back
12 comments
Dinoflaggelates have been observed to steal plastids from other marine algae. Not only this, but some species of dinoflaggelates are 'picky' in the algae they target for kleptoplasty. To perform this action, the single celled dinoflaggelate identifies a suitable algae in the environment; it must calculate and decide if the identified candidate algae represents an improvement over its current photosynthetic captive; it must then calculate a cost benefit of rejecting (or digesting) its current captive and retaining a new captive, and the expected rate of return of the partner.

Clearly, this single celled, is performing computation. 'Knowledge' of self, environment and the relationship between the two worlds has to be encoded in an active and dynamic fashion. While networks and networks of cells are an incredible and fascinating system, clearly inspirational and massively insightful for the development of intelligence outside of biological systems, one could and should consider the argument that by jumping to networks as systems for computational intelligence, we skipped a step in describing how cells (and thereby networks of cells) encode computation and intelligence. I think that a major step in theories of intelligence and mind could be possible by reconsidering intelligence at a cellular and single cell level.

https://en.wikipedia.org/wiki/Dinoflagellate https://en.wikipedia.org/wiki/Kleptoplasty https://aem.asm.org/content/78/3/813

You would be very interested in the work of independent research biologist Brian J Ford (http://www.brianjford.com/bjford.htm). Over the years he has championed the idea of the intelligent cell, and the importance of studying the entire system of the living cell.

Check out his latest publications here: http://brianjford.com/wbbjf10.htm

Thats great, I'll check that out. I took a few graduate courses in systems biology and related machine learning a long time ago, but I lost interest because it was so `human` biology focused, and that area of biology very-much does not appeal to me. There are some datasets prepared in such away as to address some of the issues in cellular intelligence, but very few out side of the realm of human biology.

I think the most interesting components to me are histology and transtrictomics as they relate to cellular behavior. The fact that cells have to rely on membranes and voltage potentials, but through compartmentalization are able to create a dynamic computational environment, is the really mind-blowing bit.

Its always funny to me when people want to reduce the behavior of non-human lifeforms to something akin to a program written on a tape drive (see below), when we know that DNA exists in a dynamical 3-dimensional form when in the somatic phase.

"'Knowledge' of self, environment and the relationship between the two worlds has to be encoded in an active and dynamic fashion." I'd say that's overreaching, but those are incredibly ill-defined terms.

In the case of the Dinoflaggelates, such a action could be performed by an incredibly simple state machine, making the definitions of the above terms incredibly weak.

ETA: I'll add that you put a lot of calculating and work on the Dinoflaggelate that is likely performed by evolution. The "calculations" would just be done by the genetic equivalent of a look up table.

DNA in its somatic state is a a 3-dimensional dynamical system. Not only does the 3-dimensional structure of DNA change with time, but in some cases, the very encoding of the DNA itself. Can you produce a simple state machine than can replicate that kind of behavior. You are aware that how DNA, transcription, translation, and cellular behavior work are still areas where we only understand the barest mechanisms (although we do know quite a lot), in an incredibly few number of systems where we know these mechanisms to exist, yes?

Based on my reading of your comment it seems like you would consider cellular behavior to be a kind of linear program, which really only shows a fundamental lack of understanding of the nature of the system and biology.

Buzz words do not make magic happen. I can describe the transistors that make up a state machine in very fancy language, even invoking quantum, to try to capture the full functioning of the atomic and quantum level of what's going on is enormously complex. It's even 3-dimensional.

However, if the part I care about is only the implementation of the state machine, I can transfer that to any other substrate or just pure logic.

If I'm looking at the "cognition" necessary to swap out the current algae for a new one, I don't need to care about what makes the state machine in the cell function, whether it's the relative concentration of two different proteins, the pH of the cell, or whatever. If the intelligence part of it can be reduced to a simple state machine, the other stuff doesn't factor into it.

You maybe right. It took about a 100 years of thousands of weavers playing with punch cards, trying to increase loom efficiency, before Joseph Marie Jacquard perfected it. Until that happened, inspiring Charles Babbage, mathematicians had not produced any worthwhile computational machines. We have probably the same problem in biology. We need thousands of people playing and producing work like this before we see major mastery at the cellular level.
When discussing abstract computations the way I interpret it is that the level of biology doesn't matter. A single cell has countless moving parts, which can be interpreted as a network of discrete computations.

The biology matters for biologists, and because perhaps it can give us insight into more abstract lessons in computation. I don't think one level of biology is more informative than the other per se.

So scanning this seems to be a summary of what's know about biological neural networks, from low to high level. It's very complex with even the lowest levels, single neurons, requiring complex models.

Further, as I understand, for each and every level here, you could find an alternate school of thought that would offer a slightly or a considerably different model.

Essentially, you reach the point where so many different human ideas have to compete in understand phenomena X in the brain that integrating them goes the mental capacity of a given human being - taking into account that every model here is going to be a very leaky abstraction.

So we basically need a computer program or interface, not even to really simulate the brain but to integrate existing models on whatever level of abstraction we're working on.

One would want something that lets one shift seemlessly between models. Essentially, a system that lets you take the information that is now contained in scientific papers and make it as interactive as a spreadsheet. Anyway, we're quite far from such a situation.

Having spent ~6 years studying nonlinear laser dynamics, it always amuses me to see bifurcation theory pop up on HN.

https://neuronaldynamics.epfl.ch/online/Ch4.S4.html

A fantastic resource on a lot of this material (basically a peer reviewed wikipedia): http://www.scholarpedia.org/article/Bifurcation

...only 1 chapter on learning ?!

It shows how little we know of how biological neural networks actually work (imo 95% of "work" would be how learning occurs, that's the really interesting bit). I know this is hard to study, but wow, there's a huuuge gap here.

Thanks for sharing this!