[svnt]

It’s an intriguingly framed paper, and others have written about how gene regulatory networks store weights, etc, but this seems to me like it is putting too much emphasis on the direct mapping of the developmental process to a generative process just because it is popular at the moment.

The encoder being evolution is an idea that has been developed by Sui Huang and numerous others.

The genetic decoder being creative/generative is an idea that has been put forth by Richard Watson and others.

What I’m more interested in at this point is how do the layers interact? Rather than waving at energy landscapes, stable attractors, and catastrophe theory, why can’t we put this to use producing alife simulations that deliver open-ended evolution?

[sam537]

Popular and incompletely understood. That analogy is unhelpful to understand processes. It's attempting to explain a incompletely understood phenomenon with another. As someone else said below, the problem of the genome being the 'compressed representation' is reductive since genomes do not exist in isolation in an organism. Genomes were environmentally selected for millenia and many unicellular organisms share genomic material with other organisms or the environment. I think the framework is useful but it does not scale.

[samuell]

I find the analogy itself super interesting and thought-provoking, and might be quite useful in studying biological systems.

The idea that we need to move away from understanding the genetic code as static machinery also aligns well with recent understanding of biology, as summarized in the highly appraised book "How Life Works" by Philip Ball [1]

But in terms of evolution, I don't see how the proposed analogy/model will get away with the fact that natural selection operates on the individual level (either you survive or you don't), while all genomic information is a package of (depending on the organism) humongous number of genes, not to mention base pairs or individual locuses (that's not even mentioning diploidy, that will mean two different copies of each locus).

With the known proportions and numbers of positive vs slightly deleterious mutations (much more of the latter, counted in the hundreds per individual), selection of positive mutations can not avoid accumulating slightly deleterious mutations.

I don't get how the proposed model is supposed to solve that.

I see that a smartly designed system could probably have processes that can change multiple loci in parallel in a beneficial way (along the lines of the theory of facilitated variation by Gerhart & Kirschner. See their papers or [2]), but that would only explain how such a fine-tuned system could be effective at adaptation, not how the system itself - including its processes - could arise from a state before these processes are in place.

To connect it to ML: In natural selection you don't have gradients and the ability to update multiple parameters based on detailed feedback on them individually. You only can provide a whole, binary feedback (survive, 1, or not, 0), to the whole set of parameters, whether they are slightly deleterious or possibly positive. The resolution is simply lacking here.

- [1] https://www.amazon.com/How-Life-Works-Users-Biology/dp/02268...

- [2] https://www.amazon.com/Plausibility-Life-Resolving-Darwins-D...

[bubblyworld]

I think it's more useful to think of natural selection as acting (probabilistically) on populations of genomes, not individuals. The feedback is individual, but the "gradients" are at the population level. It's not a perfect analogy but e.g. there are formal correspondences like this one: https://www.nature.com/articles/s41467-021-26568-2

[RaftPeople]

An interesting fact that I wasn't aware of until I read it recently is that our genes constrain the chance of mutations in critical areas of the body, which shifts the landscape.

[stonogo]

Mutation resistance is itself the result of mutation (i.e. evolution), and isn't anything particularly special among humans. And it's not just critical areas; every cell in your body has enzymes that prevent mutation, both before and after a given replication.

[RaftPeople]

> And it's not just critical areas

True, but the resistance to mutations is increased in critical areas compared to other areas. Not all changes are equally likely.

[bubblyworld]

Yeah, that's interesting, I guess it's also a simplification to think of an organism as a single genome. The reality is much more complicated! (lichens come to mind as examples of even more genetic diversity housed in a single organism, or even gut bacteria in humans maybe?)

[logicchains]

No matter how low the probability of life arising by chance, from the perspective of life the probability that it happened is 100%, because if it didn't happen then we wouldn't be around to observe it. We're operating from a massive selection bias.

[mistermann]

The probability of it happening by chance alone is not known though, in the technical sense of the word. In the colloquial sense it is though, which makes for a very interesting situation here in 2024, and prior.

[evv555]

>The encoder being evolution is an idea that has been developed by Sui Huang and numerous others.

Parallels between evolutionary systems and hill-climbing algorithms have been floating around for a long time at this point.

[jhbadger]

Although normally in the other direction -- genetic algorithms and genetic programming directly mimic evolution by natural selection, for example.

[evv555]

It was just the less controversial direction. You would be accused of anthropomorphizing and teleological thinking if you suggested evolution is a search process towards some optimum. Good luck talking to an orthodox biologist about evolutionary processes with constructs like agents and intelligence.

[bjornsing]

> why can’t we put this to use producing alife simulations that deliver open-ended evolution?

Has it been tried? I’ve said for ages: show me a representation where a random bit flip generally results in a different but viable entity, and I’ll show you artificial life. The latent space of a VAE could well have those properties.

But it’s not open-ended though (in its obvious form) since the VAE would have to be trained on various complex life forms and will probably not extrapolate well outside the support of the training distribution.

[RaftPeople]

> show me a representation where a random bit flip generally results in a different but viable entity

It's a really interesting problem I pondered quite a bit when doing some a-life hobby stuff.

I never came up with a good solution, but you can kind of "feel" that the solution needs to be more analog-ish in the way info is represented. As you say, a small change in data (bit flip) probably needs to produce a small change in the resulting form. Possibly the binary representation points to a vector space of form "primitives" (drivers of form) such that adjacent points have similar form.

[svnt]

> But it’s not open-ended though (in its obvious form) since the VAE would have to be trained on various complex life forms and will probably not extrapolate well outside the support of the training distribution.

That is always the issue in alife: we discover processes that help us explore bounded information spaces, but only that.

[exe34]

> why can’t we put this to use producing alife simulations that deliver open-ended evolution?

you might like this: https://www.gregegan.net/DIASPORA/01/Orphanogenesis.html

[theGnuMe]

Do you have any references?

[svnt]

Here is a readable article by Watson that is probably among the more relevant:

https://www.richardawatson.com/songs-of-life-and-mind

There are many, and the paper itself cites many good sources including by Huang and Watson in the references section.