Compliant Mechanisms that LEARN! - Mechanical Neural Network Architected Materials
ฝัง
- เผยแพร่เมื่อ 9 ก.ค. 2023
- This video introduces the world’s first mechanical neural network that can learn its behavior. It consists of a lattice of compliant mechanisms that constitute an artificial intelligent (AI) architected material that gets better and better at acquiring desired behaviors and properties with increased exposure to unanticipated ambient loading conditions. It is a physical version of an artificial neural network used in current machine learning technologies.
To learn more about the content of this video, I encourage you to read the following publications, which can be accessed at the provided links:
[1] Lee, R.H., Mulder, E.A.B., Hopkins, J.B., 2022, “Mechanical Neural Networks: Architected Materials that Learn Behaviors,” Science Robotics, 7(71): pp. 1-9
www.science.org/stoken/author...
[2] Lee, R.H., Sainaghi, P., Hopkins, J.B., 2023, “Comparing Mechanical Neural-network Learning Algorithms,” Journal of Mechanical Design, 145(7): 071704 (7 pages)
asmedigitalcollection.asme.or...
Part files to fabricate the mechanical neural network can be downloaded on Thingiverse using this link:
www.thingiverse.com/thefactso...
Donate to help support my channel:
If you’d like to make a one-time donation, you can use the following link:
PayPal.me/FACTsMechDesign
If you’d like to sign up to be a monthly contributor, you can sign up on Patreon:
(link pending…. Thank you for your patience)
Thank you for your support! It is much appreciated and helps enable me to make more content.
Acknowledgements:
Special thanks to Ryan Lee, Erwin Mulder, and Pietro Sainaghi who helped fabricate, test, and simulate the mechanical neural network in the video. I am also grateful to my AFOSR program officer, “Les” Lee, who funded the research that this video features.
Brain Scan Attribution:
Christian R. Linder, CC BY-SA 3.0 creativecommons.org/licenses/b..., via Wikimedia Commons
commons.wikimedia.org/wiki/Fi...
upload.wikimedia.org/wikipedi...
Microstructure Image Attribution:
Edward Pleshakov, CC BY 3.0 creativecommons.org/licenses/..., via Wikimedia Commons
commons.wikimedia.org/wiki/Fi...
upload.wikimedia.org/wikipedi...
Body Armor Attribution:
commons.wikimedia.org/wiki/Fi...
upload.wikimedia.org/wikipedi...
Disclaimer:
Responsibility for the content of this video is my own. The University of California, Los Angeles is not involved with this channel nor does it endorse its content.
![[Full Episode] Hell's Kitchen Thailand EP.12 | 28 เม.ย. 67](http://i.ytimg.com/vi/xCjbu5t25Pw/mqdefault.jpg)
at 11 minutes I realised this was a research paper in an easily digestable and widely available format, great work.
It was so easily digestible that I went to shit right after finishing this video
Accessible cutting edge research is humanity at its best
15 mins I decide to read comments .
I noticed that at 18 minutes but I like me a good research paper
@@turolretar lmao man
As a mechanical engineer who is learning computer science and machine learning, this is an amazing bridge between the two worlds! I cant wait to print some and play with the concept myself. The applications are truly endless, I wonder how long until this is made microscopically, and applied everywhere.
This sounds like fiction, but makes total sense when you think about what else we’ve done recently with neural networks.
As a mechanical engineer you should recognize that all they are doing is recognizing the displacement of one node and then directing other nodes to form the final shape. Its like if you had a human like statue, rigged with motors for human like motion and programmed it so that it would want to return to a neutral stance but if someone slowly pushed it over it would transition to a different pose based on how far it was being pushed. Its a basic transform function, like blending between two key frames.
This video is trying to make something that should be cool on tis own sound futuristic by relating it to neural nets. Its not, and it isnt. Kids these days need to learn the difference between a transform algorithm and a neural network. If anything it sounds like the first step to having a T-1000. By the way.....Skynet was so afraid of the T-1000 liquid metal it kept it in a box at the bottom of the ocean surrounded by terminator hardware..... So making some is probably not a good idea....
@@ch1pnd413 It sounds like you're either a sycophant that is buying way too hard into their swinging bologna, or a purchased comment. There is nothing revolutionary here other than the material science that allowed the springy....sorry, 'compliant' elements to be so easily manufactured.
@@zombieregime If Skynet has feelings, why doesn't it respect the feelings of others? Trauma is a poor excuse to harm innocent beings.
@@CTimmerman the part you skipped over was establishing why it should respect the feelings of others. Having feelings does not inherently imply an empathy towards other beings who express an adequate level of sentience. The cold hard truth lost on the youth of today, and honestly anyone else who hasnt given the world a think from an unbiased third party point of view, is that no entity is inherently obligated to act in your best interest. Also, it is impossible to regulate away unsavory behavior. Lastly, when the powers that be share your sensibilities its call progress, when they dont its called oppression. Oppression can come from any side, and is inched along by the refusal to consider concessions for those lifestyles you disagree with. Punishing the many for the sins of the few by way of wild assumptions compounded with the inability or unwillingness to hear and understand those whos rights, freedoms, and liberties policy built on assumptions affects. Your rights have a limit, and they end where another's begin. As theirs are limited to where yours begin.
However, while that gives us a framework for how we may approach coexistence (but is not intended to be a instructional pamphlet, telling us how we should feel or behave in general. That is up to the person to conduct themselves respectfully. And if they cant figure out how to do that for themselves, to have their own thoughts, and feelings, separate from the zeitgeist, then maybe commentary on societal convention is something they shouldn't be engaging in....) What does any of that have to do with a machine and whatever behavior it exhibits that we might classify as 'feelings' or 'intent' or 'desire'? Why should a computer 'care' about you? Or anyone for that matter?
And I do challenge you to avoid the trap of assuming any algorithm, however complex and misnomered, is actually sentient on any level....
This is an incredible combination of an entertaining youtube video and a technical paper presentation! I wish more articles were presented like this
Watch @twominutepapers it’s similar I think
@@6acosta9Thanks, for the suggestion! It was interesting in the beginning, but it feels a bit mainstream nowadays, presenting the results instead of diving into the paper's details
It immediately went off the rails at "Similar to biological brains"
@@whatilearnttoday5295 Not really
My heart goes out to the graduate students who did all this work. You guys are ferocious, you deserve only the best in life.
I went from complete "what is this I don't even" to "okay this makes sense, cool" in 20 minutes. Very well presented, super interesting and understandable even to someone with zero experience in mechanical engineering.
Interesting, yet my instinctive reaction is that using a digital computer and sensors is going to be more cost effective than this 'compliant material' stuff.
@@Dan-dy8zp I don't believe the goal is for computation, I believe the goal is more physically focused. The compliant materials are pretty necessary for utilizing this stuff at a smaller scale, especially cheaply.
In a general sense this is what bone and connective tissues do. They have built-in stress sensors that look for electrical signals that appear in weak spots in the bone and connective tissue. They rebuild the structure to fix those weak spot s and redistribute load.
Using these materials for robot bodies might help them walk better.
@@Castle3179 In the most extreme case, this is what nanotechnology would accomplish at some point. Instead of a solid bar used as a leg, for example, a composite of nano-scale versions of these nodes and beams which can independently function and reorient with, say, a goal of "optimize stress distribution", and a combination of load-bearing and stress absorbing materials around them could result in a synthetic version of bone tissue. And, if run in the opposite direction, power could be applied to the linkages in such a way that they contract on demand to make synthetic muscle tissue. Furthermore, the lattice isn't limited to two-dimensional organization. Tetrahedral lattice, I would anticipate, would likely be a highly optimal way to distribute forces throughout a volume rather than just across a plane. Though, when it comes to organic growth, like bone tissue, the structure tends to orient itself _along_ lines of stress so it's more like it would determine which paths require flexion and develop flexible connections along those lines, and which lines of stress require maximum stiffness and concentrate the most load-bearing material along them. So nano-cells within the material would periodically be redistributing stiff load-bearing material and soft cushioning material around within it to accomplish the creation of micro-struts and micro-cushions inside the composite material just as osteoblasts, osteocytes, and osteoclasts do for bone.
This is insane. Soon we'll be doing this kind of stuff with photolithography. Perhaps it'll be the next step in neural networks as a whole to increase efficiency.
There are other ways to make Neural accelerator chips, so I think miniaturising this would be better for materials science like the video suggests.
@@EmceeJoseph Yes and those other ways aren't enough of an improvement over GPUs to be worth considering right now lol. That's a sentiment from Ilya Sutskever himself. There's nothing stopping it from being used for both applications.
@@poipoi300 The fact that this team made a tool that to simulate larger nets says a lot to how this not very useful for computation.
@@generalpurposevehicl6100 The simulation tool they've made is useful because it allows for rapid prototyping without the need of physical assembly or materials. I don't understand how you arrived to the conclusion you did, because there is no link with the premise.
@@poipoi300 I am refering to the material for use in computing.
I wonder if you could use plant cells to do something like this. Have a gas-filled vacuole inflate/deflate across a uniform foam of cells, which alters the tension against the cell walls, allowing for control over the material stiffness.
plants already do this naturally to grow twards light, but imagine it being used as an organic wing. I imagine it would be made up of something like cactus flesh, filled with a microfluidic network to control local stiffness.
If nothing else plants probably serve as a good model.
@@hedgehog3180 I'm imagining you could compare it to the varioshore 3d printer filament, where you can change how dense the material can be.
Just with living material that you'd produce in cell culture. plants normally change their cell size as they grow towards light, but the mechanism I'm thinking about is how cactus will uptake water and change their stiffness.
maybe, something like those fruit molds farmers use could make a model object to compare. melon farmers sometimes use acrylic molds to make a cube shaped fruit or the like.
That's reallt smart
Interesting. What have you studied to be able to come up with something like that?
@@Joseh-le4yl my degree is in molecular biology, but I work in cell culture. I have an interest in microfluidics and 3d printing.
this video is fascinating tho, I heard about complaint mechanisms, but trying to program them as if it's a neural net is crazy.
I guess after trying the dynamic learning you could (mass) produce a hard-coded version with the same values and just 3D printing :)
I was thinking about the same thing. This is the FPGA for metamaterials.
I mean, you can mass produce ASICs made for a specific model (or type of model) and distirbute it with a copy of the software. It has many fewer points of failure.
@@ShiroKage009Less than the FPGA or less than the mechanical array? I believe what @cougarten meant was to dynamically test the array for the target conditions and then send your client or manufacturing department a 3D print or a schematic model of the final lattice.
@@claws61821 a chip has fewer failure points than a mechanical system just because it's not a mechanical system.
you probably can do it with proteins
I thought the whole point was it's a neural net where the weights have a physical meaning (i.e. the displacement), but you don't represent it that way or use gradient descent to optimize the weights. The main benefit of neural networks is that they are differentiable and so can be efficiently trained with gradient descent.
Yep. Agreed.
The weights do have physical meaning, the beam stiffness.
It is not sensible to have the weights define any dimension, since there are many impossible configurations, which would require calculation to avoid damage.
This bothered me too. In fact, if I understand correctly, there is no error back-propagation occuring in this setup. By using a genetic algorithm in the way that was described, this crucial step is simply avoided. Perhaps this is not true for the other optimisation methods studied? It seems like such a system would benefit from more rigorous weight tuning procedures.
@@dougaltolan3017how is beam stiffness information back propagted? The genetic algorithm appears to avoid this, if I understand correctly.
You should be able to manufacture a much cheaper and easier to scale version of this by using electro-osmotic cells (cellulose membrane tube with internal electrode between two plates is probably the simplest) as the stiffness altering actuator. Simply increase the voltage on the cell to increase the internal pressure.
Cool. How did you figure this out?
Just occurred to me while watching the video, but I think the slowed down thought process was something like this; ok, the minimum easy to control input is an electric impulse, which also allows us to sense the structure as well, so how to we translate electricity into force? Well there's no movement needed, so the actuator doesn't have to actually move, just increase the pressure it's exerting. Osmosis through a semipermeable membrane can be directly modulated by electric charge...
This is amazing. Miniaturized / nanoscale applications of this really could drive world-changing metamaterial developments. It's also a very helpful way to unpack and visualize the fairly opaque world of learning neural networks in general. I never mind waiting for content like this. Thanks so much for the walkthrough!
Unfortunately, neural network learning comes under the heading of don't believe the hype.
Psychosis and gaming are serious issues with neural networks, the consequences of which can, and are likely to, be catastrophic.
Other machine learning paradigms exist, and may well be more appropriate.
Materials Science and Engineering dropout here. I couldn’t hack it in academia at that level, I had the smarts but it was too much stress and pressure. But I still love the subject matter, I think it’s absolutely fascinating, and stuff like this video is what sent me into that field in the first place. Thank you for the detailed breakdown, this was awesome to watch.
In sweden you can study at university level at a slower pace if you wanted to, do you have that opportunity where you live? Such a shame to let it go if you liked it
@@flyingpotatoe1299 in theory the option exists but it would have cost me a fortune. I’m in the US, tuition for the school I was attending is around 80,000 USD annually. Most of that was covered by financial aid but that only lasts 4 years, so if I took 5 or 6 to get my degree I’d have to pay near-full tuition.
In any case, I have no desire to go back, at least not into MatSci. Career-prospect wise, it’s a bad fit for me, as I don’t have any interest in doing research and the job options outside of that are extremely competitive. I’m over having that stress in my life. Currently, I’m working on going back to college for a degree in English, with the end goal of going into technical writing. Much more my speed.
Very interesting! Though it feels like there will be a lot of problems with miniaturising this type of system. My intuition tells me that most miniature things wouldn't be tunable by the connections between nodes, but rather the nodes themselves. For example I could imagine a theoretical case where each node has some sort of "pressure" that it applies universilly to all of its neighbors. It may even be as simple as laying out a latice of beads either of different materials, or hollow with different air pressures or wall thicknesses.
Thus, what I would be most interested in seeing next is simulating a node-pressure centric model, to see if changing the adjustable factors from the beams to them would still be able to produce the behaviours that were exhibited in this video.
Absolutely fantastic! I look forward to seeing this get progressively miniaturised.
I look forward to it being developed in 3D.
What I mean is right now it processes esentially in a plane, an x, y axis. Wait until we can do this on an x, y, z axis!
@@Jamelith Being serious, wouldn't that just require a different and significantly more complex hub design?
This is fantastic way to present your paper. Very interesting research, I am looking forward to more work from your lab!
This is an idea that I had I wanted to share with yall. This idea has been partially implemented in the video but I want to extend it. What if instead of optimizing the model in the real world you created a computer simulation that would give you more accurate results and a much faster interface (because it's software software instead of software real world). Now that you are doing the simulation part purely digitally you don't really need such a complicated mechanism to vary the stiffness. Instead, you could export the result of the computer simulation in a format readable by 3D printers. Instead of your current mechanism, you could have something like a coil that could be stiffness-manipulated by varying its width. Now, yes, this is a much less "dynamic" approach because it does not allow you to change the values on-the-fly and requires you to 3D print your material every time you want to test it in the real world but as long as your Simulation -> Real World process is accurate enough you should not need to 3D print your material every time you want to test it and should be able to do it using only software and only need to 3D print it when you want to be absolutely sure that the material behaves as it should.
I feel like having a lattice of adjustable stiffness beams is the much, much more impressive feat than the neural network part. Like, does such a lattice exist in usable ways (ie building or clothing scale devices)?
It was this underplayed note that rung out through the whole video. It was the tour du force that made possible the investigation of their tunability. As I remarked above I'd be very curious to see the tuned parameters fixed (glued) in place to see if the unpowered network behaves the same way. There are other videos on variable stiffness 3d prints producing non-linear behaviors. Using these principles to predict the behaviors of given prints would go a long way to making that become a standard engineering practice.
@@ExtantFrodo2 That's pretty analogous to practical application of neural networks today. In many applications where the network needs to be deployed at the edge (not in a datacenter), the network is designed and trained on large purpose built servers. After the weights and biases are established for the network, a fixed voltage gate chip can be created that is small in size, low in power requirement, and extremely fast. That chip can then be deployed at the edge in small devices. It just requires a complete replacement if the network is later optimized.
I can see that applying here. Use a complicated setup in the video to determine optimal parameters for the network design and task, then transfer those parameters to a fixed system as you described that can be easily and cheaply deployed.
That's a fascinating idea! Instead of deploying the learning mechanics directly, we could potentially use computer simulation and optimization to design our desired model. By simulating and optimizing the design, we can determine the best configuration without the need for a trainable machine. Once we have the optimized model, we can then build it physically, thereby bypassing the energy-intensive process of training a machine from scratch. This approach has the potential to save a tremendous amount of energy while still achieving the desired final state.
Found the computer scientist
this is one of those things that is first off awesome to share with the world in this format (no way in hell i would have ever read the paper) and also an extremely interesting idea with genuine potential to change the world, thank you so much for taking the time to make such an entertaining and informative video!
Thanks for this, this is super cool!
It's given me inspiration for a potential project of my own, albeit much lower budget and tech.
Consider a PLA 3D printed lattice in a similar configuration as the triangular one you used here, but using a slight curve on the beams to allow them to bend.
Consider then pressing the lattice into a mold with a force, to the point of plastic deformation. The plastic deformation of the compliant mechanisms - the damage the beams suffer - could serve as a kind of learning process, reducing the weights of certain beams, and increasing the strain and thus weights on others.
Setting this apart from traditional machine learning - aside from the medium - is that the training is not easily reversible; the plastic deformation can't be undone, and for a weakened beam to become relevant again can only happen within the context of other interacting beams becoming relatively weaker too. Thus, I don't think it could learn many behaviours, as the system is essentially lossy.
I'm not aware of any literature that tests neural networks whose weights can only ever shift in one direction; they would naturally be less accurate, and you would have to take a very slow and conservative learning approach so as not to totally collapse the system, but I would be very interested to see how it goes. Perhaps I'll start off with that kind of computational model.
I'm also not sure how effectively it would work with simply molding it to shape, as it could use the mold as a crutch with different output forces on different locations, resulting in a different shape when not constrained by the mold. Perhaps rather than a primitive mold, then, a rig of fource gauges at the output locations could be there and seek to find where the output force is zero at the desired location; if the material is overpressuring a certain output location, you can apply a counteractive force at JUST that output location to create plastic deformation until said output force IS zero. This would have to be done conservatively and stepwise, as reducing the error at that location will inevitably create more error across the other locations; the maximal error output would have to be tweaked slightly, then the next, etc.
Would love to know your thoughts, and thanks again! :)
Is there a reason why something like gradient descent / backpropagation wasn't used to calculate the values as opposed to evolutionary search? Was the issue that the 'hard stops' prevented backpropagation from being used?
Isn't gradient descent a feature of Nelder-Mead method?
It might seem hard to compute, but in reality many neural networks are fully connected, meaning every node connects to every node in the next layer, while here each node only connects to 3 nodes in the next layer.
What do you mean by "in reality"‽
I'm actually serious, do you mean in practical use in a machine learning environment or do you mean biological systems?
In biological systems even though long axons can connect to groups of neurons at a distance, I would not in any way consider it fully connected.
If you know any references plotting actual connectivity vs proximity I'd be interested.
As for machine learning environments I'd still argue, when you get to large models that need to be distributed amongst many systems, then being fully connected is an unlikely option.
Yes, and no.....
You are right that the nodes are not fully connected, but while there is only direct connection to 3 nodes in the next layer, there are also lateral connections, the effect will propogate sideways beyond those three nodes (attenuating with distance). In a contemporary NN there is no equivalence of that lateral connection.
@@markaspen What is that 'good reason' and how does that relate to a network that is not fully connected?
As for 60 years, you are glossing over the decade+ hiatus during the 70s and early 80s, during which little or no development was done.
The post 80s work was so significantly more advanced than previous contributions, the difference is like modern cpus vs the first "computers" that were no more than programmable calculators.
It was late 80s, early 90s when I first really became aware of neural networks and machine learning.
Virtually right away I proposed the concept of NN psychoses. The idea was shot down, out of hand, by PhD researchers in the field. 20 years later there were reams of academic papers detailing exactly what I had put forward.
So do consider my extensive knowledge, understanding, and scepticism of the topic in any reply.
This format needs to be the standard for research papers going forward
Whoever made this video if I had One billion dollars I would share it with you and develop this with you.
Excellent explanation, rare I find something of such high quality.
The joys of turbulence and material science continue. It is interesting to contemplate where and how nature employs similar functions in organism behaviors.
As others have already mentioned, the weights/stiffness could probably be simulated and trained on a computer so that it would be cheaper and faster. Then, once training has finished, the resulting model could be manufactured with the determined stiffness between the nodes.
Regardless, this is a really great paper and video. Good job on getting featured on the front cover as well.
The most detailed video I have ever seen. I have only understand 1% but it is so beautiful to watch because it’s so well explained.
Wow! The idea of making a material that can change its configuration based on learned input is amazing!
I´d loved to learn more about the interface of neuronal networks and machines! It was a really cool progect!
Were other lattice configurations considered or tested at any point?
I'd be curious how something like a hexagonal lattice performs.
You could use the simulated network to generate a stiffness map for a certain behavior and then wire EDM a big aluminium plate compliant mechanism with these stiffness values. This results in a preprogrammed mechanical network that is easier to manufacture and scaleable.
This idea worth a Nobel prize! Great job, guys! In the future we will create a metamaterial that can morph into anything and be controlled by brain. This is real deal, I must say as an engineer.
This was the most mind blowing thing I've learned in the past few years! Thankyou.
I've been bamboozled. I thought this was a fully mechanical learning mechanism but it's actually a hybrid between electronics and physical mechanisms.
What I'm saying is that the neural network exists on a computer and the physical mechanism is providing feedback. Meaning that the computer doesn't have to simulate the physics in order to train the digital neural network accordingly.
I read the title and thought that the physical mechanism was the network doing the training. No, it's a mechanical network of joints that is controlled by an artificial neural network.
This almost feels like it should win some kind of award.
We need more research papers presented in more accessible formats like this. Great work.
This is really interesting and impressive! Your video also brought some methods to my attention that may be useful to me when making topology optimization add-ons for FEM packages where only the output and not the underlying stiffness matrix is available.
This is incredible. Bravo. So much to learn from this!
I was a little surprised that what seemed to be a natural next step wasn't mentioned. I thought that they would construct a simple network of beams with the resultant fixed stiffneses to demonstrate that the target behaviors would be achieved.
It would be amazing to use a lattice of nitinol to do this. If you had a system to heat and cool sections of the lattice to impart memory you could really generate a lot of final outcomes. Aircraft and spacecraft/landers/rovers will be changed so much be this technology.
21:55 recalculate this but keep the number of beams constant amongst the lattices instead of the number of layers if you wish to make inferences on the orthogonal direction limitations influencing adaptability
Also, in this study it's definitely the number of beams influencing time and monetary costs, not layers (which indirectly influences number of beams based on beam-to-node ratios), so it'd probably be helpful to do this test more fairly for practical reasons, too
This is beautiful! So many different concepts from different engineering classes are demonstrated in this video with elegant visual effects. Something especially promising seems to me to be vibration dampening. I recall that the LIGO has active vibration dampening to isolate its sensitive sensors from Earthly disturbances. If this could be trained to negate lots of different frequencies at every temperature, it would probably save some engineers somewhere a lot of work.
Excellent work and presentation! I felt hooked all the way through. I guess these lattices will at some point be etched or 3d-printed so the stiffnesses will be "hardcoded" into the geometry of the lattice elements. Basically a compliant structure with set stiffnesses. Maybe at some point the "resolution" of the lattice will become high enough so you could talk about a continuous stiffness distribution with an analytic description rather than a matrix of distinct values. Maybe behaviours could be trained in order of importance, first learning a common movement and then refining that by overlaying more precise modes of movement. Or maybe I have it backwards. On the topic of stiffness distribution, it could be represented by a bitmap. The layered compression technique in JPG format would likely go hand in hand with my previous point of layered precision.
It is a general principle in machine learning that input nodes are exponentially more than the output nodes. This will be a much more accurate representation of a neural network. It will also be very interesting if you can recreate a feedback loop node,
Amazing work and channel I'm glad to find ! A few months ago came to me this idea of mechanical computing kind of the same way you do here but with much more clumsy mechanics because I'm not engineer lol I couldn't figure out it was a neural network problem and was thinking more about a kind of crappy manual qbit processor if it makes sense to you.
So I got the idea while looking at scissor extension arm and imaging that if all hinges could slide along both scissors it links, it will vary the position of the end of arm, the X,Y outputs, the start of each links of scissors being the inputs. I have the intuition that if the hinges, or node, could be controlled by some arduino and servos to slide onto the pair the result can be interesting and maybe able to achieve some of the computing you're doing here. But I now see flaws in my design that the scissor is rigid to a line and a lenght so the node hasn't as much freedom of movement as in this clever design. Insights needed for improvements and if someone wants to realise a prototype of my design please feel free but tell me :) Subscribed !
Have you tried incorporating static friction or backlash into the computational model? (I imagine practical applications would not exclusively use flexures and voice coils.) Have you tried 3D networks? Have you tried asymmetrical stiffness functions like Rectified Linear Unit?
I was specifically wondering about the 3D networks myself. It feels like something that gets massively ignored in digital electronics and in general programming.
Very thought-provoking. Jolly good work from this team.
Amazing work, looking forward to seeing the progress on this and if it can be made using other methods.
Interestingly I was thinking about a mechanical neural network made up of nodes and springs with variable tensions and now I see this!! Fascinating!
I'm a Comp Sci grad and this is the first time in 20 years I've seen computer theory being applied physically. Usually, comp sci concepts are developed in the reverse direction. Incredible stuff. Very fucking cool.
Congratulations to all involved.
Such incredible stuff. I love it when work in one field is applied to another seemingly unrelated field. Just goes to show that everything in life is interconnected
There are people that legitimately think that going to space is so crazy that we havent done it. Meanwhile, were doing this. Its nuts
...mind totally blown! Mechanical neural networks, and you just demonstrated it! In my face! Oh goddess that's so beautiful, tyvm...x
Now to replace the magnets with ceramics for piezzo-ceramic effect to minimize the size, use the piezzo effect for both actuating and measuring the position, make it microscopic and new era of materials is here.
Amazing video, great explanation and excellent visualizations. Thank you!
i remember the feeling of seeing pictures of the earliest computers, with their large, clunky electromagnets for bits, slow clock speeds, and room sized casings and thinking "look how far we've come". i hope in 60 years, i'll walk back into this video and think the same.
A simpler integration of this is using piezoelectric materials with electrodes placed at either end to conduct energy, where a simple change in both voltage and current direction can fluctuate the volume of the crystal. Using a material with a high piezoelectric coefficient, like PMN-PT, would generate the highest volume change and possibly be more accurate in this neural network configuration. Possibly coating fixed conducting cells with this material could allow for nano sized mechanical networks, but idk.
If you are going the route of mimicking a neural network, why not also learn the weights in the same way, i.e. by using Gradient descent?
It appears that the focus of the video/project is NOT that your mechanical structure realizes neural networks. It’s more so that your structure allows configurable connections such that given some input stress, the connections will bend to form some desired output shape using only stress to determine the shape (and some barebones sensors). I would be curious to understand what are the mechanical limitations placed on each layer and how does that restrain the output shape.
I wonder if you can change the learning objective a bit, not asking it to learn a fixed number of random outputs, but instead, to explore the space of all input output combinations, and predict which inputs, together with which weights, correspond to which outputs.
For this you could employ what's called a Quality-Diversity algorithm such as MAP-Elites.
That way, once sufficiently trained, it ought to be able to more or less directly give you a pretty good guess at a configuration for some combination of inputs and outputs, likely including unseen behaviors.
You could also make it more robust by randomly "damaging" some connections (leaving them at zero or maximum stiffness no matter what), meaning the network needs to find other ways to figure out what to do.
And the other thing I wonder about is: just how small could you get this to be? Would it be possible to do this at microscopic scales? Each individual piece probably wouldn't have a very large range of stiffnesses, but across many layers that ought to be able to add up, giving you an extremely flexible material.
Third, perhaps obviously, what about 3D? I guess you'd need a network of tetrahedra in that case for the best outcomes.
that has to be one of the most amazing things I've seen in a while
Both brakets can be printed on the body, you can load manually to fit the coil magnet and you reduce the assembly error tolerance, number of parts and can be made smaller much easier
Congrats on the front cover of Science!
Interesting. Though as a software guy this still looks to me like a simple, old-fashioned neural network, nothing mechanical about it.
You have inputs (input and output forces) and outputs (the individual stiffness configurations for each beam) and try to optimize them to a certain goal.
The learning still takes place on that virtual, digital layer, not on the mechanical layer. Therefore, I wouldn't have called it a "mechanical neural network", but maybe rather something like "application of neural networks to mechanical problems" - though the former is certainly snappier.
Nevertheless, it is an interesting application of neural networks that is presented very well and I definitely enjoyed watching this.
Please make another video once you build a bigger one!
The transistor was once quite a bit larger than it is now. Hopefully this material can be shrunk down and mass produced to that degree too. Will etching be the way to go?
This is glorious. Not least of all because I thought of it forty years ago - but I shouldn't say that. The presentation is first rate and the narrator is very easy to listen to. To see a working model of this concept is truly beautiful.
This is a great video. I am going to watch it a few times to try and get my mind around it to some degree. I appreciate how this seems to be insight into both mechanical and AI learning. Adaptive physical algorithms designed to result in a specific task or outcome from specific inputs.
The four fundamental interactions of physics are the strong and weak nuclear force, electromagnetism, and gravity. To copy a relative ranking of these relative forces:
Gravitational Force - Weakest force; but has infinite range. ( Not part of the standard model)
Weak Nuclear Force - Next weakest; but short range.
Electromagnetic Force - Stronger, with infinite range.
Strong Nuclear Force - Strongest; but short range.
The four fundamental interaction of physics are interactions, these can influence on another. Consider a molecule or compound that might be adapted similar to a learning material through an influence, such as for example, electromagnetic wave interference. Can electromagnetic wave interference deliver energy or influence at an atomic or sub atomic level? Can we facilitate chemical reaction with electromagnetic wave interference? This is major question.
Would it be possible to manipulate an adaptive material with wave interference? Can wave interference serve the same function as the weights or tunable beams in the artificial or mechanical neural networks?
A specific application that came to mind is actually re-magnetization. If we have a field geometry that breaks down over time due to atomic re-alignment, might it be possible to efficiently and perhaps continuously re-magnetize, re-align in the desired alignment, a magnetic material, using wave interference? That is a marginal hypothetical question. We may now use electromagnetism to re-align and re-magnetize demagnetized material, however, this is a brut force approach. Might it be possible to use a more subtle, specifically targeted and efficient re-alignment, using wave interference interacting with atoms?
To continue the example, if we have a microwave antenna, the physical configuration of this antenna, I believe, can create a polarized field utilizing the microwave energy. This polarized field would be within an area. Might it be possible to create a similar geometry of polarization on more of an atomic or sub atomic scale through wave interference? Perhaps an approach similar to triangulation, which might result in a specific frequency and energy at a specific point, that is also aligned in a specific manner, comparable to polarization and re-magnetization?
Re-magnetization might seem a peripheral utility of adaptive materials. If we have materials that can be adapted with electromagnetism, or adapt to electromagnetism, it might have a much broader utility.
The artificial or mechanical neural networks look very reminiscent of atomic or molecular structures.
Correct me if I am wrong: but the stiffness is determined by the coil and the magnet, which in turn are controlled by the esp32 and the mosfets. At this stage it is the software that 'learns', not the material itself.
It is not shown how the mechanical structures themselves have "learned' to behave in the same way without coils and magnets.
This means that the software is a - quite complex - PID (Proportional, Integral, Derivative) routine written in code. There, all inputs (electrical, not mechanical) are read by the microprocessors, which results in the desired output being calculated and corrected after each cycle based on the comparison between the current and previous input. The power of learning is in the code: the stronger the code, the faster (fewer cycles) the desired result is achieved. Corrections - no matter how small - will always be necessary. So, in my opinion, this shows how a neural network works, but it is not mechanical. Mechanics is used to show how the learning process works. You can do that with any spring-like material. So until proven otherwise, and in the absence of any electronics, it is not a mechanical neural network. (The tuning fork shows exactly that: the fork does not learn (has not learned, will not learn) because it would change over time, which it never does)
This made me think of the experiment where a programmable array was used to "evolve" a basic radio.
In the end they bended up with a functional 2 way radio, but it contained components that weren't connected to any part of the circuit but could not be removed because the radio would stop working.
It showed that "real world devices" might get results simulations would miss because we never thought of adding them in.
I might have fuzzed up some of the details, it was some time ago I came across it.
Do you remember the name of the study or its authors?
Thank you for making this video Professor Hopkins!
Very interesting ! Thanks !! This reminds me a conference on AI 30 years ago, explaining that very first NN consisted in fully connected arrays of transistors. Due to practical resaons (i.e. the factorial growth of soldering/wiring operations) , these arrays were limited to a very small set of elements. But their performances were promising enough to enable further works...
I would love to see a physical demonstration, a built structure based on the optimized axial stiffness within chosen reason. Would be very cool to see how that preforms in practice. As for applications, not all targets will be in a triangular configuration but that is okay. As long as the target configuration can be modeled and with enough variable "nodes" as presented (which has the capacity to be optimized and exhibit desired behavior). The real trouble I imagine is tolerance in practice. This is incredible research regardless utilizing so many fields of sciences.
This is an amazing idea. Thank you for synthesizing the concept from something that existed in software to something in the mechanical world. As this technology gets developed and shrinks in size, its applications will be limitless. This is so cool!
I really commend the accessibility and transparency of this content. Excellent work, this seems like very promising technology!
Has it been minaturized into MEMS yet? If so does it have interesting properties relative to electronic and optoelectronic integrated circuits?
Interesting demonstration. Though I'd say the material itself isn't learning anything, just being tuned by a computer. Maybe I'm just nitpicking though. Either way, this is fascinating.
This is absolutely amazing! And you just give us the designs for free. Thankyou for doing such great works!
As a mechanical engineering and computer science double major with a minor in robotics and concentration in AI this is incredibly cool. Absolutely fascinating application of AI. I’m going to fall down this rabbit hole now.
literally nobody cares what you studied, there was no reason to mention it
Instead of having an outside controller, is there any way that the learning mechanism could be built into the individual modules?
Wow, this is a wondrous example of engineering.
Have you guys thought about something like a tensegrity structure but modified so that 1 string can determine the tension of the joint? That way you don't need to sense the force into to apply the magnetic field dynamically. I imagine you can just rotate an axle to tighten or relax the string.
Incredible! So cool. I'd love to see static models developed this way
This could be very interesting to the construction business. Imagine the steel frame of buildings beeing designed in such a manner, that the building has a specific desireed reaction to certain stresses. It's mind blowing!
Hello Jonathon!
I've been following your work for quite a long time, nice to find you showing some interesting things here on YT... it's so surprising to see how few mechanism designers know about, let alone use, flexural mechanisms.... Sigh.... Great to "see" you!
If it were me, I'd test inverse tan for a nonlinear stiffness function too. Not sure it'd be effective, just think it'd be interesting. Though it may need a larger network to really shine.
You could simulate a network that is modelled on what you can 3d print. So you devise a smooth compliant flexure you can print, with some element of its structure thinner or thicker to effect it’s resistance. Then print a solved result from a simulation and test it in the real world. While not a device that solves its self, you could prove the principle with way more nodes.
The design would probably be mostly 2d but in the vertical direction alternating wall thickness could give a resistance that has a not so integer difference from one flexure to the next. Ether way you could print a truely detailed result of a simulation with hundreds of nodes without blowing the budget. A 3D network could be printed but would probably need to be printed in a flexible resin. It would be an impressive result to have a cube that you press down on and get various 2d shapes on the other side depending on the direction of force.
In the real world this is closer to the use case. You solve a smart structure then print a static structure with smart behaviours.
Could the node arrangement ment change to accommodate different crystal structures?
Yes!
As a computer engineer, I am amazed by this. The possibilities are endless. How is it going to get smaller? for instance in a body armor we can't have processor units for controlling each node, so I'm really curious how it can be done on a smaller more practical level.
Photon-sensitive materials which expand or contract according to LED lights can also be made.
Really fascinating topic.
Thank you
no, I am quite certain that this will *not* replace classical materials for multiple reasons: (1) cost (2) miniaturize this, there is a limit as to how small you can make it and it never can be as small as a purposely made material; in particular the actuators and gaps are a problem (3) it can fail to learn
This is fascinating research with amazing use cases via further development.
I fantasize that one day we will be saying "hold on let the building load before you enter"
But ... isn't this learning done by software, via feedback from the mechanical setup ... and not the actual mechanical setup?
Perhaps the end goal is to use the learned stiffness values, to then produce physical objects, without sensors, actuators and computers to exhibit the learned behavior ... or??
I expect they could fix or glue the tuned parameters in place to see if the unpowered network behaves the same way. Of course it wouldn't be amenable to learning new behaviors, but maybe that's not the point.
@@ExtantFrodo2 they explicitly mention in comments that "learning on the fly" is a goal, this contradicts fixed parameters
@@toxicore1190 If you say so. I'd suggest that being able to establish variable responces by "fixing" the learned tensions and compressions is just as valid.
Questions:
1. In practical use, how does this perform in 3D space and not just on a single plane?
Is there plans to make a 3D version of this?
2. How far from the input node can an output node be effected?
Great work. It is basically a truss system: F = Kx where K is the stiffness matrix and F and x are applied force and resulting displacement vectors. The structure of the matrix K is given by the geometry of the linkage system (basically a banded symmetric matrix.) The question is: given a set of F and x vectors find the elements of K. In the linear case it looks like this can be achieved using gradient descent or least squares methods. With too many actuators, this is a complicated control problem: The system matrix and input matrix are part of the control. One would think this allows solutions where we can isolate the inputs with a layer of zero stiffness linkages thereby making the input matrix zero. Since the inputs are force specification this will result in mechanical stops in the linkages being engaged allowing the input force transmitted to the next layer. I would think this is more of an impedance control problem than an NN problem. If the inputs were the displacement imposition and the output was the desired force distribution then the problem statement would be more inline with impact problems, vibration isolation and earthquakes.
Neural Net application gives the impression that the stiffnesses will be computed once this will basically be a static system. The test apparatus developed looks a lot more capable than the static applications.
Can this be used to make something that would be pierced by a bullet to be manipulated in such a way that it can essentially translate the kinetic energy to somewhere else. Like if you had a bit of Kevlar Hooked up to a large array of tiny versions of this in body armor, could it *newtons cradle* the energy to the ground or out your back? Or simply disperse it amongst your body?
Just to make sure I'm getting things right here, this essentially means we will be able to have materials that act as we want them to, rather than us having to design/discover materials that have the particular functions we desire? For example, we may use silk for now as a extremely soft material, but, using this technology we would (after sufficient scaling) be able to make a material act as soft as silk or even softer? Or, you could design a button that when pushed, could be trained to give a specific 'feeling' (ie hard to push, or easy to push).
Would appreciate it if someone could clarify for me!
Coincidentally, for the last few months, I've been "backburner" working on something like this but using ambient light through a vascular network to control/actuate SMM (shape memory material) robots
definitely helped me understand neural networks more even if not intended, Bravo!
As a mechanical engineer i like the idea of this setup, but what is the benefit compared to just jumping straight to simuting the network? We have very good FEM and generative design tools. Instead of changing the shape you would change the beam stiffness in order to reach the desired output. Once you find a solution you can create a 3d printable network that has a specific material thickness for each beam to reach the desired stiffness. This network made of one solid piece can then be used for real world validation. This way you would be able to create much bigger networks with much lower cost.
Can someone please explain me why would one need to play around with these mechanical beams to train the network instead of simulating them all as simple model, calculate all the weights similarly to how classical neural networks are trained, and then simply apply the calculated weights (as stiffness) to the beams effectively using them only as actuators?