Slime molds’ complex ability to detect mass from a distance
harvardmagazine.com
harvardmagazine.com
It's not clear to me if they are using 'mass' as we understand it from physics, or just the general notion of size of an object.
To me the stacking experiment seems to invalidate that its the former. It would probably be useful to do some analysis on airflow inside the petri dish, there could be a very subtle convection and/or radiant heating from the white disks. There's also potential phototropic effects like we would see with plants bending towards light. Both of these inputs would actually be affected by the stacking, where mass is clearly not. (Yes they 'turn off the lights', but you can't take a time lapse movie in absolute darkness)
Regardless of the cause, the effects are going to be extremely subtle, and the fact that the slime mold seems to select a direction based on this incredibly subtle effect is still extremely cool.
There is a fantastic book that talks about this kind of emergent intelligence in slime molds, called--ironically--"Emergence: The Connected Lives of Ants, Brains, Cities, and Software".
The first 1/4 of the book is about the freakiness of slime mold. I thoroughly enjoyed this book, very well written and accessible and covers multiple domains (a lot like G.E.B.); Link to non-amazon:
https://www.barnesandnoble.com/w/emergence-steven-johnson/11...
This is as detailed a description as the article gives in how they detect objects? This explanation does not seem very helpful since I assume the glass disks were not jumping up and down. Is there any more information available on how they sense?
I only glanced at it, but noticed that they hypothesized that the organism uses mechanosensation involving rhythmic pulling on the substrate and interpretation of changes in things like tension, compression or mechanical strain.
https://www.researchgate.net/publication/340090261_Mechanose...
Today is Sci-Hub anniversary the project is 10 years old!
I'm going to publish 2,337,229 new articles to celebrate the date. They will be available on the website in a few hours (how about the lawsuit in India you may ask: our lawyers say that restriction is expired already)
https://nitter.kavin.rocks/ringo_ring/status/143435621720862...
> Physarum is widely known to grow in a pulsatile manner, which consists of a forward growth phase and a reverse streaming phase during which the cytoplasm is retracted away from the growth regions (Figure 8A). We have observed that this oscillation is critical for Physarum mechanosensation and that interrupting it by changing the substrate stiffness (Figure 4) or interfering with the oscillations (Figure 6) prevents accurate decision making. Thus, we propose a theoretical model of Physarum navigation where this oscillatory behavior acts as a sample-and- integrate function (Figure 8B): the growth regions sample the environment during the growth phase, optimize the direction of the network tubes during the reverse streaming phase by inducing internal tension in the Physarum network which then aligns future growth of the growth regions.
So basically the slime mold pulls back on the substrate and feels for strain gradients.
Imagine if we knew how to train ML models that could reliably perform feats like this (learn after training, not train to withdraw every hour), while still managing all other parts of being a slime mold like hunting for food, distributing nutrients etc.
https://www.sciencedirect.com/science/article/abs/pii/S03032...
I imagined it and I don’t find it that special.
Picking up a strong and regular periodic signal and predicting the next iteration is not a hard problem.
> while still managing all other parts of being a slime mold like hunting for food, distributing nutrients etc.
ML models don’t do those things because ML models are abstract constructs without appendages or any other physical embodiment.
We cannot manufacture artifical slime molds because we cannot manufacture artifical cells. This doesn’t seems to be an ML problem. (Or if it is, then everything become an ML problem too and the term lost its meaning.)
I am talking about it doing this in a simulation. We don't have AI that are as smart as a slime mold at the moment, it will be an achievement when we get there. Also we can't discount the processing power required to process raw signals detected, our AI models needs to also learn the signal processing as well. The raw data a slime mold has to work with is extremely bare, if you get that input do you really think an AI could do the same things as it? And this is a single cell organism, this is the capability of a single biological cell.
> Picking up a strong and regular periodic signal and predicting the next iteration is not a hard problem.
Not if that is the only thing you do, no, but if you also do a lot of other things at the same time then it gets extremely hard. Edit: For example, how does the mold know that it just didn't move into a sunny spot? Or maybe this is just a cloud and it retracted unnecessarily? There are a lot of things to account for in a real world scenario that the slime mold has to understand that your naive signal processor can just ignore.
Why do you think that it does know that?
> Or maybe this is just a cloud and it retracted unnecessarily?
I bet that it does retract sometimes unnecessarily. Every organism has these kind of false positives.
> We don't have AI that are as smart as a slime mold at the moment, it will be an achievement when we get there.
When you say we don’t have this AI do you mean you can’t download it from Github right now, or do you mean it is something humanity couldn’t build if they tried? I assume you mean it in the second sense. Why do you think it is so hard? Let’s say someone, a research project or an alien inteligence gives you such an AI on a thumbstick. How would you evaluate it to test if it is indeed a simulation of a slime mold? In other words what do you expect from the simulation?
"Could build if they tried" is a strange description, we haven't made an AI with similar capabilities as a slime mold so we don't know how to do it. Maybe a research team could do it if they tried, and we would learn things from it, but before we have done it I'll say that we can't do it. Similarly 5 years ago I'd say that we don't know how to make a super human GO AI at the time. There was a team in the process of making a GO AI in the works, but we still didn't know how to build it since we didn't know if their tech would be good enough to do it at the time.
> I bet that it does retract sometimes unnecessarily. Every organism has these kind of false positives.
Sure, but it can't get into an infinite loop or hang, or it would die a lot. It needs to be smart enough to survive in most real world scenarios. That is the hard part. It can't just do something like "if I get cold two times in a row I'll retract and stop moving in preparation for the third time", that would probably make it stop way too much. And we do know that slime molds remembers locations of food, so it has a proper memory, likely it has some sort of reasoning about temperatures and locations as well to avoid going to its death. This just has evidence that it even reasons about time and temperature as well, and even detects recurring time series events in addition to everything else it does, that is a lot of features that needs to interact well.
Edit:
> How would you evaluate it to test if it is indeed a simulation of a slime mold?
Simple, you make another team set up some tests for it, and see if it behaves similar to a real world slime mold in those scenarios. Slime mold environments are very quick and cheap to set up and run, so it would be neither hard nor expensive to do.
https://www.scientificamerican.com/article/veggies-with-visi...
https://www.nationalgeographic.com/science/article/the-sneak... (for this there is the claim that it copies plastic plants, too. couldn't find anything factual though)
If Slime can detect a mass horizon gradient, maybe experiments need to be redesigned to account for it?
Why would a slime mold even evolve this capability of detecting tiny mass differences?
None of it makes sense. It's much more likely there's an ultra-thin tendril they haven't noticed.
Mold is saprophytic so it’s more often than not advantageous to grow toward larger stationary objects in its environment to eat.
The stack of discs deformed the agar gel, the mechanoreceptors on the periphery hyphae detected the tiny deformation and set off a chemical cascade for growth in that direction.
What if the substrate is lumpy and harder than agar? The researchers disabled the mechanoreceptors and there was no “mass detection”. If there’s no deformation in the substrate by discs on a hard bumpy surface then I’m guessing those mechanoreceptors won’t detect the mass.
So mold uses gradient deformation of mechanoreceptors to move toward possible food sources.
This explains why they compared the ability to a blind folded person on a trampoline. I wish they had explained it this way.
The mold may not even be moving toward a food source - it could be moving toward the deformation because larger objects cast bigger shadows and it needs to prevent desiccation.
A 100 times for each variant.
> How many trials did they have to do before they got this result?
Looks like it picked the side with 3 disks 70% of the time and never picked the side with 1 disk. They tested a ton of different scenarios so the results seems really robust.
"When presented with the 3-discs versus 1-disc choice, the Physarum grew toward the 3 disc regions 70% of the time, while never choosing the 1 disc region alone."
They weren't touching, tasting, smelling or hearing.