Finally, Differentiable Physics is Here!

[@Anton-cv2ti]

Came here for the impressive papers, stayed for the positive and inspiring commentary.

[@secretgames1906]

@Zarion 11 wasnt meant to be funny

[@OctavianTech]

positivity is contagious

[@dryued6874]

This is just straight-up black magic. It's like knowing where exactly to tap a rock to make it crumble into Venus de Milo.

[@teovinokur9362]

exactly!!!!!!! (except venus de milo is a painting)

[@ninjamaster224]

@@teovinokur9362 *exactly*

[@mrrock0ut821]

@@teovinokur9362 Indeed, a painting that lost its arms god knows how. c(=

[@aiksi5605]

@@ultimacy100 I too have learned a thing or two from watching Spongebob

[@matthiaseisl5623]

@@teovinokur9362 you mean boticellis venus

[@syfou725]

When your program is so powerful it can make a SQUIRREL into a GOLDFISH

[@TwoMinutePapers]

Transmutation at its finest. 👌

[@judgeomega]

@@TwoMinutePapers should have transformed a lead bar into a gold coin

[@johnsherfey3675]

Now can it hack people to make a squirrel look like a gold fish.

[@stefanoscolapasta]

Lol

[@sliceofbread2611]

i feel like the squirrel already looks quite a bit like a goldfish, the head has the right shape and the orange thing behind him looks like it could be a tail. i wonder what results it can give with other pictures.

[@AsmageddonPrince]

If you could constrain it to only apply some amount of force, you could use this to create a video game in which the player is continuously unlucky, with the physics simulation ending up with(in 2D, for example) rocks tumbling onto the player, shrapnel bouncing just right to almost strike them, wobbly rope bridges vibrating just right to make navigating them difficult... it could be absolutely amazing.

[@AmeshaSpentaArmaiti]

An RPG where even the physics is effected by your stats sounds like it has awesome potential.

[@nerobernardino88]

​@@AmeshaSpentaArmaiti Player: I'll walk twenty feet *Rolls 1* Game: Aye, you've chosen death!

[@mixer0014]

A game when you collect good and bad karma for doing things, and world itself can help you or hate you depending on your overall score

[@Rotem_S]

@@mixer0014 I imagine maybe having some kind of "narrator" like in Rimworld that controls your luck through slight variation to the world's physics. ...Actually I've met a few people who claim that's how god works... So yeah I think it'll be real cool

[@AaronRotenberg]

Unfortunately, since this is still a gradient descent technique, it's unlikely that a game would be able to run physics simulations fast enough to magically "adjust" towards a goal in real time using this method. But maybe someday... those Luck stats will actually be reality.

[@gamerboygaming]

“Dear fello scholars...” Me sitting on the couch half asleep: “ye.”

[@HaugeBauge]

damn i was about to make this comment

[@DeSinc]

this is what I imagined when I first saw one of those physics demonstrations that drops thousands of differently coloured marbles and has them all fall perfectly sorted by colour in the bottom. the way it actually works is you run the physics calculation first and then colour the marbles second, but I imagined the insane challenge it would be to actually do it for real - to start a simulation in just the right way so that the balls all fall in the correct order, or in this case, so the ink gets moved around in just the right way to form a yin yang symbol.

[@pecfexfextus4437]

didnt know desinc was a deae fellow scholar

[@neillcoetzer9133]

Oh damn it's DeSinc didn't know you'd watch this. Yeah that was also what I was thinking they originally did with those marble things

[@rubenpartono]

A cool challenge to convince the audience that colouring wasn't done as a "second step" would be to require the marbles to be sorted at the beginning too, instead of only at the end. That's sort of what the yin yang example did: it was a regular checkerboard to begin with, and the forces directed it to form the yin yang after.

[@PythonPlusPlus]

So we can now control chaos? Isn’t this the definition of magic in the Witcher?

[@GeneralPet]

Hell yeah! Though chaos is not well defined and entropy doesn't change. The yinyan arrangement is equally unique to the checker one as does every other possible arrangement.

[@arendellecitizen208]

Finite Impossibility Generator, we are soon to create the hell of a spaceship with that

[@SpaghettiToaster]

@@GeneralPet Entropy most certainly does change. A high-entropy image would not have all the black pixels clumped together.

[@GeneralPet]

@@SpaghettiToaster Entropy is related to the number of possible states, not the state itself.

[@SpaghettiToaster]

@@GeneralPet Yes, and if the pixels in the image followed real physics, say, brownian motion, there would be more states the image could subsequently reach than from an evenly distributed mix.

[@kibrika]

"published 2 minutes ago" best time to notice 2 minute papers :D

[@anunaccountablescience6464]

x10 20 minutes ago

[@chrisguy6301]

x10 20 minutes ago + 10

[@StoneGear]

Plot Twist.

[@gullit97]

Guess I just found the subject of my project for the machine learning course this year. Thank you so much.

[@shaimach]

Read up on "optimal control theory" - it has been around for over 50 years (e.g. application to quantum computing: arxiv.org/abs/1612.04929 or a more explicitly analytic approach journals.aps.org/prl/abstract/10.1103/PhysRevLett.120.150401)

[@rayray6548]

I subscribe to this channel a while ago for 3 simple reason : 1) Content is incredibly good and fascinating 2) the narrator is excellent and has a Slavic accent, (if i wrong im sorry.) easy to listen and perfect for talking science. 3) the word ''paper'' is use often and i like it , but i dont know why...lol

[@hunted4blood]

I might be misunderstanding the concepts, but I feel like this would probably be a perfect way to supplement character animations.

[@tomc.5704]

Don't know about that. All of these examples had a very specific defined goal. "Wiggle forwards." "Make a yin-yang out of smoke" "Make this image appear like a goldfish" Some of those are astoundingly difficult problems, but they're fundamentally different from "make an animation that looks good" On the other hand, you should check out DAIN, it's an AI method to enhance existing animations (or old TV shows) to higher fps

[@tiamagus6641]

@ Yeah, and then the arm decides that moving 2 inches to the left requires spiraling around to the right, adding several inches. You still need intermediary key frames, and depending on the kind of math you use, you end up with various quirks for intermediary steps (my example comes from using quaternions and Euler equations for 3D motion). It's not impossible, but it definitely requires more than you're implying. Source: Took 2.5 years of game dev (specifically, coding) in college. In my experience, the time you're implying is saved by the animator is actually spent making intermediary frames and tweaks to avoid algorithmic animation quirks. The math is beyond my level (note the .5 of a year...), but fixing those quirks is mathematically and programmatically some genius level stuff on par with accurate n-body simulations (accurate billiards simulation or cosmic motion are just different terms for hell) and (dis)proving Reimann's Hypothesis. The iterative capabilities shown in this video are some extremely impressive works, but I don't think well achieve realistic, animal-like animations without some heavy intermediary work. Key frame tweaking is going to be an unfortunate reality a while yet.

[@tomc.5704]

@ That's exactly what DAIN does, and it does it better than current interpolation methods

[@martiddy]

It seems that this is more oriented to fluid dynamics into certain shapes, so I guess you could use it to create realistic animations of characters that are made of smoke or water or any other fluid.

[@eelcohoogendoorn8044]

Contrary to what the other replies say, I fully agree; this seems to me like the most obvious commercial application of this awesome work. As an animator, all youd specify is that the bad guy needs a foot in his face so that his head spins around dramatically; and given a realistic muscoskeletar model and some reasonable but coarse initialization, this system should be able to give you a fully physically correct roundhouse kick, without any mocap or keyframes involved.

[@jimmyhsp]

i hope this gets used a ton for cgi in movies, like how you said the smoke bunny can be set up so it looks like it could actually happen, that would be cool in movies/art

[@salman_chowdhury]

the program is so powerful, it learns to roll a natural d20. every time.

[@ukaszMarianszki]

Well you could technically use this to generate a physics symulation that while looking like a random roll would always result in a 20 :p

[@janzacharias3680]

Simply beautiful... thank you very much for this channel!!

[@lorenzoa.ricciardi4264]

Very cool, but... this was known several decades ago as "optimal control" or "differential programming": to steer a general dynamical system from an initial state to a final state, so that the system's trajectory is optimal in some user defined sense. The emphasis of the paper seems to be on the efficient implementation of the algorithm, especially the automatic differentiation of the physical model. Many languages allow this, but apparently their implementation is more efficient. If you find these results stunning, check out what it is done for the Global Trajectory Optimisation Competition (GTOC), or any complex trajectory optimisation ;)

[@ceoyoyo]

There are several auto differentiating math libraries now, so differential programming is exploding. You don't have to write a bunch of math code and compute a bunch of gradients anymore, you just install TensorFlow or PyTorch, set up your problem and run it.

[@lorenzoa.ricciardi4264]

That's true, I was commenting on what I think the actual contribution of the paper is. At the moment I'm very curious about the capabilities of Julia, which pretty much has native automatic differentiation of any code, and it's very fast. But as you say, there are many libraries to do it in several languages

[@tracyh5751]

@@lorenzoa.ricciardi4264 It may have been an academic concept decades ago, but now it is technological reality. That's what's cool.

[@lorenzoa.ricciardi4264]

@@tracyh5751 Not at all, it was a technological reality already decades ago. There was already running code in the late 70s for space/aerospace applications, and results for some relatively simple (but still technologically relevant) problems were known even way before that: Robert Goddard solved the problem arising from the maximisation of the altitude of a sounding rocket early in the previous century, and rocketry was born. There are even older examples. Simply, they didn't have computers back then. The real difference is that today we have way more powerful machines at a much lower price and way easier to use software.

[@EnginAtik]

It looks like they interpret the weights in the neural net as control variables and apply Dynamic Programming. I saw a blurb in the article that they use “source code differentiation” and “just in time compilation.” I interpreted this as a method to overcome “curse of dimensionality” a problem arises in optimizing many parameters. Otherwise this looks very much like an application of Hamilton-Jacobi-Bellman although the article does not mention these names.

[@afterarrival8169]

This is like doctor strange looking at all possible outcomes and then going with the one that works. Brilliant!

[@kellynolen498]

I would like it for magic and combat in games so it can make you accomplish your actions and make it look like that would realy happen instead of a disconnect or as a setup for cut scene cinamatics you get hit and happen to fall in the right place without feeling forced More detail use headshot skill that acually flies and hits head properly

[@lusheng2346]

this channel is the best

[@xl000]

bestest

[@irok1]

Mafia boss man has spoken

[@zakuro8532]

Worstn'tst

[@Drachensslay]

Wow. Can this be used to predict how perturbations to a flow of real fluid can be manipulated in some way? (Say use a camera to track a droplet of ink or dust in water and learn to make a simple shape or pattern it with a controllable piezo)

[@sssfsfdfsdsdffsfsdf4]

It could mabye be used to aide in automation and manufacturing. Tell it that we have this robotic arms etc, and this is what we want our end product to be. Optimise for quickest cycle time etc. This seems like it is going to have real uses in the future for any number of things.

[@TheNewton]

Possibly in a large volume of still water, otherwise too many factors including the things making the perturbations

[@johnviljoen271]

Probably yes, but due to real flows non- linearities you would need to redo your droplet test for every example, so you'd basically just be doing an experiment for every situation which means you don't need to simulate like this. I think maybe lol

[@morpher728]

This could be made more complicated by predicting how quantum particles act or predict the future

[@morpher728]

It can predict space launches and the paths of astral entities

[@AaronRotenberg]

I wasn't _that_ impressed until you showed the adversarial attack with the water ripples. Then my jaw just dropped.

[@chaselewis5372]

Yeah having a way to actually generate adversarial attacks is actually truly amazing. This potentially could make traditional AI more robust.

[@mostafabalboul3966]

Where's the adversarial attack with ripples? I'm not quite seeing it here.

[@Mactakun]

Mostafa Balboul they made it so a ripple on an image would make another image recognization software would see the squirrel with the ripple as a goldfish instead of a squirrel.

[@mostafabalboul3966]

@@Mactakun Oh, THAT'S an adversarial attack! That's cool! My apologies, I didn't know the terminology beforehand; thanks for explaining!

[@bluezz5002]

@@Mactakun i dont understand the implication, did the AI think it was a goldfish because the ripple looked like water?

[@Zen-rl5pv]

channel is as underrated and powerful as the papers that are mentioned on the channel! always love the content, keep it coming professor!

[@Speak_Out_and_Remove_All_Doubt]

Hold onto your papers, here's another video!

[@WestOfEarth]

It would be interesting to apply this toward cosmology. We have a fairly good understand of the state of the universe as it exists today. With this program, you could input parameters for the Big Bang and inflation to see which original conditions most accurately describe what the Universe is like today.

[@ixion2001kx76]

Uses: designing of trajectories through chaotic systems, for instance, zero energy spacecraft trajectories through multi body systems (interplanetary transport network).

[@turkergoktas777]

After seeing this all i can imagine is that how easy neutron scattering is become in reactor physics damn the optimization

[@blackshard641]

I thought almost the same thing. A couple years ago I remarked that AI is advancing so quickly now, it is likely to be an integral part of the next huge advance in physics. I wouldn't be surprised if in 5-10 years, it's an AI that solves Quantum Gravity.

[@TheArtikae]

Tim Haldane then we get to spend the next century studying the ai to try to figure out how the hell it works

[@xGOKOPx]

@@TheArtikae If we get to the point where AI improves itself without our control then we're out of loop; we'll never catch up

[@buckrogers5331]

Incredible! Pls keep these papers coming!

[@drmilkweed]

I think the most impressive part of this paper is the sheer speed at which this rendering is accomplished. Not only can it iterate quickly (200 iterations in 120 seconds!) but it approaches the target very quickly. That Tachi-from-checkerboard render was great at 10 iterations and basically perfect after 100. This is really mind blowing stuff, thanks for sharing!

[@nononono3421]

Useful in systems where you can afford a certain level of imprecision and where you want to limit the number of interventions (energy consumption) necessary to reach an objective.

[@JKKnudsen]

Ooo, fun! Predicative hydraulic/ferro-fluid suspension, smart window screens, soft robotics, non-planar 3d printing, real-time heat transfer models, are some just from the top of my head. But there are so many possibilities. Keep the videos coming!

[@DriesduPreez]

CIA: DiffTaichi, we need to naturalize a target with just a paperclip and a rubber band. DiffTaichi: Hold your papers.

[@PythonPlusPlus]

Dries du Preez this needs more likes

[@TheNewton]

3:02 forming Ying-yang is has a similar look to the art technique Ebru paper marbeling. Ebru uses ink suspended on water and tools (like a human hair) are dragged through it to create designs

[@nickclarkart]

The implications of this for all levels of 3D asset creation are insane, but what is really exciting to me is the combination of the Differentiable rendering and physics. It makes me dream of a world in the not too distant future (maybe houdini 23) where i can sculpt or model a basic form, bring in a photo of the thing i am creating. The differentiable rendering does it's thing and then uses a knowledge base of physically based materials. It then assigns the material properties (wood, metal, glass) to the object based on it's Reflectivity etc, and applies the physics properties , and pushes it back to the 3D software. you then have a photogrammetically correct hi res model that has not only the visual properties of the material but the physical properties like mass and density as well. all adjustable and able to be worked on further. And then I can do it all in VR. SIgn me up! Luv the videos, thanks for keeping me informed! :)

[@hdswashere]

It felt like basic background information was missing from this. Most of the video was spent singing praises over the difficulty of various tasks done by this paper. Yes, it's impressive. What does this do more broadly? What novel technique is applied to perform these feats? A quick and direct explanation of what differentiability refers to would have helped. It would make the link with the previous paper clearer. Overall, there was major hype here, but not as much substance.

[@ViiKZzz]

The jello is few lines of code because it’s written using language and framework that researchers developed. A classified in PyTorch is few lines of code for the similar reason.

[@just_peace]

I'm no scientist by any means, but all of your videos are still so fun to watch. The kind of science popularization KZbin needs! "Two Minute Math Papers" and "Two Minute Psychology Papers" would make great channels. They would probably require a little bit more than two minutes though haha Thank you Two Minute Papers!

[@Shreya-fv8xl]

Love the way he says "hello scholars"

[@viibridges]

feel myself being flattered and complimented. haha

[@kryptoniterazor]

Impressive. The very first example you showed has quite obvious utility a a tool to help 3d asset creation - the challenge for game makers and fx artists is always to take some concept sketch or photo and create a textured 3d model of it. This tool could massively simplify that process if it could use noisier inputs and export standard formats.

[@wilkins67890]

Really like this paper has a useful effect while still being simple and easy to use and understand. One major use I can think of for this is perhaps using it as a way to make an AI able to take far more information in when looking at high FPS and/or high resolution inputs without a lot of shortcuts that could remove information. For example it could be used to know the outcome of a interaction that triggers a set outcome several seconds before a "Normal" AI does thus allowing it to ignore certain parts of the input until after it knows the outcome is resolved. One variable that could be used for this the average time it takes to resolves said outcome. This could drastically increase the speed in which AIs can respond to other inputs and perhaps even have a system where a primary AI uses this method and is less precise due to the removal of what it thinks is "Unnecessary" and a secondary AI which uses more traditional methods that are more precise even if they are slower.

[@perihelion7798]

This gives us a very important window into our perceived reality. I shows us that our 'reality' does not really exist, but is the result of our mental iterations.

[@SuperGastrocnemius]

Awesome channel Karoly, it's the best source for staying up-to-date with all the cool stuff happening in machine learning. And a great source of inspiration for new projects too! :)

[@u.martin6917]

That's mind-blowing. That's how much life evolved movement and form patterns, through millions of iterations tested against the cruel environment where only the fitted made it.

[@imjody]

Excellent news! Thank you very much for sharing!

[@zachb1706]

I love when you say “hold onto your papers”

[@jellyfishmachinist]

This could be used to quickly program self assembling machines whose different shapes perform different functions from nanoscale on up. Amazing.

[@moustholmes]

OMG this could not have come at a better time! well i have just found my bachelor thesis. thank you

[@realcygnus]

Heavy Duty & amazing stuff is right. It has only fairly recently dawned on me just how effective & especially how incredibly accessible ML has become.

[@z3d12]

Nice paper! Just wanted to point out that it seems to work well on continuous systems, the billiard balls are setup close together and function like a continuous domain, the possible pairwise collisions are small because they're so close together. For example if billiard balls where more spread apart and one had to do a particular sequence of collisions to get some ball in the pocket, I don't think this technique would scale since there would be a lot more possibilities of collisions to consider. Anyway, nice paper, and nice video!

[@zeitgeisttv5312]

I will not give up on math and machine learning. These videos are amazing, informative, and awe inspiring

[@dennisrichards2540]

3:36 . . . im still seeing 3 squirrels

[@Katatonya]

but the neural network detecting what's there sees a goldfish

[@chaselewis5372]

It's called an 'adversarial attack'. It is one of the reasons that AI is still confined to very specific domains. AI tends to wig out and return false results for some inputs that are CLEARLY wrong to us humans. Having a way to generate them is actually insane.

[@getsideways7257]

@@chaselewis5372 And it's getting worse. You can create a "master fingerprint" that unlocks biometric devices...

[@revimfadli4666]

@@chaselewis5372 generate them without having the networks themselves to backprop with, you mean?

[@rampage14x13]

I was lucky enough to attend a seminar last semester by Yuanming who is one of the authors of this paper. Nice to see it being covered on here!

[@rampage14x13]

Wrote this halfway haha maybe I should’ve waited until the end..

[@Timsturbs]

if it can be applied to solve some problems and then be applied to the end code to optimize it as much as possible.. or even applied to optimize itself.. thats just amazing we're getting close to answer to life, the universe and everything

[@sieyk]

I felt unreasonably impressed by the jello simulation. That was astounding.

[@BatBeardGames]

Oh man, the future of live rendering of anything is so bright its like a supernovae.

[@Noah_AWICB]

Crazily enough my dad called me over a few minutes ago to tell me about a talk he's doing about AI and machine learning and stuff that can find cracks in pavements, hazards, cool stuff like that

[@ayandas874]

Noah Edwards what a cool dad you have.

[@souslicer]

I bet you wish you were as good as your dad

[@fungussa]

Extraordinary! Thanks for the excellent videos

[@cvspvr]

man, this guy keeps coming out with amazing papers! i wonder hu he is

[@nathangamble125]

3:38 tbh, the squirrel's head on the right does actually look a bit like a goldfish's head.

[@ShinsekaiAcademy]

only if you're as smart as an AI LOL

[@Theminecraftian772]

This is amazing!!! I can't wait to put this in a drone to have it follow me with a camera so that A: I'm always in shot, B: The view is smooth, and C: It properly frames me in shot. Or, couple this with a speech interpreter and synthesizer, hook it up to some smart devices, and Bingo Bango Bongo, I have a Jarvis. I think this is the closest thing we have right now to AGI, and I'm super excited. I'm just waiting for some open source/spyware free AI technology(that I know how to use) so I can have my own smart assistant without having to worry about data on my life being sent to a corporation that wants to sell me things and make money.

[@combcomclrlsr]

Reminds me (just a little) of the FFTW. FFT is sped up greatly through the use of mathmatical identities and code transformations.

[@matthewwerblud1366]

Wow! Absolutely amazing, what a great way to end the day!

[@6ix157]

that's actually crazy... can't wait to see what the future holds.

[@Igneshto]

I sometime go around the house randomly shouting your FULL NAME at my wife... KarolJolnayfaher!!!! (I know that's not how it's written but that how I shout it) With the accent on point I must add 😁😁😁 Awesome video as always, it's a pleasure listening to you... Especially! ESPECIALLY!!! "What a time to be alive". Love that mentality. Keep being Awesome.

[@albertoviceconti8579]

That looks like a game changer. SO powerful

[@purpleghost106]

You got to the pool balls, and I was like use it for SPAAAAACE! If this gets accurate enough, and if the physics of it are solid enough, it could be highly useful for estimating trajectories of known objects, such as asteroids in the belt between us and jupiter, to plan flight paths for missions. Already we have simulators for that of course, but if applied correctly, this could make it easier by a lot. -- Everything you show can help make awesome games, but also much of it could be useful For Science! :D

[@georgiostsirtsidis1125]

This gave me legit goosebumps. Loved it!

[@TheGrinningSkull]

I love your phrase "What a time to be alive". It so is!

[@xf1961]

I see a possibility of this program being applied to fluid dynamic studies. I haven't read though the details yet but it seems that if you know the physical attributes that you are trying to achieve and the general shape you will need you can allow the program to make many iterations to find the perfect shape for each use case. Obviously there is still a lot of development that will be needed like first off making it 3D and as well being able to set constraints of things that can not change but everything else is up for grabs for the tool to work with.

[@GodOfReality]

Use this in a meta way: design the game such that you have an idea of what the outcome of the game should be: how it should play and what the final package of the design should look like, and then use this differentiable system to figure out what mechanics and mechanics-adjustments need to be created to get it just right. Plausibly this could be applied to currently existing video games like fighting games, and within a strict framework, it would be possible to give the algorithm free reign to end up with a final version that has something like "no character has a better than 55% winrate against any other character".

[@SprDrumio64]

This is going to be immensely useful for ballistic forensics. Just think: take a picture of a crime scene, map it out precisely, and you basically have all you need besides listing materials to know where and how the bullets came from

[@chaoticprogramming]

If you could bake the simulation and make it with particles that move randomly and then tell it to remove any particles that were ever out of the boundary you might get a pretty good looking one.

[@trex70]

The movment Iteration is cool. Maybe usefull for Maschines on other planets to move on any kind of surface (would be nice to see it handle a frozen surface)

[@Lumitex]

Thank you for your effort. After few evenings of these short videos I've finally got the point of ML/AI industry. Basically you just give it a certain task which it was trained for, and it gives out result which could take few years to complete by high qualified team of people. Similar paradigm shift to how Operating Systems changed manual calculations. Sad thing is that people will find ways to abuse it, so we will have to deal with whole new kind of problems. Tech is neutral, but people behind it.. not always, but questionable

[@mayorc]

I came here for "What a time to be alive!", and "I see you next time!" :)

[@moth.monster]

Creating adversarial images for neural networks using fluid simulation sure isn't what i expected to see today, but I'm into it.

[@manyirons]

Wow! Okay, here's a suggestion: We get a lot of snow here in Canada, and it can get in the way. For many years I wanted to simulate how it fell, to match reality, and then use the simulation to help predict placement of snow fences that would best deflect snow from settling in undesirable locations. Could your techniques be effective at this task?

[@jonesbbq307]

This is what I have always dreamed about: no matter how impossible it might seem, there is must a way to do it. Now this algorithm can find that way.

[@garrettrichards9028]

Wow! This is amazing! I'm currently working on a hobby project and I bet that if I'm able to implement this it'll be much better. I'm trying to build a simulation where it simulates the trisolarian's solar system from the book The Three Body Problem. I was able to write a simple physics engine for the solar system to work in however the issue is that I don't have the exact conditions of the solar system so I can't implement it. I tried using a genetic algorithm to guess what the solar system may look like however the agents cheated my fitness function! The second issue is that it takes way to long to train. I know this paper won't help with the fitness function but I do hope that it'll make things much faster and hopefully would have better control with the chaos in the simulation. I can't wait to try it!

[@Vontux]

Heh, the neural network and the goldfish misidentification, I think I get what that neural network looks for, it looks for a head and eye in an image distorted by water ripples. One of the few times I understand an ANNs "thought" process...

[@ceskehry]

You are very talented, thank you for sharing with us.

[@fossil98]

why doesnt this system get stuck in local minimums?

[@Sphinxrave-dev]

> We present DiffTaichi, a new differentiable programming language tailored for building high-performance differentiable physical simulators. Based on an imperative programming language, DiffTaichi generates gradients of simulation steps using source code transformations that preserve arithmetic intensity and parallelism. A light-weight tape is used to record the whole simulation program structure and replay the gradient kernels in a reversed order, for end-to-end backpropagation. We demonstrate the performance and productivity of our language in gradient-based learning and optimization tasks on 10 different physical simulators. For example, a differentiable elastic object simulator written in our language is 4.2x shorter than the hand-engineered CUDA version yet runs as fast, and is 188x faster than the TensorFlow implementation. Using our differentiable programs, neural network controllers are typically optimized within only tens of iterations. It's a different way of representing a Neural Net that makes backprop way cheaper than contemporary frameworks. I think the magic in not getting local minimum'd is the way it generates the "source code transformations". I wonder how they define "arithmetic intensity / parallelism" here.

[@neurophilosophers994]

Had the same question algebraic intensity kind of makes sense but parallelism ? Between node layers ?

[@Sphinxrave-dev]

@@ebaziuk kzbin.info/www/bejne/kGLbp3SQq8psl6c they go into it in their paper video, a little. I'm not sure how this absolves them from every single local minimums, e.g. the adversarial attack which is a black box... how is that differentiable? But it does make a bit of sense for the physics-driven ones.

[@HunteronX]

Not sure if this is completely wrong, but: I assume because the physical laws apply consistently throughout the simulated space. No glitches, so to speak. For example, Hooke's Law as applied to the springs in the paper. The unreasonable effectiveness of mathematics... For the adversarial attack on the VGG network, it may just be that the water ripples induce enough noise (some contradictory non-linear effect in the arrangement of pixels, which its neurons cannot disentangle) - maybe a bit like attempting classification of a circular distribution of points with a line.

[@chaselewis5372]

In general I bet it can, but given the source code transformations it guarantees your derivatives are actually SMOOTH. You get a lot more local minimums due to discontinuities in your domain then you would expect. The paper and the video describing the paper from the creator go over it a bit. Essentially with accurate & smooth derivatives the convergence is MUCH quicker and more accurate.

[@mauz791]

Thank you for simplifying this!

[@AhmedSalah-vy8vn]

i can see a lot of potential from this method in structural optimization.

[@jumbi5333]

My mind is blown! This is straight up controlling the butterfly effect! I wonder if this could have weather applications later on down the line

[@StormwaterIsOneWord]

I have no idea what I just watched, but I’m gonna subscribe for more intriguing topics.

[@dernicolas6281]

please explain what differentiable physics means in this context - I still have no clue what is differentiable here. Lovely pictures though...

[@ceoyoyo]

Take the billiards example. You can specify the speed and angle of the white ball, the physics simulation grinds through, and the blue ball ends up some distance from the black spot. You then find the gradient (multidimensional derivative) of the initial speed and angle parameters with respect to the distance from the black spot. That gradient is an estimate of how much you have to change the speed and angle to precisely hit the black spot. So you apply some fraction of that gradient and try again. You keep repeating, getting closer and closer. That kind of optimization is as old as computers, but there are several software packages now that make it really easy.

[@dernicolas6281]

@@ceoyoyo yes - of course you can do that - but there's really nothing new about that. So nothing to publish... okok - I will read the paper when I've got some time.

[@ITR]

@@dernicolas6281 I think he's claiming that they made an algorithm that is a lot faster than older ones, in which case it's definitely something to publish, lol

[@maverick9300]

It seems that this technique is the basis for generating instructions for any end goal in a given simulated environment. If the environment is accurate enough to reality, the possibilities are endlesssssssss.

[@andrewkelley7062]

It would be useful to train a algorithm like this to be used as a slicer for a 3d printer. All you would need would be a few test prints to calibrate the printer then the filament. Then it could run simulations until it produced a gcode that would always work and always have the properties you set it for. Within reason of course. Then after a few generations and a module on the printer along with a few cameras, as well as some parts. You might be able to get to the point where you can do mold less injection molding.

[@conoroneill3826]

This is fantastic. Please create lots of elastic gui objects, give them a football and make goals their goal.

[@maidpretty]

Wow, this wizardry has a potential of making a great impact in lots of areas.

[@Mikey-mike]

Cool. Interesting. Thank you for sharing.

[@jurian0101]

This tech kinda enables the future. For example, magnetic confinement nuclear fusion could benefit by using this algorithm to tell where, when, and exactly how much is to correct for stable plasma current. Generally, in real time manipulate all sort of complex systems to our wish on the fly (given enough calculating power).

[@NehadHirmiz]

It would be amazing to see this method deployed to look into the self-assembly in soft condensed matter. Interesting physics can be investigated from several important processes such as protein aggregations (pore formers, ion channels, etc) diblock copolymer formation, multiphase behaviour,

[@lu-dx6oh]

Any sufficiently advanced technology is indistinguishable from magic. Arthur Clarke

[@ragnabob]

Excellent as always!!!

[@Ar3Ar3]

Such a great paper （⊙.⊙） That conversion of chess board to yin and yang is awesome.

[@StoneGear]

Every video of Two Minutes Papers makes me want to live more.

[@c.t.d.r.a.8295]

Nice one, as always!

[@kkkaran786]

What a time to be alive!

[@PeterAuto1]

The last one looks great for creating new example data for training

[@purpleghost106]

Imagine if you could actually get this to run *in* a game, that would be amazing! Hypothetically then it could be used dynamically! Imagine if a character model in a game didn't just perform a preset animation! A player model, or an NPC could approach an environment based on the physics, meaning that if you give them a lump of rock instead of a climbing animation they could actually try to scale it. Also those rocks could break if there are weak-points when they're generated and it wouldn't have to be like a planned 'this always breaks when you climb it' story event. It could mean a voxel-ish sandbox with superbly immersive adventure elements. Also in games with fighting imagine an AI shifting their posture to more accurately target your characters location based on the key constraints of the physics. (ex: their leg can't stretch, meaning they can miss if the timing is off, but they won't just kick the same way each time instead they'll shift however their skeleton allows a kick.) There's just so many possibilities, so many! I agree, what a time to be alive!

[@trejohnson7677]

The security sector will absolutely love the differentiable water renderer