Is the Mind a "Natural Intelligence" Large Language Model (LLM)?
The mind/brain has been compared to the software and hardware in a digital computer by decades of "computationalist" cognitive scientists.
We suggest that a much better, and much simpler, computer science parallel with human mental activity is the large language model (LLM) in today's artificial intelligence (AI). A chatbot reply to a question is prepared from pre-trained sequences of words and sentences similar to (with high "transition probabilities" from) the sequence of words and sentences in the question. Brain hardware may not be computer hardware, but brain software may closely resemble today's state of the art AI software.
Consider the past experiences reproduced by the
Experience Recorder and Reproducer (ERR) 1. They are past experiences which are stimulated to fire again because the pattern of current somatosensory inputs, or simply our current thinking in the prefrontal cortex, resembles the past stored experiences in some way. The ERR model is an extension of
Donald Hebb's hypothesis that "neurons that fire together get wired together." The ERR model assumes that neurons that have been wired together in the past will fire together in the future, as suggested by
Giulio Tononi. 2
We can say that the brain is being "trained" by past experiences, just as a large language model is trained with sequences of words and sentences. And like the LLM, a new experience or our current decision making will recall/reproduce past experiences that are statistically similar, providing the brain/mind with the context needed to interpret, to find meaning in, the new experience and to provide options for our decisions.
2
A new experience that is nothing like any past experience is essentially meaningless. It is without context.
Does this parallel between artificial intelligence software running on digital computer hardware and human natural intelligence software running on analog brain hardware make sense?
In the most popular consciousness models, such as
Bernard Baars' Global Workspace Theory or the Global Neuronal Workspace Theory of
Stanislas Dehaene and
Jean-Pierre Changeux, the fundamental idea is that information is retrieved from its storage location and displayed as a
representation of the information to be processed digitally and viewed by some sort of executive agency (or Central Ego as
Daniel Dennett called it).
Unlike computational models, which have no idea where information is stored in the brain, the ERR explains very simply where the information is stored. It is in the thousands of neurons that have been wired together (in various Hebbian assemblies). The stored information does not get recalled or retrieved (as computers do) to create a representation that can be viewed in a mental display. We can more accurately call it a direct reproduction or re-presentation to the mind.
Our hypothesis is that when (perhaps several) Hebbian assemblies of wired-together neurons fire again because a new experience has something in common with all of them, they could create what
William James' called a "blooming, buzzing confusion" in the "stream of consciousness." They would generate what James called
alternative possibilities, one of which will get the mind's "attention" and its "focus." Since each Hebbian assembly is connected to multiple regions in the neocortex, e.g., visual, auditory, olfactory, somatosensory cortices, and to multiple nuclei in the sub-cortical basal ganglia, like the hippocampus and amygdala, when one experience is freely
3 chosen all those brain regions that were activated by the original experience will be immediately bound together again.
Compare the way computational neuroscience stores and recalls an experience in memory. A central serial processor, or many parallel distributed processors in the connectionist picture, must digitize all the sensory inputs and transmit the data over the neural network, storing the bits at digital addresses that can be used to retrieve the data when it is needed later. The logical bits of data are presumably stored in individual neurons whose "all-or-none" firing corresponds somehow to the ones and zeros of digital information.
When the computer brain recalls a past experience, it must reverse the above, copying the stored data and sending the copy back across the neural network to form a "representation" of the original experience to be viewed. At this moment there must exist two copies of the original information at different places in the brain. We know very little about the cycle time of each instruction in the presumed algorithms used to store and recall memories of experiences, but it is likely very long compared to the instruction cycle time of modern computers. So computational neuroscience requires twice the memory store and is very slow.
A serious additional problem for computational neuroscience is how would the brain know which of the myriad past experiences to recall? A brute force approach would be to build and continuosly maintain an up-to-date index database of all possibly salient properties. Consider the
taste and smell of Marcel Proust's madeleines! Such essential properties are biologically embedded (not digitally encoded) in the Hebbian assembly of the original experience.
By comparison, the experience recorder and reproducer (ERR)
re-presents everything in an original experience
instantly, though no doubt with images weakened compared to the original, as
David Hume feared for his "impressions." The mind is "seeing" an original experience, not because the brain has produced a visual representation or display for a conscious observer to look at. This would require the computer equivalent of Descartes' homunculus! The brain/mind is also "feeling" the emotions of the original experience, as well as seeing it in color, solving
David Chalmers' "hard problem" of the subjective qualia.
As to how the ERR knows which past experiences are relevant to the current experience, no massive database of indexable properties is needed. Just as ChatGPT returns text that is statistically close to the language content of a query, those Hebbian assemblies that fire again in the ERR are those with neurons statistically close to, perhaps firing in, those in the new incoming experience.
The ERR is simply reproducing or "re-presenting" original experiences in all parts of the mind connected by the neural assemblies, solving the so-called "binding problem." This unification of each experience is because the information stored is distributed throughout each Hebbian assembly. All the cortical and subcortical centers that its neurons were connected to are immediately connected again.
The ERR is a presentation or
re-presentation to the conscious mind, not a separate representation on a screen as in a Global Workspace Theory and its "theater of consciousness." The fundamental philosophical question of how and where information is created, stored, and utilized in brain memory is answered very simply. It is in the strengthened synapses of the neurons that get wired together by each new experience. That information is never "processed," nor is it communicated or transmitted to other regions of the brain to be processed. It lives forever in a Hebbian assembly as long as the synaptic connections remain healthy. Much, if not all, lasts a human lifetime, although recalling it, reproducing it, "playing it back," may fail for reasons of "long term depression" (LTD).
In a break from computational neuroscience models of the mind, we can assert that man is not a machine, the brain is not a computer, and although the mind is full of immaterial information stored in the material brain, there are no "mental representations" as such, and mental information is not being processed digitally by a central processor or distributed parallel processors. A number of relevant experiences are simply "turned on" again, in an instant, likely faster than any computer program, considering the large number of ideas that simply "come to mind," automatically detected by elements in past experiences that are salient to the current experience.
We can also identify the Crick and Koch neural correlates
6 of a conscious experience. They just those neurons that were wired together in the Hebbian assembly created by the experience.
Our Natural Intelligence LLM is
human intelligence, built on the ERR model and the
two-stage model of free will endorsed by
Martin Heisenberg in 2010 as explaining "behavioral freedom" in lower animals such as fruit flies and even bacterial chemotaxis.
4 As such, the human mind can be seen as evolved from the lowest animal intelligence and even from single-celled organism intelligence, although bacterial experiences are not learned but acquired genetically.
In summary, we ask why decades of computational neuroscientists have imagined multiple digital processors moving bits of information from place to place in the brain, when the multiple neural pathways through the body and brain that are activated by an experience can simply be re-activated on demand, our memories played back in full by the experience recorder and reproducer, a purely biological capability created by natural evolution in almost all animals?
It's all because two brilliant thinkers (
Warren McCulloch and
Walter Pitts) imagined a logical machine could be built someday that would rival the ability of humans to solve problems in propositional logic put forward by
Immanuel Kant,
Gottfried Leibniz,
Bertrand Russell,
Ludwig Wittgenstein,
Rudolf Carnap,
Ruth Barcan Marcus,
Willard van Orman Quine, and
Saul Kripke. See our history of
computational models.
I wish experimantal neuroscientists would spend more time studying the mechanism that strengthens synapses (long-term potentiation) and weakens them (long-term depression).
Stephen Wolfram has concisely explained the workings of an LLM...
What Is ChatGPT Doing... and Why Does It Work?
It’s Just Adding One Word at a Time
That ChatGPT can automatically generate something that reads even superficially like human-written text is remarkable, and unexpected. But how does it do it? And why does it work? My purpose here is to give a rough outline of what’s going on inside ChatGPT— and then to explore why it is that it can do so well in producing what we might consider to be meaningful text. I should say at the outset that I’m going to focus on the big picture of what’s going on — and while I’ll mention some engineering details, I won’t get deeply into them. (And the essence of what I’ll say applies just as well to other current "large language models" [LLMs] as to ChatGPT.)
The first thing to explain is that what ChatGPT is always fundamentally trying to do is to produce a "reasonable continuation" of whatever text it’s got so far, where by "reasonable" we mean "what one might expect someone to write after seeing what people have written on billions of webpages, etc."
So let’s say we’ve got the text "The best thing about AI is its ability to". Imagine scanning billions of pages of human-written text (say on the web and in digitized books) and finding all instances of this text—then seeing what word comes next what fraction of the time. ChatGPT effectively does something like this, except that (as I’ll explain) it doesn’t look at literal text; it looks for things that in a certain sense "match in meaning". But the end result is that it produces a ranked list of words that might follow, together with "probabilities" 5
1. Doyle, B. (2016) Great Problems of Physics and Philosophy, Appendix E, Experience Recorder and Reproducer, p.394
2. Tononi, G. (2008) "A BOLD window into brain waves." PNAS (41) 15641-15642
3. Doyle, B. (2011) "The Two-Stage Model of Free Will," in Free will: The Scandal in Philosophy, p.186
4. Heisenberg, M. "The Origin of Freedom in Animal Behaviour", in Is Science Compatible with Free Will? pp.95–103
5. Wolfram, S. (2023) What Is ChatGPT Doing... and Why Does It Work?
6. Crick, F. and C. Koch, S. (1995) Nature 375, 121–123
Normal |
Teacher |
Scholar