That last pattern is so wild, it almost looks like a microscopic fluid simulation. I bet you could use it in optimizing circuit board layouts or something.
3 things: 1) your work is awesome, keep it up! 2) it's fascinating how these patterns are visually so similar to those produced by reaction-diffusion systems (e.g. how big cats get their stripes/spots). I have to believe there's a connection there. Convolutions, Fourier Transform, and differential equations are all tightly linked, so this isn't too surprising. 3) any insight into what's up with the dithering / checkerboard pattern? It seems crucial for stability in the worms world parameter-space. It also seems like it facilitates the different color density "species". Would love to see a cross section of what's going on.
These videos and this knowledge and this path of study needs to be shared!! I don't think most people would understand how important this is but I can see so much potential here! This IS how life works.. this IS life and nature itself..
Thanks for demystifying neural cellular automata for me. It's really quite amazing and reminds me of reaction-diffusion more than conway's game of life.
Im wondering if you had to do some kind of sahder programming, or if one can program something just as fast with just by populating a new array for all the pixels every frame
@@EmergentGarden Does this mean that with wider range of cell values and more complicated activation functions there might be a way to implement at least something that vaguely resembles a simple natural selection?
This is usually due to the skip frame display feature, rather than genetic diversity. Well, depending on your definition of genetic diversity. All cells pulse between a bright state and a darker state and depending on the phase you'll see only one of those when the skip frame function is enabled.
@@galator443 with enough floating point precision (would need a modification of the engine, I think) and a complex enough activation function you can perform arbitrary calculations so I don't see why not
Fascinating. I'd be quite excited to see a life simulation based on these complex NCA behaviors. There would have to be quite a few parameters to be able to inherit/mutate I'm sure.
@@EJM505 Neural cellular automata, or complex algorithms that dictate what the pixels surrounding it will look like. The NCAs highlighted in this video are examples that resemble organic life and some of it’s cellular function (like mitosis, around the 1:20 mark)
the last slime mold example (6:10) reminds me of ant colony simulations where they start off with an unoptimized path from hive to food and optimise it iteration by iteration. Really fascinating stuff!!
If you can implement some sort of pattern with a kind of conservation of energy with a little bit of free energy income, that could lead to some incredibly realistic patterns.
The takeaway from this should be the confidence that organic shapes and complex behaviors arise easily in situations where the elements of a large system are influenced by their local environment and not more distant information. To say that these shapes are 'surprising' to see coming out of these cellular automata is to say that the a-priori expectation was that organic shapes and complex behavior relies upon some complicated design. If conventional wisdom incorporated the understanding that these shapes and behaviors are inevitable in dynamic systems, we would be better off since people would not find it necessary to look for supernatural explanations for life and such.
I wonder if you could use an edge detection algorithm to find if a new "organism" was born and then have a different filter for each one. Each organism could have a slightly different filter than its parent and thus natural selection could ensue.
Perhaps you do this: Each organism lives in their own grid that evolves according to a rule. After each step, you run something to detect if there are multiple blobs in a given grid. Perhaps you could try to see if the function implied by the values in that grid has multiple peaks and apply a threshold to define each blob. Than you would separate the blobs each in their own grid and slightly change the update rule (filter + activation function).
Perhaps interaction between organisms by using filters that run in other grids and contribute to one's own next values... I don't know how you would introduce competition though. It seems that the only pressure here would be to form the biggest number of "blobs" possible. Perhaps you could limit the total sum on each site of the grid (like a spatial limitation) or introduce some energy expenditure that affects the dynamics.
I would be interested to see contiugious masses storing the genetic code for the algorithm, and when they expand there is that fractional chance of a mutation in that small part of the mass. What colour differentiation threshold defines a contiguous mass would likely have to be a parameter of the system, and any mutation would likely have to come with a colour change in order to seperate it from the original, but. Otherwise it would be a relatively clean and elegant way to implement a kind of natural selection, I think.
please add some food/resources on the way these patterns grow, to add complexity. in general it would be 2 patterns : one for organism (based on its genetics) and food map/pattern (based on how food distributed around) I think main advantage of this overlay would be the conditions you can't program into organisms so this overlay should expand maximal complexity of this system.
Be cool if we could somehow have multiple different ones occupy the same simulated space to see what would happen. Maybe different colours could cause blank spaces to act differently.
What about creating a slime mold colony and an anti-slimemold colony that explode when they come into contact, destroying a large portion of the area around the explosion, but also leaving trace amounts of energy that interacts with both colonies in various energy like ways? Sort of like the matter/antimatter collisions if the early universe.
At about 2:10 to the left side of the collective blob. How does those small "mutations" arrise, as you said, there is no genetic information passede on, but it sure seem like the small ones produce small ones & the big ones big ones. So there must be something to their pattern that acts as a semi information. Or is it a YT compression error I read to much into?
I feel like this would be a different kind of AI than what most would imagine. Instead of one superintelligent AI overlord, it would be an innumerable group of singular thoughtless AI, using evolution processes to colonize the internet. It would be a conquest of plague
Suppose that these patterns could be used as patterns of growth or expansion for self-organizing programs, robots, or AI creatures in a simulation. In theory you could run the algorithm for a few steps and compare the resulting pattern with the current pattern and wether there's sufficient resources to grow into that new shape. You could also have different layers of these patterns to help decide different aspects of the growth. Would interest me greatly to see more AI used in videogames.
if the computer generated automata are the relativly the same looking as the eukeryotic bacteria then whats saying that when the bacteria go through mitosis they dont just use the relativly same algorithim for when they decide to multiply/grow in certain areas?
I'd really like to try this out but was unable to get anything but the main menu and a black screen. I tried loading some of the examples but nothing displayed. Do I need to enter some data into the activation window? Pointing and clicking the left mouse button had no affect either - thanks.
I've watched so many videos on countless different cellular automata and still every time , I'm just , overwhelmed by the complexity that emerges
That last pattern is so wild, it almost looks like a microscopic fluid simulation. I bet you could use it in optimizing circuit board layouts or something.
3 things: 1) your work is awesome, keep it up!
2) it's fascinating how these patterns are visually so similar to those produced by reaction-diffusion systems (e.g. how big cats get their stripes/spots). I have to believe there's a connection there. Convolutions, Fourier Transform, and differential equations are all tightly linked, so this isn't too surprising.
3) any insight into what's up with the dithering / checkerboard pattern? It seems crucial for stability in the worms world parameter-space. It also seems like it facilitates the different color density "species". Would love to see a cross section of what's going on.
These videos and this knowledge and this path of study needs to be shared!! I don't think most people would understand how important this is but I can see so much potential here! This IS how life works.. this IS life and nature itself..
Thanks for demystifying neural cellular automata for me. It's really quite amazing and reminds me of reaction-diffusion more than conway's game of life.
Im wondering if you had to do some kind of sahder programming, or if one can program something just as fast with just by populating a new array for all the pixels every frame
There does seem to be genetic diversity for the green mitosis cells, since some of them are dark and others are light.
Ha I knew I should've mentioned that! You're right, though it doesn't affect their rate of replication I believe
@@EmergentGarden Does this mean that with wider range of cell values and more complicated activation functions there might be a way to implement at least something that vaguely resembles a simple natural selection?
This is usually due to the skip frame display feature, rather than genetic diversity. Well, depending on your definition of genetic diversity. All cells pulse between a bright state and a darker state and depending on the phase you'll see only one of those when the skip frame function is enabled.
@@galator443 with enough floating point precision (would need a modification of the engine, I think) and a complex enough activation function you can perform arbitrary calculations so I don't see why not
the visuals and the music make this some of the most beautiful bits of generative art I’ve ever seen
I absolutely love this. Check out this pattern I found: all weights set to 1, activation function -(x-1.98)*(x-4.02)
This is great! Your content is really high quality, i sincerely hope your channel grows!
Fascinating. I'd be quite excited to see a life simulation based on these complex NCA behaviors. There would have to be quite a few parameters to be able to inherit/mutate I'm sure.
What does the acronym 'NCA' stand for?
@@EJM505 Neural cellular automata, or complex algorithms that dictate what the pixels surrounding it will look like. The NCAs highlighted in this video are examples that resemble organic life and some of it’s cellular function (like mitosis, around the 1:20 mark)
I tested this thing, very cool!) I changed the parameters many times and it turned out very interesting things sometimes.
the last slime mold example (6:10) reminds me of ant colony simulations where they start off with an unoptimized path from hive to food and optimise it iteration by iteration.
Really fascinating stuff!!
would be so interesting to actually see evolution in NCAs, it would be just next level fascinating
If you can implement some sort of pattern with a kind of conservation of energy with a little bit of free energy income, that could lead to some incredibly realistic patterns.
DO IT!! PLEASE PURSUE SELECTION BASED ON NCAs!!!
PLEASE PLEASE PLEASE.. I would if I could but can't.. I don't have the knowledge or resources.. please do this please.. I need to see it!
I don't know how to contact you directly.. I hope my comments reach you.. this area of research is so important
These are so cool! You should definitely simulate a natural selection process for these patterns
I would watch a hour compilation of these neural patterns. Would be great to have as a visual stimulus on my desktop as I was working in my office.
Everything truly is MATH. Wow the patterns are like real life. That is insane.
I found this as fun to watch as to watch real microscopic footages of organisms , but why?
this is just beautiful. and thank you, mr. emergent garden.
The thing at the end looks like roadways yes, but actually looks even more like a circuit board
Beautiful descriptions and images. You've truly expanded my mind, thank you.
The takeaway from this should be the confidence that organic shapes and complex behaviors arise easily in situations where the elements of a large system are influenced by their local environment and not more distant information.
To say that these shapes are 'surprising' to see coming out of these cellular automata is to say that the a-priori expectation was that organic shapes and complex behavior relies upon some complicated design.
If conventional wisdom incorporated the understanding that these shapes and behaviors are inevitable in dynamic systems, we would be better off since people would not find it necessary to look for supernatural explanations for life and such.
1 step closer to making a true a.i. that thinks and feels like organic beings do
I wonder if you could use an edge detection algorithm to find if a new "organism" was born and then have a different filter for each one. Each organism could have a slightly different filter than its parent and thus natural selection could ensue.
Perhaps you do this:
Each organism lives in their own grid that evolves according to a rule. After each step, you run something to detect if there are multiple blobs in a given grid. Perhaps you could try to see if the function implied by the values in that grid has multiple peaks and apply a threshold to define each blob.
Than you would separate the blobs each in their own grid and slightly change the update rule (filter + activation function).
Perhaps interaction between organisms by using filters that run in other grids and contribute to one's own next values...
I don't know how you would introduce competition though. It seems that the only pressure here would be to form the biggest number of "blobs" possible.
Perhaps you could limit the total sum on each site of the grid (like a spatial limitation) or introduce some energy expenditure that affects the dynamics.
I would be interested to see contiugious masses storing the genetic code for the algorithm, and when they expand there is that fractional chance of a mutation in that small part of the mass.
What colour differentiation threshold defines a contiguous mass would likely have to be a parameter of the system, and any mutation would likely have to come with a colour change in order to seperate it from the original, but. Otherwise it would be a relatively clean and elegant way to implement a kind of natural selection, I think.
Bluhuhuh reminded me of ecoli when you first introduced the white and pink mitosis
Love the video please make more.
please add some food/resources on the way these patterns grow, to add complexity.
in general it would be 2 patterns : one for organism (based on its genetics) and food map/pattern (based on how food distributed around)
I think main advantage of this overlay would be the conditions you can't program into organisms so this overlay should expand maximal complexity of this system.
out of this world
The last ones looked as the patterns of the plane of a city streets, it could be usefull for some software that generates cities for videogames.
6:10 stunning!!!
Be cool if we could somehow have multiple different ones occupy the same simulated space to see what would happen. Maybe different colours could cause blank spaces to act differently.
What about creating a slime mold colony and an anti-slimemold colony that explode when they come into contact, destroying a large portion of the area around the explosion, but also leaving trace amounts of energy that interacts with both colonies in various energy like ways?
Sort of like the matter/antimatter collisions if the early universe.
This is the coolest shit I've seen in ages
At about 2:10 to the left side of the collective blob. How does those small "mutations" arrise, as you said, there is no genetic information passede on, but it sure seem like the small ones produce small ones & the big ones big ones.
So there must be something to their pattern that acts as a semi information.
Or is it a YT compression error I read to much into?
I feel like this is the first step towards an AI takeover
I feel like this would be a different kind of AI than what most would imagine. Instead of one superintelligent AI overlord, it would be an innumerable group of singular thoughtless AI, using evolution processes to colonize the internet. It would be a conquest of plague
@@micahconnor8954 yeah like someone does an evolution simulator with a supercomputer but then they realize they’re in a simulations and break out
I love this video so much
now you need to write one which acts as a predator, eating the slime mould
What is the underlying logic behind ? Some kind of modified Grey scott reaction ?
Incredible job
So cool thankyou for sharing!
art
6:50 looks like a circuit board made out of liquid
Man, u really gotta post the filter and activation function. Id like to play around with it
Great work!
Suppose that these patterns could be used as patterns of growth or expansion for self-organizing programs, robots, or AI creatures in a simulation. In theory you could run the algorithm for a few steps and compare the resulting pattern with the current pattern and wether there's sufficient resources to grow into that new shape. You could also have different layers of these patterns to help decide different aspects of the growth. Would interest me greatly to see more AI used in videogames.
if the computer generated automata are the relativly the same looking as the eukeryotic bacteria then whats saying that when the bacteria go through mitosis they dont just use the relativly same algorithim for when they decide to multiply/grow in certain areas?
YAY NEW VID
lovely
combine neural cellular automata with evolution
fun fact: no one has watched the full video yet because its only 5 minutes old
edit: someone probably has now
Hello, fellow fan.
What if the parameters were not the 8 adjacent cells but instead the 16 cells in the ring around the 3x3 grid. And 8 adjacent cells do not matter.
Love your vids
Hope you've heard about Lenia
Local vibrational networks?
What is the filter and activation for all of these
What is this coded in? I'm quite experienced with neural networks in python and I'd love to have a go at something like this.
u cud prolly sell these full screen images as nfts lol
first song used in video?
That almost sounds like stem cells.
I'd really like to try this out but was unable to get anything but the main menu and a black screen. I tried loading some of the examples but nothing displayed. Do I need to enter some data into the activation window? Pointing and clicking the left mouse button had no affect either - thanks.
오 :)
2:22 natural selection IS happening here; it's just that it's not "genetic" (although it is also code, lol),
natural selection is a benign phenomenon. it is something which happens all around us every day. it is a product of the passage of time
So is this your project?
Your discord link is broken
ok
fucking badass
hell yeah brother