A Cognitive Architecture for Machine Consciousness and Artificial Superintelligence: Thought Is Structured by the Iterative Updating of Working Memory

[KnutJaegersberg]

A Cognitive Architecture for Machine Consciousness and Artificial Superintelligence: Thought Is Structured by the Iterative Updating of Working Memory

This is a lengthy paper describing a cognitive architecture.

https://arxiv.org/abs/2203.17255

The youtube video is easier to grasp imo. The theory is from 2016, also described here:
http://www.observedimpulse.com/2018/05/a-neural-model-of-working-memory-and.html

https://www.youtube.com/watch?v=R2H2Pl0I6EA

It describes a mechanistic way by which working memory might work. I think it could explain in part why GPT works or why it might be an imperfect imitation of the human mind.

[KnutJaegersberg]

it's not mainstream theory, and there is no definitive evidence. I think it's missing aspects, but I feel it is a roughly good approximation how it might work.

[KnutJaegersberg]

it explains the narrative nature of working memory

[andreaskoepf]

Added it, but still need to read. My first impression is that the ideas are less different to others than presented... seems to deal with different forms of memory and attention, associative search .. which probably fall mostly in the system-1 area.
Will update once I have fully understood how he proposes the actual reasoning to happen...

[KnutJaegersberg]

check out his video presentation, it's really good. it's a bit lengthy, but well done.
I find his papers are not very accessible, though they tell the same story. The video made it easy to digest.
according to this perspective sys2 is sys1 operating / biasing itself in a certain way.
he makes connections between consciousness, general intelligence, narration, fractals.
I don't think there is enough hard evidence so that this theory is generally accepted in the scientific community,
but it makes A TON of sense to me. sometimes one maybe doesn't have to take him literally, but the gestalt of the theory, it feels very real to me.
Heidegger suggested, that which must be thought commends us to think it. Reser describes a mechanism that explains how that happens

. How that happens may only happen by approximation as he describes it, as kinda search for similar contents. I think he is not enough of a dynamist in his thinking, but overall, I think the thing works. I think it is useful for building AGI and it even explains why GPT works as well as it works.

this is the key concept. the words that describe it are a nested concept, but the video makes it actually intuitive to understand.

https://www.youtube.com/watch?v=R2H2Pl0I6EA

[KnutJaegersberg]

what's fascinating is that the later part is all about the brain and structural underpinnings which make the thing work. that again is compatible with newer accounts of dynamicism in cognitive neuroscience.

here is another video on dynamical systems tehory, which should be watched with this one.
these two together describe a practical perspective on mind which I've adopted for now.
the latter part of that video also is somewhat on the interplay of sys1 and 2 though they don't talk about it explicitly (more about how central attention works in that perspective).
That part is not really reflected in how static LLMs work.

https://www.youtube.com/watch?v=19ZqeQzXVV4&t=2271s

[KnutJaegersberg]

I've always liked Cowans and Oberauers theory on working memory, those are fairly popular.
https://de.wikipedia.org/wiki/Embedded_Processing_Model_of_Working_Memory

Reser developed that perspective further.

[KnutJaegersberg]

a long time ago I bought "Models of working memory: Mechanisms of active maintenance and executive control"
that book gives you a fair idea about now older theories of working memory. That said, a lot of psychology introductory books are still stuck on the old concepts of Baddeley and Hitch.
That book although outdated is still worth reading, as it helps with the big picture.
Another concept that influenced my understanding is long term working memory, also from that book.
It's helpful for understanding skilled access and expert performance. I think thinking in those lines can help improving AGI composite systems to have better information processing flows, i.e. in RAG.

https://annas-archive.org/md5/2747c0f55a48a73e9b5f6df94b6841ee

[KnutJaegersberg]

It doesn't contain LIDA, because that came thereafter, which is a big miss. Reser also implies LIDA in his account, it's quite powerful, while the actual LIDA tells a better digestible story on cognition as a whole. LIDA stitches together the main concepts of cognitive psychology, including placing system 2 in iteration steered by attention.

[KnutJaegersberg]

overall there are some more structuralist perspective on mind and some which are more connectionist. first the structuralist accounts became popular, because it was easier to come with, it's intuitive to think of parts of a mind.
But I am biased towards accounts that allow for a lot of emergence, Resers account is one of those.

[KnutJaegersberg]

When I first read his other paper, I didn't really get it either, it's very artificially expressed. That's why that video is so helpful.

[KnutJaegersberg]

talking about this paper, where he first described it.

https://www.sciencedirect.com/science/article/pii/S0031938416308289

[KnutJaegersberg]

the thing is, even if one rejects this account for the human mind, I think it is very relevant for AGI, still.

[KnutJaegersberg]

I feel the distinction between sys1 and sys2 is a conceputal tool for humans, but actually it's a continuum of processing style.

[KnutJaegersberg]

in a lot of daily circumstances, you are neither 100% sys1 or sys2. it's just a helper to talk about these things, and although clearly sys 2 requires special brain regions involvement, I don't think it's not really 2 different apparati.

[KnutJaegersberg]

sometimes, sys2 components are in use rather unconsciously, too. the line is more blurry than the discinction of dual process theory suggests.

[KnutJaegersberg]

that's just my feeling about this.

[KnutJaegersberg]

I feel if one is preferring system 2 style processing, one is more likely to see this as a binary thing, because it's useful to think like that. But I don't think it's the truth. Reflective, intellecutally inclined people also quickly adopted the concepts of Freud and Jung when they were new. They were still wrong in some ways.
We want to give order to our environment, we crave for it. When I think about Resers perspective, I think about cognition biasing cognition for evolutionary advantage giving rise to system 2 style processing. That's way more fuzzy than having 2 categories of something, like sys1 sys2, the unconscious and conscious parts of mind, etc.
I'm sold to the idea that our mind is a chaotic system, a very elegant one.

[KnutJaegersberg]

for example, when you're in default mode, that can still have characteristics of system 2, i.e. when you're engaging in self-reflection. same for imagination, while that is more associative in nature, it's often kinda steered into a direction. sometimes a daydream can tell you something about your interests. it can also be very random. but not always.
I think being on the extreme end of the two is not evolutionarily advantageous. if we're totally sys 1, we quickly make stupid mistakes, some can cost our lives (they do occasionally). if we're totally in sys 2, we might be so narrowly focussed that we miss something important that changes what we ought to be busy with for our own surival, cognitive flexibility, an executive function, requires both sys1 and sys2. It's often mixed imo.