Это видео недоступно.
Сожалеем об этом.
What we Learned about Compliant Mechanical-neural-network Learning through Failure
HTML-код
- Опубликовано: 18 авг 2024
- This video details what we learned from our first and failed attempt at building a compliant mechanical neural network for the purpose of demonstrating the ability for an architected material to autonomously learn desired mechanical behaviors and properties. Although, the attempt failed and the machine never successfully learned any behaviors, it ironically taught us more about the learning process than the next improved design that did successfully demonstrate learning. Thus, this video explains all that was wrong with our first attempt and what improvements were necessary to enable a physical compliant mechanism to autonomously learn.
To better understand the contents of this more advanced video, it is recommended that you first familiarize yourself with the fundamental concepts by watching this seminal video titled, “Compliant mechanisms that LEARN! - Mechanical Neural Network Architected Materials.”
Video link: • Compliant Mechanisms t...
Also, we encourage you to read the following journal publication, which first introduced this content and can be accessed at the provided link:
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.or...
Donate to help support my channel:
If you’d like to make a one-time donation, you can use the following link:
www.paypal.com...
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 who built both mechanical neural networks featured in the video. I am also grateful to my AFOSR program officer, Byung “Les” Lee, who funded the research that this video features and was patient with our failures.
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.
Documenting what was tried and did not work is an underappreciated contribution to Science
It's actually a result of our economic system, when science is done not for educational and progress purposes but for profit the only thing that matters is "success" and it can be incentivised to act as if you never failed because those who "get it right the first time" are more likely to achieve funding for future programs even though truely the best option would be to be honest, but the people funding the science want a return on investment so lying about "success rate" which drags down the industry as a whole.
The reason I'm saying this is because we need to move past capitalism so we can incentivise honesty and actual progress over artificially inflating "successful" indevours that have a short term, direct "return on investment" in actuality the ROI is the act of moving the whole of scientific knowledge forward even at the cost of the investment
Edit: that doesn't mean we can't encourage and reward efficiency and cost effective testing of course, and by doing it this way we can actually find better and more effective ways of doing so, because we are focused on overall progress rather than just profitable ventures
@@smileyp4535Bro, science is mostly funded publicaly, the criteria for recieving grants is not profit, it's publications. The thing that's corrupt is not the economic system, it's academia itself.
Dishonesty and only publishing positive results do not lead to more profits, the opposite would be true if that research was then used for any kind of production.
I've worked in academia and although the problem is money related it's not as simple as "it's capitalism".
The academics are corrupt
Saving this
@@evelioguaperas while I agree it's more complex than just "it's capitalism bro" the problem is acedemia (while often funded publicly) is still a business in most of the world if it were universal and encouraged to all it would pay for itself and then some leaving more money for research and funding overall and less need for "results" based research and more room for "general research for researches sake" type research so really the answer is basically just "it's capitalism" but for more general society level reasons outside of just the intra-acidemia parts
Did I miss, or there is no learning curve shown for the working design. It would have been interesting to see the final learning curve of the working design.
As mentioned in the video, you can find this in the previous video linked to in the description. The previous video goes into more detail and is in fact way more informative, I highly suggest you watch it.
at this point, these videos feel like the beginnings of a cult
@@dine9093a cult of mechanical minds
All hail the mechanical neural network
“Reality is never overwhelmed by computation limits.” That’s really interesting, I never thought of it that way.
I've heard that motion force, a push/pull or torque, travels through a medium at the speed of sound in that material. There is an upper limit surely, if nothing else the speed of light. Funny how he said it as if mechanical computing is free of computational limits imposed by the properties mechanism
Idk, making 1 cuda core on gpu is cheaper than making one beam on mechanical neural network...
@@nedoran5758analog systems in general are much faster than comparable digital systems; what digital has on analog is repeatability, i.e., less sensitivity to noise
@@charlesmartin1972 Oh, no doubt. I've been interested in analog computing for years, like the antikythera and the "plumbing" model of the US economy. There are certainly use cases where analog approaches would be worthwhile even today, at least as useful as quantum computing. The more varieties the better!
Make it out of transistors and electromagnets lol. It's digital speed analog
Absolute one of the coolest things I've seen. Failures are always the best teachers.
For a number of years now, work has been proceeding in order to bring perfection to the crudely conceived idea of a transmission that would not only supply inverse reactive current for use in unilateral phase detractors, but would also be capable of automatically synchronizing cardinal grammeters. Such an instrument is the turbo encabulator.
Now basically the only new principle involved is that instead of power being generated by the relative motion of conductors and fluxes, it is produced by the modial interaction of magneto-reluctance and capacitive diractance.
The original machine had a base plate…
Yooooo its tim I didn't know you liked this stuff too
Umm, yeah I totally agree.
did we just become best friend!! yeeesss lol
😮
Id like to point out that the character of the learning algorithm was not considered in your evaluation of the failure to converge on the right response policy satisfactorily. Graduated progress and catastrophic loss of progress are common in even synthetic purely virtual implementations of neural networks training, and evolutionary algorithm training. Im an expert in deep reinforcement learning and I have seen many forms of failure to converge in virtual environments even where sensors are perfect, noiseless, and cannot fail. It is a feedback mechanism after all, and your fitness function itself may be nonlinear, and there's exploration and optimal policy preservation issues in genetic algorithms where the population is too small.
In order to really isolate this variable, you would need to train the network on a simulated model of the device first as you did in your first iteration. Interestingly, you seem to have rediscovered failure modes within modern deep learning, but within an orthogonal domain. amusing tbh fam
THE OMNISSIAH APPROVES OF THIS WORK
I'm not sure saying the universe can't be overwhelmed by computation is accurate. The universe will produce waste heat from "computational" interactions - that is, all physical interactions produce waste heat - and that does things like making stars glow. If pushed far enough, the universe does strange things like producing black holes or cosmic inflation. There is an upper limit to the amount of computations the universe can produce. So, it _can_ be overwhelmed, just not at the levels that would matter to us.
No, the above isn't really serious.
Yeah, as soon as you described the structure, I suspected that the creep exhibited by PTFE was causing the problem. It's such a weird material, mechanically and electrically. It's such an odd choice for a mechanical device but obviously you figured it out. I love the wedge fasteners. It reminds me of the tapered pins that were so common in mechanical design 100 years ago. Have you considered using a simplex optimizer with parameter descent? I used SIH in graduate school. We found that it did better than simplex alone.
Try using electro-osmotic cells to make your variable stiffness flexible units. Should let you make the whole system WAY cheaper.
Whoa this is the second video ive seen from you guys, I'm going to go through all the videos you've uploaded. Thank you.
This isn't a mechanical thing, there's boatloads of electronics driving. Just sayin.... 🤔
Now it would be truly amazing to see something that was completely mechanical that was capable of doing this. That would be absolutely astounding.
That's what we are working on currently so stay tuned!
Great work Dr. Ryan Lee! Way to pivot and deliver a beautiful result.
Thank you for making this video. This is one of my favorite & most interesting projects I've ever followed!
Did you try any training runs with the range of stiffness values limited to positive? are the negative stiffness values required for the learning to work? If you could get away with only positive stiffness values, you could do cool stuff like, train a model on the full machine, and then bake the final stiffness values into an entirely passive mechanical neural network, just a single piece of aluminum cut to form a lattice of flexures with different weights. having an entirely passive structure implement the final behavior would be pretty cool, even if you lose the in-situ training benefits, because that makes me imagine some future project scaling that to MEMS-based mechanical neural networks. I'm guessing the magnitude of possible output node forces would decrease for each additional layer in the network without anything to amplify the forces, but that's just a hunch.
Additionally, if learning is possible with positive-only stiffness values, that seems like it would open up a ton of engineering opportunities to do reduce node cost significantly by replacing the whole control loop with some passive physical element with an active means of adjusting stiffness, instead of the whole element being active.
This video is invaluable. Thank you for documenting everything that didn’t work and why. ❤
It is exactly ongoing self modification of the pieces through their use which is learning.
That pieces either becomes more compliant in use or less compliant because disuse.
I’d love to see a "hardcoded" version made with laser cut/3d printed springs after training in a simulation
They did better than that. They used wire EDM for incredible precision.
I love the explanation of why using the physical device to learn is so much more efficient than a computer simulation! Would it be difficult for these to have a feedback loop so they could constantly adapt to changing variables in real time? Like in the rover example, where it would need to adapt to conditions it's never experienced before?
When you have spent that much time and money on sending a rover to a planet you don’t want to just let a training algorithm do it’s thing. You can’t replace any broken parts and you often can’t fix any mistakes you make, you may only get one shot at it. You want everything to be tested and verified before attempting it on the actual hardware. Training a mechanical neural network in a real environment has the potential to cause damage or make a situation worse. With space missions they often have a second rover on Earth to test things before they try it on the rover on another planet.
The same approach of testing it on the functioning physical system could be used for all kinds of robotics and would be easier than creating a simulation (in theory) but people instead create the simulations and train it entirely virtually, why? It is because it is much safer, faster and more efficient. Fixing a broken mechanical part or broken piece of electronics is much harder than fixing something in a simulation or just resetting it.
You can kind of think of it like human pilots using simulators, just putting someone in a jet and getting them to learn that way is much more likely to end badly, hence they train in simulators first where there is no risk of damage to physical components.
When you have a mars rover and no way to repair it you don’t want to train on the physical system.
I wonder if it would be possible to build it on a micro-level, possibly with a quartz-based material
Maybe it could be done on silicon, and take advantage of existing chip manufacturing methods? IMU's for phones and stuff have some similar flexures, actuators, and measuring capabilities, and are manufactured that way.
@@TiagoTiagoT at that point wouldn’t you be better just running a neural network on a processor? The main point of systems like this is to have a mechanical network that can respond to and cause mechanical movements and forces. If you shrink it down to the size of MEMS devices then you lose the ability for it to interact with its environment in the way it is intended to.
@@conorstewart2214 Even if you make a big array, like about the size of a silicon wafer, and perhaps stack several layers of that?
5:46 This highlights one of the key features of learning and memory in humans and animals:
Earlier data are weighted less heavily than recent data. This grants adaptability to a changing environment.
On the other hand, earlier established patterns are statistically reinforced more times (by repetition in the environment) than recent patterns. Repetition-based weighting grants stability to mental models, by rejecting new patterns based on rare edge-case data, "fluke" changes in the environment, misperceptions, or similar kinds misinformation that would be inappropriate to generalize. The accumulation of stubborn close-mindedness explains why adults are less impressionable than children.
For example, if you try a new upscale restaurant and have poor food or service, you'll likely never return. However, if it's a new franchise location, and you've enjoyed the other locations, you'll associate the bad experience with the new location, not the whole chain.
Do you think it's possible for the learning process to also be through mechanics?
You should make one out of steel instead of aluminum. Aluminums' fatigue properties mean that its UTS will ultimately decrease to oblivion over time.
This sounds like someone just cut out parts of science magazine headlines with a pair of scissors and taped them together. Compliant mechanism neural networks??? Insanely cool.
now my wrench is going to be sentient just what i need
The latter consisted simply of six hydrocoptic marzlevanes, so fitted to the ambifacient lunar waneshaft that side fumbling was effectively prevented.
impressive. thanks for sharing
You should try making it out of nitinol, as nitinol exhibits no permanent hysteresis at all, but is still flexible.
I think you can use something like Nitinol Wires... maybe...
those esp32 voice coil controllers are interesting, have the designs/firmware been shared?
I wonder if this will see us making something like the ”Calculate Fortress” material found in the fictional Japanese ”Toaru” franchice, at least given enough iterations on this concept.
Basically, the idea is a material that pushes back when pressure is applied, such that it could withstand something like a bullet.
A mechanical neural network might function as a sensor for such a material.
it would be a million times easier to just use conventional mechanicals to move mechanical plates at pre calculated optimal rates to "catch" the bullet
what about those parafin wax linear actuators
Tell me, and everyone ...
What is a possible practical application of this?
Imagine this in place of the intersections of support beams in a skyscraper in an earthquake prone area, they can act as a muscle of sorts that counterbalances the shockwaves permeating the building and preventing it from collapsing.
@@Krimhful they can earthquake proof buildings without anything like this and without any kind of training, so what would be the advantage of this? This needing trained is a large disadvantage, what are you going to train it on? You could train it virtually or using some simulated mechanical system but it wouldn’t be the same as a real earthquake, so that only leaves training it in operation at which point training might not happen fast enough to protect from an earthquake.
Using it for what you describe is cool, in theory, practically though it most likely would not work and there are already decent methods of earthquake proofing that would probably come out cheaper and more effective than this.
Regarding the simulation training question; what if you used a simulation with simplified physics, ideal materials etc, to obtain values to initialize the physical network, and let the physical network just take care of fine-tuning the values to account for the complexities and imperfections of the real world? Would that let you reach the end result faster, or the values needed for the real network are so different from what's obtained in simulation that you might just as well start with random values?
i'm guessing it's a matter of: a very small discrepancy at one layer is amplified for each subsequent layer, until it's no longer small?
@@spambot7110 Won't that be taken care of in the physical training stage?
Amazing!!
Props to Ryan
Your mars rover example isn’t a good one, this is a relatively simple neural network that could easily be simulated and trained by an on board computer, not requiring it to be transmitted from Earth.
It would have less stress, less unusual conditions during training and less wear if the mechanical system was just characterised and then simulated and trained on the rovers on board computer and would most likely be much faster. You want to keep experiments on a working piece of hardware to a minimum as it reduces the likelihood of damage and on a rover that is that expensive and that far away you definitely do not want to run any kind of training or experiments on the hardware since you can’t replace it. You want everything to be verified before trying it, that leans into the approach I mentioned above, characterise the system mechanically and then test and train it in a simulation.
Edit: or you could do what the engineers already do which is to have a second rover on Earth that they can test in different environments and situations before they try anything on the rover itself.
Forward Forward Algorithm?
Is it really a compliant mechanism when it has a electromagnetic actuator this is still quite awesome. Thanks for the great content
I think the idea is that you can train, get the values then make flexures with the desired stiffness.
@satibel ya it took me watching the rest of the video to understand lol. But its a cool concept i think. Itd been nice to see some practical applications to help people understand the use cases
@@98f5 it's still experimental there isn't a specific application in mind afaik.
a thing I can think about is a material that has variable stiffness, like a bed that you can change the stiffness on by sliding a plate.
It might also be useful to counter the torque of a motor, where you have side load from the moment but vertical load from the bumps.
Another one could be a lock, where only the right key allows the lock to flex in the right way, or you have a key with the flexure, and it allows a question-response authentication (basically the lock presses on the key and the key deforms in a set way, which makes it extremely hard to copy, as you'd have to copy a majority of the combinations on the key, instead of a fixed bitting.)
@satibel yes i imagine it could be used for self leveling and many other interesting uses in manufacturing processes, maybe as a type of motion control of a flat plane too (not an airplane .
The actuators seem to create feedback loop. Wish i was smart enough to build software for things like this too let alone physically creating it. Huge amount of work. I gues thats the advantage to academia
@@98f5 tbh it's stuff you can learn to do on the go mostly, ml is relatively simple to code, but yeah time consuming so if you don't have the budget and time to experiment with that, it's hard to do, which is indeed the advantage of being in academy, if you're convincing you can get payed to fuck about and find out.
It's certainly interesting, but it's no more a mechanical neural network than the muscles of my hand.
You really need to test this in zero-g
think of it like this your brains a deep neural network and your body's a mechanical neural network and together they form a neural net that can learn and adapt to coming circumstances without one or ther other conscious choice does not exist as it cannot effect reality therefore if your trying to improve this your gonna need to learn how to combine neural networks together to form something truly compliant because it mimic's the brains neural plasticity and enhances the structure's shape changing and load bearing features. i wish you luck hope this helps😁 and if this doesn't i guess i'm the one learning through failure 🤣🤣
The universe is the most powerful simulator known to man.
I wanna print one
So THAT'S what's inside a Rockwell Retroencabulator
Failure may lead it to optimization
The argument(s) you use as to why directly training of the physical model is superior to computational simulations are the same points that made the first model fail. That is quiet odd.
The main reason the first design failed to learn was hysteresis, which caused lack of repeatability. It's also true that direct training takes hysteresis into account more accurately than simulation, but you can leverage the benefits of reality while still designing systems that have tolerable hysteresis so they can learn where simulation fails or is not as accurate. Note that despite all the issues with the first design, it still almost learned around all the problems! Pretty cool
👍👍👍 💐💐💐💐💐
Me, wondering what it might feel like to actually understand any of this:
🤔
Laser cutting teflon is about the most disgusting thing I've heard of all day
It was such an odd thing for them to choose. Teflon or PTFE since Teflon is a brand name, is generally used for its low friction or non stick properties, it would be interesting to know why they chose it for this, since it isn’t being used for its main properties.
@@conorstewart2214 My idea is that it's lightweight, relatively cheap, and easy to prototype.
Aluminium is heavy, hard to cut, and relatively expensive.
Most prototypes are made out of plastics and cardboard...
The role of a prototype is to identify all the problems that you do not anticipate, and this prototype perfectly do it's job, allowing the aluminium product to be fonctionnal quickly without spending days cutting and ajusting aluminium.
Cool
No no no, it's the plenum pressure in your hyperturboencabulator that's causing hysteresis in the teflon components.
An analogic silicon based eletronic ANN is the ultimate Mechanical-neural-network design.
>Mechanical
>Uses electronics
Hmm
That's a lot of failure points that could have been avoided with a little bit of thinking....
Teflon? Really? of course it'll sag
'mechanical neural network' how did you even get the funding for this project rofl
So... you built it wrong.
I think it is an error to use electricity, it will make the low cost and scalability of this dream to fall.
Hmm... I have come early it seems.