Some Criticisms of the Neo-Darwinian Theory of the Mind

Dennis Hackethal has proposed a theory he calls “The Neo-Darwinian Theory of the Mind”, which is intended to be a low-level explanation of how intelligent minds work. The central concept of the theory is that ideas within a mind are replicators that compete with one another for prevalence in the mind. The evolutionary dynamics that arise from this evolutionary competition supposedly explain creativity, just as the competition between genes within gene pools explains the origin of adaptivity in biological evolution.

While I am glad to see people exploring Critical Rationalist approaches to AGI, I don’t think that this theory holds up to criticism. I have several major criticisms of theory which I think show that the theory is neither a good approach to implementing artificial general intelligence, nor a plausible evolutionary account of how general intelligence arose in humans.

The Inefficiency of Replication

According to Dennis’s theory, the ideas which manage to survive in a mind are those which are most successful at replicating themselves, and thus minds will tend to have many copies of each individual idea, especially those which are most evolutionary successful. However, from a computational perspective, creating a new copy of an existing piece of data, such as an idea, is generally wasteful. A mind that constantly made exact copies of ideas that already exist would end up unnecessarily wasting a lot of memory space.

There’s a way to modify the theory that solves this issue with memory space: Rather than allowing ideas to replicate, simply have an integer associated with each idea, and increment that number each time the idea would normally replicate. Tracking these values, which could be called the “weight” of the corresponding ideas, would allow the mind to behave exactly as it would if it contained replicators, but without any of the wasted space.

I’ve mentioned this criticism to Dennis in the past, and in responding to it in his article explaining the theory, he said:

There has been an ongoing, friendly point of contention between the intelligence researcher Ella Hoeppner and me about the role of replication in any theory of evolution. Citing Donald Campbell’s evolutionary epistemology, she argues that replication can drive evolution, but is not necessary. All that’s needed, she argues, is variation and selective retention. I am agnostic on the question of necessity and think that if we introduce replication in a mind, we can explain a great many things that we couldn’t (at least not as well and as easily) without it.

Dennis Hackethal, The Neo-Darwinian Theory of the Mind in Conjecture Magazine

However, I don’t agree that introducing replicators helps explain anything that isn’t explained just as well, and just as easily, by the variant theory that I’ve proposed. The point of the variant theory that I’ve suggested is that, aside from having a smaller memory footprint, the way it behaves should be precisely the same as the way Dennis’s theory would behave. Whenever an idea would thrive in Dennis’s theory, it should also thrive in the variant theory. The difference is merely that “thriving” means increasing the weight of an idea, rather than creating additional copies of the it. So the including replicators doesn’t help the theory explain anything at all, and only serves to make the theory less computationally efficient.

The Evolutionary Disadvantages of Idea Replication

Dennis has argued that his theory helps explain how general intelligence evolved in humans, like so: The minds of some pre-human ancestral species contained set of ideas that remained relatively unchanged over the course of a lifetime, and at some point a genetic mutation occurred that allowed one or more of the ideas within a mind to begin replicating. Once this replication started, a competition between different kinds of replicators began, and this process eventually gave rise to general intelligence.

I think that there is a big problem with this explanation. When an idea began to replicate in the mind of a species that had never contained replicating ideas before, it would likely be detrimental to the organism’s survival, and so this trait wouldn’t persist. As I mentioned in the last section, replication can quickly eat up a lot of memory, since each new copy of an idea requires some dedicated memory space. It’s unlikely that an brain not yet adapted to containing replicating ideas would have any free memory space for the new ideas to use. If any ideas were to successfully replicate, they would take up space that would normally be used by other ideas, and this would almost certainly be detrimental. The emergence of replicating ideas in a mind not suited for them would be akin to a form of brain cancer.

Emergence of Higher Level Epistemological Dynamics

The problems that I’ve pointed out so far center around the issues caused by replication. However, I’ve already proposed a variant theory that behaves identically to Dennis’s theory, but without relying on replication. Is this variant theory, which avoids several of the problems original theory that the original theory had, a plausible account of general intelligence?

Unfortunately, I don’t think so. Dennis’s theory is intended to describe a low-level system out of which the high-level properties of intelligent minds, as described by Popperian epistemology, will emerge. However, I don’t see why the high-level dynamics of Dennis’s system would resemble general intelligence.

From Popper’s epistemology, we know that intelligent minds work by an iterative process of conjecture and criticism. Dennis’s theory explains how conjecture can occur: Occasionally, during replication, the newly created idea ends up being slightly different from the original idea in some way, rather than being perfectly identical. This is analagous to how new genes are created via genetic mutation in biological evolution.

However, I don’t think that Dennis’s theory satisfactorily explains how criticism occurs. Criticism, in the Popperian sense of the term, refers to finding a contradiction between some set of ideas. Intelligent minds work by looking for such contradictions between their ideas and attempting to resolve them. So, for Dennis’s theory to be correct, the replicator system that he proposes would need to somehow lead to the mind searching for and attempting to resolve contradictions between ideas.

While it’s possible that such dynamics may develop in a system like Dennis’s, it isn’t obvious exactly how they would, or why there would be any selection pressure for them to occur. Without an explanation of how and why such dynamics would evolve, I don’t think the Neo-Darwinian Theory of the Mind provides a good explanation of general intelligence.

5 Replies to “Some Criticisms of the Neo-Darwinian Theory of the Mind”

  1. Interesting thoughts, Ella.

    I like that Dennis tried to find a connection between Neo-Darwinism and AGI. And I actually think that is an interesting analogy worth exploring. But I think your criticisms are strong as to why this isn’t the most plausible approach.

    I also might add one other criticism that came up from David Deutsch on this. He points out that even if the way human general intelligence works via replicators were true, that isn’t a good reason to model AGI that way if there is a more efficient way. This seems to be a confusion between levels of explanation equivalent to trying to study neurons to understand AGI.

    There is a stronger problem, in my opinion, however. Neo-Darwian Mind theory doesn’t seem to solve any problems or really explain anything directly related to AGI. The few cases where it might shed some light, such as having a memory on the tip of your tongue, aren’t directly related to AGI and have alternative explanations that are quite strong. (This might just be a search problem in the mind, for example.)

    So this theory isn’t compelling because it doesn’t really solve anything that seems to matter that much and what it does solve is not directly related to AGI and has strong competitors. It goes against Popper’s epistemology which starts with problems and then tries to solve them. This is more like a solution looking for problems to solve.

    In short, the “Neo-Darwinian Theory of Mind” isn’t really yet a good hard-to-vary explanation at all and so it isn’t yet a theory. (Except in the very loose sense that all conjectures of all sorts are ‘theories.’) It’s an ambitious research project masquerading as a good explanation at this point.

    However, research projects always start out weak like this. It’s impossible to know if it might still lead to bearing fruit down the line. “Let a thousand flowers bloom.” But to bloom, the ‘theory’ needs to take it’s criticisms seriously and try to solve them. I think your criticisms are good examples of the sorts of strong problems a final theory must solve. Moreover, it must solve some compelling problems about AGI. Right now I don’t see it doing that, it just draws an interesting analogy.

    But if I might compare to Einstein, his initial idea that led to GR was not really that much better. It was a simple realization that moving through curved space would be identical to having a force act on you. So I don’t think this means this isn’t a good research project that Dennis should pursue further and see if he can solve the problems the conjecture currently has.

  2. Most genes don’t know how to replicate, they just tag along for the ride. Even replicating genes do it indirectly through the machinery in the cell. Anyway, I guess the appeal in using replication and mutation to explain creativity lies in the astonishing reach it has. Just about anything can be brought into existence, just like creativity is capable of unbound knowledge creation. Despite this strength, idea creation by mutation seems to me to be a triviality compared to idea selection. Just use any random process. There must be faster ways than endless self supported copying waiting for a mistake to be made. And as we have seen with genes, it is the whole genome that gets copied, not genes in isolation. It all adds up to a lot of ideas. Being on top of the copy count doesn’t seem to be very important either, as a single copy with the right knowledge easily outcompetes the bunch.

    Over to selection. Reality kills non viable genes. What kills non viable ideas? Could it be problems? The mind is at least in part a reality simulator. Reality is problem free, therefore a good simulation strives to be problem free as well. Problem recognition then, is it nothing more than deduction that ends in (p and not p)? The precision in representation that is required for this seems almost unreal. How is the same p deduced from two sets of unrelated ideas, created by a random process? This p, like qualia, must be some kind of fixpoint that supports its own stability through feedback. Maybe creativity is not random at all, but very precise in what it does.

    Genetic mutation creates knowledge about nothing in particular and succeeds as long as it doesn’t cause the knowledge to be destroyed. The mind creates knowledge about reality to simulate it faithfully.

    1. Wow! Good thoughts Marcus! I like where you are going with this.

      “Despite this strength, idea creation by mutation seems to me to be a triviality compared to idea selection. Just use any random process. There must be faster ways than endless self supported copying waiting for a mistake to be made. “‘

      Let me add here. What you are really getting at is that the big mystery we need to solve is not “how to replicate idea” nor “how to randomly create new ideas” (we know how to do both of those!) but “how do we come up with a conjecture process that is non-random and even directed but is still open-ended enough to potentially solve any problem?” (Where you are defining problem as mismatches from reality. Which I like.)

      This is what I see as the funamental flaw with both Ella’s and Dennis’ theories. I believe they are trying to solve the wrong problems. What we need are theories about how conjectures are made. We should not expect the random process of Ella’s theory and Dennis’ theory to be the correct one. In fact, we have very good reason to believe them to be wrong.

      That doesn’t mean both theories don’t have something important to contribute, I might add. I actually think Ella’s theory is fabulous in how it solves the question of what is a problem. But I think that is progress against something that wasn’t really the central problem of AGI.

      Or that is my opinion right now. I might be wrong. But I think this is what you might be getting at as well.

      1. What reasons do you think there are to believe the random process that I’ve proposed for idea generation is wrong? I don’t think we’ve talked about that before. I’d be interested to hear you expand on that.

        Also, I think you might be misinterpreting what Marcus was saying. What I interpreted Marcus to be saying was:

        The way in which ideas are created isn’t so important. The selection process, the process by which ideas are narrowed down once they’ve been created, is much more important.

        And I agree with that. So long as the process by which ideas are created is at least partially blind, it involves taking bits and pieces of existing ideas and combining them into something new, and the range of ideas that can possibly be created is sufficiently large, the details are relatively unimportant. And achieving all those properties is quite trivial. So coming up with theories of how conjectures are made isn’t the hard part. The hard part is about exactly the mind does with ideas once they’ve already been created: how it decides which ideas to keep and which to throw out.

    2. Thanks for sharing your thoughts Marcus.

      I agree that the details of the mutation process for creating new ideas are relatively unimportant, and there are a huge variety of random processes that would work. Presumably some random processes will turn out to be better than others, in the sense of allowing the mind to learn more efficiently, but worrying about that right now is premature. We’re still at the phase where we’re just trying to figure out how to build a single, proof-of-concept AGI, and it will be much easier to figure out which mutation processes are most efficient once we’ve got at least one example of a working AGI.

      I do think that problem recognition boils down to deducing things like “p and not p”. I don’t think I understand why you think the precision required for that seems unrealistic, I see no problem here. If you start with an empty mind and begin adding new, random ideas, I agree its unlikely that any of the first few ideas will lead to contradictions with one another, since its unlikely that any of them will produce precisely the same p. But as the number of ideas grows, the likelihood of that happening only goes up, as the ideas in the mind deduce more and more unique ps. Eventually, in a mind with enough ideas, contradictions will be happen quite frequently (especially since, I think, minds have some kind of preference for “bold” ideas, ideas with a lot of logical content).

Leave a Reply

Your email address will not be published. Required fields are marked *