of pure computer engineering, and a
brainstorm about obtaining nano-scale computing elements using
crystals...]]

If this post ever sees the light of day, it will talk about the major
revolution desperately needed in DNNs right now and argue that point.

https://heartbeat.fritz.ai/deep-learning-has-a-size-problem-ea601304cd8

The excessive size and hardware requirements of contemporary AI is
basically driven by a number of factors, first there's the extreme
demand for AI and AGI technologies, so we have this one somewhat
good-ish learning algorithm that has become the proverbial hammer that
makes everything look like a nail. This has given rise to an industry
around reaching higher and higher performance through expanding the
scale of DNNs using ever more hardware.

One school of thought seems to be that "Well we are still only 1/100th
the scale of the actual biological brain, we just need to scale up".
No... Not at all. WROONG!!! =P

What is going on is we are trying to re-build your desktop computer with
a feed-forward network of NAND gates... try to imagine that. While it is
mildly breathtaking what has been accomplished with the current
paradigm, the approach has hit a brick wall.

WE NEED TO CLIMB THE CHOMSKY HEIRARCHY, PEEPS!!!!!

MUST HAPPEN!!!!

So what will this look like? We need to solve the cortical column
problem. We need a trainable sub-net of some reasonable size and
complexity that can be tiled, and a meta system that can solve tasks
given to it by recruiting sub-sets of these columns and operating
ITERATIVELY using those columns until the problem is solved. We can
continue to use some of the existing learning algorithms but the
meta-system needs a learning solution too. I know a lot of you just post
reflecxively, but actual human beings are supposed to stop and think
about things AND THEN post. That is what we need to do, instead of a
feed-forward network, we need a system that can ruminate properly.

Such a system would be radically more efficient than existing systems
just as your computer which operates sequentially is radically better
than a feed-forward network of NAND gates.

AGI is still a few steps beyond that but the goal posts are in sight.
Ten hut! hut! HIKE!!

-- 
The vaccine is a LIE. 
#EggCrisis     
White is the new Kulak.
Powers are not rights.


------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T15fa07153b838749-M8409def430ea8c5bec44bb15
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to