Scientists have put a simulated worm’s brain into a Lego robot’s body(sciencealert.com)
sciencealert.com
Scientists have put a simulated worm’s brain into a Lego robot’s body
http://www.sciencealert.com/watch-scientists-have-put-a-worm-s-brain-into-a-lego-robot-s-body-and-it-works
12 comments
> I had hoped this story was about a physical worm's brain being transplanted into a robot.
isn't that so much less interesting with respect to technology, biology, and philosophy? isn't putting a biological brain into a robot just an exercise in mapping inputs and outputs? we have already done this with human prosthetics.
how the brain communicates with other parts of your body is not much of a mystery.
synthesizing a brain in software seems so much more profound. not only is it a tool for helping understand the brain itself, but more importantly, consciousness.
isn't that so much less interesting with respect to technology, biology, and philosophy? isn't putting a biological brain into a robot just an exercise in mapping inputs and outputs? we have already done this with human prosthetics.
how the brain communicates with other parts of your body is not much of a mystery.
synthesizing a brain in software seems so much more profound. not only is it a tool for helping understand the brain itself, but more importantly, consciousness.
"... cultured rat neurons are placed on an array of electrodes and trained to power a robot..."
Wait what. That's sounds amazing. Are they still working on this?
Wait what. That's sounds amazing. Are they still working on this?
Kind of misleading when the video goes "this is not a program driving the robot" - of course it's a program. Just because they copied the interconnection structure of a worm brain doesn't make it not a program. It's simply very hard to reason about because it's structure and mode of computation isn't readily apparent, so we can think of it as the world's largest perl one-liner.
The text can be given the more charitable interpretation that they mean "program" in the sense of "human created instructions." This is how the word is generally used.
As this type of brain modeling becomes more widespread, we may have to make our language more precise to distinguish "programs" copied from brains vs "programs" written by brains!
As this type of brain modeling becomes more widespread, we may have to make our language more precise to distinguish "programs" copied from brains vs "programs" written by brains!
I wasn't misled one bit. Why be so pedantic? Sure, you can call it a program, but in that sense you're also a program.
Modeled a worms brain
Yes, the title is misleading. The software runs a simulation of the worm's brain using a software defined mapping between the 302 neurons (which I am assuming are software modeled) of C. elegans brain. The (software and hardware) reacts in a way similar to that which has been observed of C. elegans (the worm).
This brings up a question which I have often wondered - whether our consciousness is just a combinatorial explosion of the echoes of the external world - that our abilities and thoughts are predetermined by our environments. If I believed that, then perhaps I could model the human brain and say it was my own. But there is also the way humans choose to chop up the world, where the line is drawn between 'neuron' and 'neuron connection' and 'stimuli'. So, shrug.
This brings up a question which I have often wondered - whether our consciousness is just a combinatorial explosion of the echoes of the external world - that our abilities and thoughts are predetermined by our environments. If I believed that, then perhaps I could model the human brain and say it was my own. But there is also the way humans choose to chop up the world, where the line is drawn between 'neuron' and 'neuron connection' and 'stimuli'. So, shrug.
We do have assumptions to chop up the world, but they only to help give us a place to start exploring (as we know, we find better assumptions that make old assumptions obsolete). I think the chopping up problem is one people run up into a lot with discrete spaces vs continuous spaces.
Exploring the permutation space of 302 elements is one thing (302!), but exploring the permutation space of 20 billion+ elements is another[0].
Though there are some interesting approaches I have seen with exploring such spaces like projecting discrete spaces onto higher dimensional spaces and then using mathematics more suited to continuous spaces for exploring regions of interest [1].
[0] http://www.neurology.org/content/64/12/2004
[1] http://jmlr.csail.mit.edu/proceedings/papers/v15/plis11a/pli...
Exploring the permutation space of 302 elements is one thing (302!), but exploring the permutation space of 20 billion+ elements is another[0].
Though there are some interesting approaches I have seen with exploring such spaces like projecting discrete spaces onto higher dimensional spaces and then using mathematics more suited to continuous spaces for exploring regions of interest [1].
[0] http://www.neurology.org/content/64/12/2004
[1] http://jmlr.csail.mit.edu/proceedings/papers/v15/plis11a/pli...
It likely is just a machine. But incredibly complicated. C elegans itself is quite complicated with 300 neurons. It responds to various chemical stimuli, each neuron has thousands of activated genes, it responds to food and nutrition sources, touch, etc. As humans we have over 100 billion neurons.
More interesting than seeing motor function - being that C Elgans has been shown to exhibit depressive behaviour [1], I wonder if they can observe or trigger depression in the model?
That would, to me, show something much more thought provoking.
[1] http://scholar.google.co.uk/scholar?q=c+elegans+depression&h...
That would, to me, show something much more thought provoking.
[1] http://scholar.google.co.uk/scholar?q=c+elegans+depression&h...
Next thing for mad computer science to do: take a simulated C. Elegans, give it the toolkit of Stuxnet and Regin, set it loose to chew data. First real internet worm?
The article subtitle claims C. elegans is a parasite. That is incorrect. C. elegans tends to live in places like apple orchard floors where they eat decaying fruit. They are quite harmless and have been instrumental in genetics research.
C. Elegans is related to many species of parasite. I used to use it to help understand strongylioides.
Imagine using this "worm simulation" as a real world "computer worm" to do your bidding!
Without having checked the code (it is open source I believe); this is a neural network right? Do they understand why it reverses when it bumps into the wall from the activity observed in the neural net?
Here's a short video of the worm and how it moves https://www.youtube.com/watch?v=GgZHziFWR7M
Amazing stuff, the start of inorganic brain emulation/uploading? Maybe programmers could devise a program format for neuronal programming/uploading, which could be run on libraries for robotic platforms. Now we can truly code the "ghost in the machine".
Going a step further, what if we caused neuronal programs to evolve according to some fitness function, and gave them new kinds of sensors (senses?) and environmental challenges? We could end up with a complex brain, and some very curious emergent properties indeed...it makes me excited and worried all at once.
Digital brain + objective/incentive functions + evolutionary mechanics ==> hybrid natural-artificial life?
Quote from the creator about the roboworm's behaviour upon encountering an obstacle:
Obviously, the robot is a very simplistic form of the worm and there are many sensory inputs that aren't being taken into consideration. The worm's "nose" has sensory neurons for chemo, oxygen, touch, temperature, mechano, etc. I'm only using the "avoidance touch" but we have no idea how these other sensory functions might influence the behavior overall. Like any animal, the worm won't back up each and every time it comes to an obstacle; i.e. it might kind of "sniff" around it for a bit before moving on so we are taking a very narrow approach mostly due to the limitations of robotic sensors. However, I have probably ran this experiment over a hundred times and each time, the results are very consistent. Sometimes. I'll let the robot just wander around for a long period of time and it's like we have a cat checking out the environment.
On memory/plasticity of the simulated brain:
Actually you are the first to ask and I'm surprised no one has asked before. We actually have a mechanism we call the "poison pill" whereby we send the individual neurons of the simulated connectome a weighted value of -99999 and when the neuron reads this weight, it kills itself. So to answer your question, this current research is based purely on the connectome as the means to operate the robot and there is no plasticity other than the recursive nature of the connectome itself. No memory storage and each time you start the connectome, it starts new but we have probably did this experiment a 100 times by now and have shown that the connectome itself gives the wormbot it's ability to maneuver it's environment. Thanks for the question.
On human control of such robots:
One of the things that we want to explore is adding a Deep Learning app to the connectome to be able to guide the robot so we can make some more practical use for the system. The current research gives the robot the ability to do it's own thing like any animal. As I have already stated, it's like a cat whereby you can try to command a cat all day but it will do what it wants to do and the way we get cats to do what we want them to do is to change the environment; e.g. put out food. For example, if we were to create a bot specifically for search and rescue, we might want to give the bot some incentive to go a specific direction and if life is found, we may need it to perform specific tasks to ascertain the situation. My connectome gives the bot the ability to maneuver around in an environment and keep it safe from harm but it would do, like the worm, it's own thing and not pay much attention to what we wanted it to do. By interjecting some control (ala a DNN?), this might give us the best of both worlds: a bot that can fend for itself as well as some control that gives us a means to obtain certain information. So your idea is very much in line with some of our thoughts as well. The only difference would be that you must realize, this is truly an autonomous system and as it is right now, you can't control it any better than you could control a real worm or your cat.
Going a step further, what if we caused neuronal programs to evolve according to some fitness function, and gave them new kinds of sensors (senses?) and environmental challenges? We could end up with a complex brain, and some very curious emergent properties indeed...it makes me excited and worried all at once.
Digital brain + objective/incentive functions + evolutionary mechanics ==> hybrid natural-artificial life?
Quote from the creator about the roboworm's behaviour upon encountering an obstacle:
Obviously, the robot is a very simplistic form of the worm and there are many sensory inputs that aren't being taken into consideration. The worm's "nose" has sensory neurons for chemo, oxygen, touch, temperature, mechano, etc. I'm only using the "avoidance touch" but we have no idea how these other sensory functions might influence the behavior overall. Like any animal, the worm won't back up each and every time it comes to an obstacle; i.e. it might kind of "sniff" around it for a bit before moving on so we are taking a very narrow approach mostly due to the limitations of robotic sensors. However, I have probably ran this experiment over a hundred times and each time, the results are very consistent. Sometimes. I'll let the robot just wander around for a long period of time and it's like we have a cat checking out the environment.
On memory/plasticity of the simulated brain:
Actually you are the first to ask and I'm surprised no one has asked before. We actually have a mechanism we call the "poison pill" whereby we send the individual neurons of the simulated connectome a weighted value of -99999 and when the neuron reads this weight, it kills itself. So to answer your question, this current research is based purely on the connectome as the means to operate the robot and there is no plasticity other than the recursive nature of the connectome itself. No memory storage and each time you start the connectome, it starts new but we have probably did this experiment a 100 times by now and have shown that the connectome itself gives the wormbot it's ability to maneuver it's environment. Thanks for the question.
On human control of such robots:
One of the things that we want to explore is adding a Deep Learning app to the connectome to be able to guide the robot so we can make some more practical use for the system. The current research gives the robot the ability to do it's own thing like any animal. As I have already stated, it's like a cat whereby you can try to command a cat all day but it will do what it wants to do and the way we get cats to do what we want them to do is to change the environment; e.g. put out food. For example, if we were to create a bot specifically for search and rescue, we might want to give the bot some incentive to go a specific direction and if life is found, we may need it to perform specific tasks to ascertain the situation. My connectome gives the bot the ability to maneuver around in an environment and keep it safe from harm but it would do, like the worm, it's own thing and not pay much attention to what we wanted it to do. By interjecting some control (ala a DNN?), this might give us the best of both worlds: a bot that can fend for itself as well as some control that gives us a means to obtain certain information. So your idea is very much in line with some of our thoughts as well. The only difference would be that you must realize, this is truly an autonomous system and as it is right now, you can't control it any better than you could control a real worm or your cat.
-- How the robo-worm finds its food (it also responds to touch):
Actually we used Sound sensors to activate food (chemo) sensory neurons - you can hear me whistling in the video. Once the sound gets above a certain decibel threshold, the food sensory neurons are sent a weighted value. We also used touch sensors to simulate anterior and posterior touch. After I did a specific experiment, I used to sometimes let the robot wander around my house. My wife would call the robot - activating food sensory neurons - and the robot would start to move towards her, I'm not sure who was training who at that point :-) It was very interesting to have the robot just wander for a number of minutes or an hour+. It was like a real creature wandering around interacting with it's environment. Most like a cat than any other animal.
-- The researcher also wrote about plasticity and the challenges to developing connectomes (virtual brains) further, e.g. to model higher organisms:
Scaling my efforts up is a huge magnitude of effort - to the human level is beyond my vision at this time. I am looking to scale up only to the next higher order of animals and would be very happy to get to the mouse brain in my lifetime.
A couple of things that make scaling up a difficult issue:
1) C Elegans does not grow any new neurons or synapses over it's lifetime so it is a static model as far as the connectome. Plasticity is a whole new paradigm that my current model does not consider. I do have ideas for growing new synapses that I actually worked on around 8 years ago and with this research, should allow me to move forward.
2) It is hard for current computing power to handle the C Elegans nervous system with 302 neurons. The next level would be the Jelly Fish with 800 neurons. Drosophila has 100,000 neurons, and other higher order animals would be a hard problem indeed.
3) My model works on a fully mapped connectome. C Elegans is probably the only animal that has this attribute. The Allen Brain Institute is doing detailed mapping on the mouse brain so this might be a good direction. The Human Brain project is doing the human brain but they would be first to tell you it's a multi-decade endeavor.
However, there are so many benefits to this research beyond what I have accomplished. The recursiveness of the connectome that I have found is very significant; which by the way compounds the complexity of brain science 10 fold, I am also exploring the cross connections of neuron pairs to try and find patterns and distinctions. Like all good science, we take a step forward and find many new things to explore.
The OpenWorm[1] project is doing exactly what you are suggesting - trying to map the connectome to a virtual body. I encourage you to look into what they are doing in more detail.
...However, my robot-connectome is fully autonomous and this is something we rarely see in science where we give a machine a level of intelligence we cannot control so in this regard, I firmly believe that if the singularity is met, it could be caused by a guy in his garage and the "old" humanity could be doomed forever more. But this is pure speculation and we won't know until we get there but it is something to ponder and whether it's a bad thing or good, I guess depends on your point of view.
Thanks for your kind words and please continue your thoughts, Tim
[1] http://www.openworm.org/ - creating a virtual C. elegans nematode in a computer.
Actually we used Sound sensors to activate food (chemo) sensory neurons - you can hear me whistling in the video. Once the sound gets above a certain decibel threshold, the food sensory neurons are sent a weighted value. We also used touch sensors to simulate anterior and posterior touch. After I did a specific experiment, I used to sometimes let the robot wander around my house. My wife would call the robot - activating food sensory neurons - and the robot would start to move towards her, I'm not sure who was training who at that point :-) It was very interesting to have the robot just wander for a number of minutes or an hour+. It was like a real creature wandering around interacting with it's environment. Most like a cat than any other animal.
-- The researcher also wrote about plasticity and the challenges to developing connectomes (virtual brains) further, e.g. to model higher organisms:
Scaling my efforts up is a huge magnitude of effort - to the human level is beyond my vision at this time. I am looking to scale up only to the next higher order of animals and would be very happy to get to the mouse brain in my lifetime.
A couple of things that make scaling up a difficult issue:
1) C Elegans does not grow any new neurons or synapses over it's lifetime so it is a static model as far as the connectome. Plasticity is a whole new paradigm that my current model does not consider. I do have ideas for growing new synapses that I actually worked on around 8 years ago and with this research, should allow me to move forward.
2) It is hard for current computing power to handle the C Elegans nervous system with 302 neurons. The next level would be the Jelly Fish with 800 neurons. Drosophila has 100,000 neurons, and other higher order animals would be a hard problem indeed.
3) My model works on a fully mapped connectome. C Elegans is probably the only animal that has this attribute. The Allen Brain Institute is doing detailed mapping on the mouse brain so this might be a good direction. The Human Brain project is doing the human brain but they would be first to tell you it's a multi-decade endeavor.
However, there are so many benefits to this research beyond what I have accomplished. The recursiveness of the connectome that I have found is very significant; which by the way compounds the complexity of brain science 10 fold, I am also exploring the cross connections of neuron pairs to try and find patterns and distinctions. Like all good science, we take a step forward and find many new things to explore.
The OpenWorm[1] project is doing exactly what you are suggesting - trying to map the connectome to a virtual body. I encourage you to look into what they are doing in more detail.
...However, my robot-connectome is fully autonomous and this is something we rarely see in science where we give a machine a level of intelligence we cannot control so in this regard, I firmly believe that if the singularity is met, it could be caused by a guy in his garage and the "old" humanity could be doomed forever more. But this is pure speculation and we won't know until we get there but it is something to ponder and whether it's a bad thing or good, I guess depends on your point of view.
Thanks for your kind words and please continue your thoughts, Tim
[1] http://www.openworm.org/ - creating a virtual C. elegans nematode in a computer.
While the title itself is a little misleading, i find the experiment incredibly inspiring.
Your scientists were so preoccupied with whether or not they could that they didn't stop to think if they should.
Here is a link to some info on Hybrots:
http://en.wikipedia.org/wiki/Hybrot