Placeholder Image

字幕表 動画を再生する

  • Hi, everybody.

  • And welcome to this session where we're gonna talk about breakthroughs in machine learning.

  • I'm Laurence Maroney.

  • I'm a developer advocate.

  • That's Google of working on tensorflow with the Google brain team.

  • We're here today to talk about the revolution that's going on in machine learning and how that revolution is transformative.

  • Now I come from a software development background.

  • Any software developers here, given that it's I Oh, sure, on this, this is this transformation.

  • This revolution is, particularly from a developers perspective, is really, really cool because it's giving us a whole new set of tools that we can use to build scenarios and to build solutions for problems that may have been too complex to even consider prior to this.

  • It's also leading to massive advances in our understanding of things like the universe around us.

  • It's opening up new fields and arts on it's impacting and revolutionizing things such as health care on so many more things.

  • So should we take a look at some of these so first of all, astronomy at school?

  • I studied physics.

  • I wasn't the com side part since I'm a physics and astronomy geek on it wasn't that long ago when we learned how to discover what, how new planets around other stars in our galaxy, the way that we discovered it was that sometimes would observe like a little wobble in the star.

  • And that meant that there was a very large planet like Jupiter, size or even bigger, orbiting that star very closely and causing a wobble because of the gravitational attraction.

  • But, of course, the kind of planets we want to find out the small, rocky ones like Arthur Mars, where you know there's a chance of finding life on these planets on DDE.

  • Finding those and discovering those was very, very difficult to do because small ones closer to Starr, you just wouldn't see.

  • But there's with research that's been going on in the Kepler mission.

  • They've actually recently discovered this planet called Kepler 90 i by sifting through data and building models for using machine learning and using tensorflow on Kepler 90.

  • Eye is actually much closer to its host star than Earth is so that its orbit is only 14 days instead of our 365 and 1/4 in a bitch on.

  • Not only that which I find really cool that they didn't just find.

  • This is a single planet around that star.

  • They've actually mapped and model the entire solar system of eight planets that are there.

  • So these are some of the advances.

  • It's to me.

  • I find it's just a wonderful time to be alive because technology is enabling us to discover these great new things.

  • And even closer to home, we've also discovered that looking at scans of the human eye, as you would have seen in the keynote, you know, with machine learning trained models on this, we've been able to discover things such as blood pressure predictions or being able to assess a person's risk of a heart attack or a stroke.

  • Now, just imagine if this screening can be done on a small mobile phone, how profound is the effects going to be?

  • Suddenly, the whole world is going to be able to access easy, rapid, affordable.

  • A noninvasive screening for things such as heart disease is it'll be saving many lives but also be improving the quality of many, many more lives.

  • Now, these are just a few of the breakthroughs and advances that have been made because of tensorflow and tensorflow.

  • We've been working hard with the community with all of you to make this a machine learning platform for everybody.

  • So today and when we want to share a few of the new advances that have been working on this.

  • So including will be looking at robots on Vincent's gonna come out in a few moments to show us robots that learn and some of the work that they've been doing to improve how robots learn.

  • And then Debbie is going to be from nurse.

  • She's gonna be showing us cosmology advancements on including showing how building a simulation of the entire universe will help us understand the nature of the unknowns in our universe, like dark matter and dark energy.

  • But first of all, I would love to welcome from the magenta team we have dug.

  • Who's the principal scientists, Doug.

  • Thanks, Florence.

  • Thanks, Doug.

  • Thank you very much.

  • All right.

  • Day three.

  • We're getting there, Everybody.

  • I'm Doug.

  • I am a research scientist at Google, working on a project called Magenta.

  • And so before we talk about modeling the entire known universe, so we talk about robots.

  • I want to talk to you a little bit about music and art and how to use a machine learning potentially for expressive purposes.

  • So I want to talk first about a drawing project called Sketch Are n n where we trained a neural network to do something as important as draw the pig that you see on the right there.

  • And I want to use this as an example, actually highlight a few, I think important machine learning concepts that we're finding to be crucial for using machine learning in the context of art and music.

  • So let's dive in.

  • It's gonna get a little technical, but hopefully to be fun for you, all we're gonna do is try to learn to draw, not by generating pixels, but actually by generating pen strokes.

  • And I think this is a very interesting representation to use because it's very close to what we do when we draw so specifically, we're gonna take the data from the very popular quick draw game playing Pictionary against machine learning algorithm.

  • At that was captured his Delta X delta y movements of the pen.

  • We also know when the pen is put down on the page and when the pen it's lifted up and we're gonna treat that as our training domain.

  • One thing what I would notices that are observed is that we didn't necessarily need a lot of this data.

  • What's nice about the data is that it fits the creative process.

  • It's closer to drawing, I argue.

  • Then pixels are to drawing.

  • It's actually modeling the movement of the pen.

  • Now what we're gonna do with these drawings is we're gonna push them through something called an auto encoder.

  • What you're seeing on the left, the encoder Networks job is to take those strokes of that cat and encode them in some way so that they could be stored as a latent vector, the yellow box in the middle.

  • The job of the decoder is to decode that late in vector back into a generated sketch and the very important point.

  • In fact, the only point that you really need to take away from this talk is that that late in vector is worth everything to us.

  • First, it's smaller in size than the encoded or decoded drawing, so it can't memorize everything.

  • And because it can't memorize, we actually get some nice effects.

  • For example, you might notice if you look carefully that the cat on the left, which is actual data and has been pushed through the trained model and decoded, is not the same as the cat on the right, right?

  • The cat on the left has five whiskers, but the model regenerated the sketch with six whiskers.

  • Why?

  • Because that's what it usually sees.

  • Six whiskers is general.

  • It's normal to the model where it's five.

  • Whiskers is hard for the model to make sense of.

  • So this idea of having a tight, low dimensional representation this late in vector that's been trained on lots of data.

  • The goal is that this model might learn to find some of the generalities in a drawing, learn general strategies for creating something.

  • So here's an example of starting each of the four corners with a drawing done by a human.

  • David, the first author, and those are encoded in the corners, and now we just move linearly around the space, not the space of the strokes but the space of the Layton vector.

  • And if you look closely, what I think you'll see is that the movements and the changes from these faces say, from left to right, actually quite smooth.

  • The model has dreamt up all of those faces in the middle.

  • Yet to my eye, they really do kind of fill the space of possible drawings.

  • Finally, as I pointed out with the cat whiskers, these models generalized, not memorize.

  • It's not that interesting to memorize a drawing.

  • It's much more interesting to learn general strategies for drawing.

  • And so we see that with the 5 to 6 Mr Cat, I think more interestingly, and I think it's also suggestive.

  • We also see this with doing something like taking a model that's only seen pigs and giving it a picture of a truck.

  • And what's that model going to do?

  • It's gonna find a pig truck because that's all it knows about right?

  • And if that seems silly, which I grant, it is in your own mind.

  • Think about how how hard it would be, at least for me.

  • If someone says draw a truck that looks like a pig, it's kind of hard to make that transformation, and these models do it.

  • Finally, they get paid to do this.

  • I just want to point that out as an aside, so it's kind of nice.

  • I said that last year.

  • It's still true.

  • Um, okay, so these late in space analogies, another example of what's happening in these late in space is obviously, if you add and subtract pen strokes, you're not gonna get far with making something that's that's recognizable.

  • But if you have a look at the light and space analogies, we take the late in vector for a cat head and we add a pig body and we subtract the pig head.

  • And of course, it stands to reason that you should get a cat body and we could do the same thing in reverse.

  • And this is real data.

  • This actually works, and the reason I mention it is it shows that the late in space models are learning some of the geometric relations between the forms that people draw.

  • I'm going to switch gears now and move from, uh, drawing to music and talk a little bit about a model called Incense, which is a neural network synthesizer that takes audio and learns to generalize in the space of music you may have seen from the beginning of a Iot with bathing that has been put into a hardware unit called Instant Super how many people have heard of hand since super?

  • How many people want an instant?

  • Super good.

  • Okay, well, that's possible, as you know.

  • Um, okay, so I want for those of you that didn't see the opening, I have a short version of the making of the instant.

  • Super like to roll that now to give you guys a better idea of what this model's up to.

  • Um, let's let's roll it.

  • That's, like, wild.

  • There's a flu.

  • Here's a snare.

  • Now I just feel like attending a corner.

  • What could be new possibility?

  • It could generate a sound that my inspire ism.

  • The fun part is like, Even though you think you know what you're doing, there's some weird interaction happening.

  • Can give you something totally unexpected.

  • Wait, why did that happen that way?

  • Okay, So what you see here, by the way, the last person with the long hair was Jesse Angle, who was the main scientist on the instant project.

  • This grid that you're seeing this, uh, square where you can move around the space is exactly the same idea as we saw with those faces.

  • So the idea that you're moving around the late in space, and you're able to discover sounds that hopefully have some similarity and because they're made up of learning what makes humans how sound works for us in the same way as a pig truck gives us maybe some new ideas about how sound works.

  • And, as you probably know, you can make these yourself, which I think is my mind.

  • My favorite part about the Instant Super project is that this is open source.

  • Get hub for those of you who are makers and like to tinker, please give it a shot.

  • If not, we'll see some coming available from tons of people who are building them on the room.

  • So I want to keep going with music, but I want to move away from audio and I want to move now.

  • Thio Musical scores, musical notes something that you know, think of last night with justice driving a sequencer and talk about basically the same idea, which is, can we learn a late in space where we can move around what's possible in in in a musical note or a musical score?

  • Rather so what you see here is some three part musical thing on the top and someone part musical thing on the bottom and then finding in a late in space something that's in between, okay?

  • And now I put the faces underneath this.

  • What you're looking at now is a representation of a musical drum score where time is passing left to right.

  • And what we're going to see is we're gonna start.

  • I'm gonna play this for you.

  • It's a little bit long, so I want to set this up.

  • We're gonna start with a drum beat one measure of drums, and we're gonna end with one measure of drums and you're gonna hear those.

  • First you're gonna hear A and B, and then you're going to hear this late in space model.

  • Try to figure out how to get from a to B, and everything in between is made up by the model in exactly the same way that the faces in the middle are made up by the model.

  • So as you're listening, basically, listen for whether it makes musical sense or not, that the intermediate drums let's give it a role.

  • So you have it moving right along.

  • It turns out, take a look at this command.

  • Um, this make sense to some of you may be, we were surprised to learn, after a year of doing magenta, that this is not the right way to work with musicians and artists.

  • I know I laughed too, but we really thought he was a great idea.

  • Guys, it's like taste this into terminal and they're like, what's terminal?

  • And then, you know you're in trouble, right?

  • OK, so, um, we've we've moved quite a bit towards trying to build tools that musicians can use.

  • This is a drum machine, actually, that you can play with online built around tensorflow dot Js and I have a short clip of this being used.

  • What you're going to see is all the red is from you.

  • As a musician, you can play around with it and then the blue is generated by the models.

  • So let's give this a roll.

  • This is quite a bit shorter.

  • So this is available for you as a code pen which allows you to play a round of the HTML and the CSS and the JavaScript and really amazing a huge shout out to Tero Parviainen who did this.

  • He grabbed one of our train magenta models and he used tensorflow dot Js and he hacked a bunch of code to make it work.

  • And he put it out on Twitter and we had no idea this was happening.

  • And then we reached out to my research A month for terror.

  • You're my hero.

  • This is awesome.

  • And he's like, Oh, you guys care about this?

  • Can Of course we care about this.

  • This is our dream to have people not just as playing with this technology.

  • So I love it that we've gotten there.

  • So part of what I want to talk about today actually close with, We've cleaned up a lot of the code.

  • In fact, Tero helped.

  • And we've Now we're able to introduce Magenta that J s, which is very tightly integrated with tensorflow dot Js.

  • And it allows you, for example, to grab a checkpoint in model and set up a player and start sampling from it.

  • So in three lines of code, you can set up a little drum machine or music sequencer, and we're also doing the same thing with sketch aren't and so we have the art side as well.

  • Um, and we've seen a lot of demos driven by this a lot of really interesting work, both by Googlers and by people from the outside.

  • And I think it highly lines well with what we're doing in Magenta.

  • So too close.

  • We're doing research in generative models were working to engage with musicians and artists.

  • Very happy to see the JavaScript stuff come along.

  • Which is really seems to be the language for that, hoping to see better tools come and heavy engagement with the open source community.

  • If you wanna learn more, please visit Judah Coast last magenta.

  • Also, you can follow my Twitter account.

  • I post regular updates and try to be a connector for that.

  • So that's what I have for you.

  • And now I'd like to switch gears and go to robots.

  • Very exciting with my colleague from Google Brain.

  • Vincent.

  • Anouk.

  • Thank you very much things, Doug.

  • So my name's been sent and I lead the Brain Robotics research team, the robotics research team at Google.

  • We when you think about robots, you may think about precision and control.

  • You may think about robots, you know, leaving factories.

  • They've got one very specific job to do, and they gotta do it over and over again.

  • But as you saw in the keynote earlier, more and more robots are about people write their self driving cars that are driving in our streets, interacting with people.

  • They essentially now live in our world, not their world.

  • And so they really have to adapt and perceive the world around them and learn how to operate in this human centric environment.

  • Right?

  • So how do we get robots to learn instead of having to program them?

  • Um, this is what we've been embarking on.

  • And it turns out we can get robots to learn.

  • It takes a lot of robots.

  • It takes a lot of time.

  • And but we can actually improve on this if we teach robots hard to behave collaboratively.

  • So this is an example of ah team of robots that are learning together how to do a very simple task like grasping objects right at the for the beginning.

  • They have no idea what they're doing there.

  • Try and try and try.

  • And sometimes they will grass something every time they grass.

  • Something would give them a reward and over time to get better and better at it.

  • Of course, we used the planning for this um, basically have a convolution all network that maps those images that the robot C of the work space in front of them to actions and possible actions.

  • And this collective learning of robots enables us to get to levels of performance that we haven't seen before.

  • But he takes a lot of robots.

  • And in fact, you know, this is Google.

  • We would much rather use lots of computers if we could instead of lots of robots.

  • So the question becomes, Could we actually use a lot of rope simulated robots, virtual robots to do this kind of task and teach those robots to perform tasks?

  • And would it actually matter in the real world, would they?

  • What would they learn in simulation, actually apply to real tasks?

  • Um, and it turns out the key to making this work is to learn simulations that are more and more faithful to reality.

  • So on the right here, you see what a typical simulation of a robot would look like.

  • This is a virtual robots trying to grasp objects and simulation.

  • What you see on the other side here may look like a real robots doing the same task, but in fact, it is completely simulated as well.

  • We've learned a machine learning model that maps those simulated images to real images to real looking images.

  • They're essentially indistinguishable from what a riel robot would see in the real world.

  • And by using this kind of data in a simulated environment and training assimilating model to accomplish tasks using those images, we can actually transfer that information and make it work in the real world as well.

  • So there's lots of things we can do with theis kinds of simulated robots.

  • Uh, this is Rainbow Dash, our favorite little pony.

  • And would you see it here is him taking his very first steps and our very first hop?

  • So you could say he's very good for somebody who's just starting to learn how to walk.

  • And the way we accomplished this is by having a virtual rainbow dash running in simulation.

  • We train it using deep reinforcement, learning to run around in the simulator, and then we can only basically Donald the model that we've run the simulation onto the real robots and actually make it work in the real world as well.

  • There are many ways we can scale up robotics and robotic learning.

  • In this way, one of the key ingredients turns out to be learning by itself.

  • Self supervision, self learning.

  • This is an example of, for example, what you see at the top here is somebody driving a car.

  • And, um, what we're trying to learn in this instance is the three D structure of the world, the geometry of everything.

  • What you see at the bottom here is a representation of how far things are from the car.

  • You probably are looking at avoiding obstacles and looking at other cars to not collide with them.

  • And so you want to learn about the three D geometry based on those videos?

  • The traditional way that you would do this is by involving, for example, or three D camera or a lighter or something that gives you a sense of depth here.

  • We're gonna do none of that.

  • We're going to simply look at the video and learn directly from the video the three D structure of the world.

  • And the way to do this is to look at the video and try to predict the future of this video.

  • You can imagine that if you actually understand the three D geometry of the world.

  • You can do a pretty good job of predicting what's going to happen next in a video.

  • So we're going to use that signal that tells us how well we're doing at predicting the future, to learn where the three geometry of the world looks like.

  • So, at the end of the day, what we end up with is yet another big convolution, all network that maps what you see at the top to what you see at the bottom without involving any three D camera or anything like that.

  • This idea of self learning are just learning without any supervision directly from the data is really, really powerful.

  • Another problem that we have when we're trying to teach robots how to do things is that we have to communicate to them what we want.

  • What we care about right and the best way you can do that is by simply showing them what you want them to perform.

  • So here's an example of one of my colleagues basically doing the robot dance and robots that is just looking at him, performing those tasks and trying to imitate visually what he's doing and What's remarkable here is that, you know, even though the robot, for example, doesn't have legs, it tries to do this crouching motion as best they can, given degrees of freedom that has available.

  • And all of this is learn entirely self supervised.

  • The way we go about this is that if you think about imitating somebody else, for example, somebody pouring a glass of water kind of coke it all relies on you being able to look at them from 1/3 party of you and picturing yourself doing the same thing from your point of view, what you would look like if you did the same seeing yourself.

  • So we collected some of this data that looks like that where you have somebody looking at somebody else do a task and you end up with those two videos off, one taken by the person doing the task and another one taken by another person.

  • And what we want to teach the robots is that those two things are actually the same thing.

  • So we're going to use again machine learning to perform this match up.

  • We're gonna have ah machine learning model days going to tell us.

  • Okay, this image on the left is actually off the same task as this image on the right.

  • And once we've learned that correspondence with lots of things we can do with this one of them is just imitation like this.

  • Imagine you have somebody pouring a glass of water.

  • The robot sees them.

  • They try to picture themselves doing the same task and try best they can to imitate what they're doing.

  • And so, using again deeper enforcement learning, we can train robots to learn those skins off activities completely based on visual observation without any programming of any kind.

  • So I won't let that well, but for me they are quite yet.

  • But it's very encouraging that we can just look, have robots that understand essentially what the nature with the fundamentals off the task is regardless off, whether they're pulling a liquid or they're pouring beads or whatever the glasses look like, or the containers all of that is abstracted, and the room would actually really understand deeply what the task is about.

  • So I'm very excited about this whole perspective on teaching robots how to learn instead of having to program them right.

  • At some point, I would want to be able to tell my Google assistant a Okay, Go, go.

  • Please go for my laundry.

  • Right.

  • And for that to happen, we're gonna have to rebuild the science of robotics from the ground up.

  • You're going to have to, um, base it on understanding and machine learning and perception.

  • And of course, we're going to have to do that.

  • Google's go with that.

  • I'm going to give the stage to Debbie.

  • Who's going to talk to us about cosmology?

  • Thank you.

  • Okay.

  • Thank you.

  • Good afternoon, everyone.

  • My name's Debbie bards.

  • I'm talking about something a little bit different from what you've heard so far.

  • S O.

  • I lead the data science engagement group.

  • That nurse a nurse is the National Energy Research Scientific Computing Center.

  • Were a supercomputing center up, But Lawrence Berkeley National Lab, just over the bay.

  • From here we are the Mission Computing Center for the Department of Energy.

  • Officer Science on what this means is that we have something like 7000.

  • Scientists are using our stupid computers to work on some of the biggest questions in science today.

  • On what I think is really cool as well is that I get to work with some of the most powerful computers on the planet.

  • One of the things that were noticing, especially the last couple of years, is we've seen that scientists are increasingly turning to deep learning in machine learning methods to solve some of these big questions that they're working on.

  • We're seeing these questions showing up in our workload on our supercomputers, So I want to focus on one particular topic.

  • Area is very close to my heart, which is cosmology, because I'm a cosmologist by training.

  • My background is in a cosmology research because I've always been interested in the really the most fundamental questions that we have in science about the nature of the universe from.

  • That's one of the most basic questions you can ask about the universe is what is it made off on dhe thes days.

  • We have a fairly good feel for how much dark energy there is in the universe, how much dark matter how much regular matter there is in the universe.

  • And there's only about 5% of regular matter, which is everything that you and I and all the stars and all the dust nor the gas and all the Galaxies out there, they're made of regular matter and that makes up a pretty tiny proportion of the contents of the universe.

  • The thing that I find really interesting is we dose just don't know what the rest of it is.

  • Dark matter.

  • We don't know what that's made off, but we see indirectly the gravitational effect.

  • It has dark energy.

  • We don't know what that is.

  • A tool that was only recently discovered about 15 years ago on dark energy.

  • Just the name that we give to an observation, which is the accelerated expansion of the universe on.

  • And this is, I think, really exciting.

  • The fact that there is so much that we have yet to discover means that they're tremendous possibilities for new ways for us to understand our universe and we are building a bigger and better telescopes were collecting data all the time, taking images and observations of the sky.

  • Thio, get more data to help us understand this because we only have one universe to observe.

  • So we need to be able to collect as much data as we can on that universe and we need to be able to extract all the information we can from our data from observations and cosmologists are increasingly turning to deep learning to extract meaning from our data.

  • And I'm gonna talk about a couple of different ways that we're doing that.

  • But first of all, I want to kind of ground this in the background of how we actually do experiments in cosmology.

  • Cosmology is not an experimental science in the way that many other physical sciences are.

  • There's not a lot we can do to experiment with the universe.

  • We can't really do much to change the nature of space time, although it would be fun if we could.

  • But instead we have to run a simulation.

  • So we run simulations in supercomputers off theoretical universe is Andi different physical models and the different parameters that control those physical models?

  • And that's how the experiment we run these simulated universes.

  • Then we compare the outputs of these simulations to our observations of the real universe around us.

  • So when we make this comparison, we typically using some statistical measure, some kind of reduced statistic like the power spectrum which is illustrating this animation here.

  • The paras spectrum is a measure of how matter is distributed throughout the universe, whether it's kind of distributed fairly evenly throughout space or whether it's clustered on small scales.

  • This is illustrated in this.

  • The images on the top of this light here, which is snapshots of a simulated universe, run in a stupid computer, and you can see the overtime gravity is pulling matter together on so that start matter and regular matter.

  • Gravity is acted upon that collapsing the matter into very typical cluster and phylum entry type structures.

  • Where is dark energy is expanding space itself, expanding the volume of the this miniature universe?

  • And so, by looking at the distribution of matter, we can start to learn something about the nature of the matter itself, how gravity is acting on mats and what dark energy is doing.

  • But as you can imagine, running these kinds of simulations is very computation, the expensive.

  • Even if you're only simulating a tiny universe, it still requires a tremendous amount of computer power.

  • Andi, we spend billions of compute hours on super computers around the world on these kinds of simulations, including the supercomputers that I work with on one of the ways that we're using deep learning is thio reduce the need for such expensive simulations similar to the previous speaker was talking about Vincent talking about when robotics were exploring using generative networks to produce in this case, this example two dimensional maps of the universe.

  • So these are two dimensional maps of the mass concentration of the universe.

  • So you can imagine the three dimensional volume collapsed into a two dimensional projection off the mass density in the universe.

  • As you're looking out at the sky when we used again which is basically fairly standard D c gan topology to produce new maps off these new mass maps based on simulations S O, this is an augmentation we're using this network.

  • Thio augments on existing simulation to produce new maps.

  • We see that it's doing a pretty good job.

  • So just by looking by eye at the generated images, they look pretty similar to the real input images, the rial simulated images.

  • But as a scientist, kind of squinting at something and saying, Oh yeah, that looks about right is not good enough.

  • What I want is to be able to quantify this, to be able to quantify how are the network is working and quantify how, like the rial images our generated images are This is, I think, where scientific data has a real advantage compared to your natural image data because, uh, scientific data of usually very often has associative statistics with it.

  • So statistics that you can use to evaluate the success of your model.

  • So in this case, we were looking at reduced statistics that describe the patterns in the maps, like the power spectra, another measures of the topology of the maps.

  • We see that not only do the maps look about right, but the statistics that contained in those maps match those from the real simulations so we can quantify the accuracy of our network on this is something that potentially could be useful for the wider, deep learning community.

  • Using scientific data that has these associative statistics could be of real interest.

  • I think too deep learning practitioners in trying to quantify how well your networks are working.

  • S o.

  • I mentioned before that this is an augmentation again that we've been working on so far.

  • It can produce new maps based on a physics model that it's already seen on.

  • We're working.

  • It's scaling this up on producing physics models that the network has never seen before.

  • So making this into a true emulator on and this will help reduce the need for these very computation the expensive simulations.

  • And I are cosmologists to explore parameter space a bit more freely, and I'd like to explore a little bit further what this network is actually learning.

  • I saw a really interesting talk this morning here was touching on this kind of thing how we can use machine learning to gain insights into the nature of the data that we're working with.

  • Eso in the work that I'm showing here we were looking at which structures in our mass maps are contributing to the model of most strongly contributing to the model by looking at a quantity called saliency and so by if you look at the map of saliency, which is the black and white image here, you can see that the peaks in the salient see Matt correspond to peaks in the mass map on dso these peaks in the mass About these are concentrations of matter.

  • And these corresponds to galaxy clusters, typically in the real universe.

  • And this isn't news to cosmologists.

  • We've known for decades.

  • That Galaxy cluster is a really good way of exploring cosmology, but the shapes of the features that this network have learned are not nice round galaxy in balls, they are irregular and they're showing some structure, and this is something that's really interesting to me.

  • There's also indications that some of the smaller mass concentrations of showing up is important features in this network, and that's perhaps a little bit unexpected.

  • So by taking this kind of introspection into the features that our network is learning, we can start to learn something about the data and get insight into some of the physical processes that are going on in our data and learn what kind of structures perhaps our most sensitive to the interplay of gravity and dark energy.

  • I think this is something that's really, uh, a real strong point of deep learning when you are allowing the network to learn features for itself.

  • Rather than imposing features doing feature engineering or telling it any particular statistics, you can allow the network to tell you something about your data that might surprise you.

  • So far that was looking at two dimensional maps, but of course the universe is not a two dimensional place.

  • At least four dimensions, perhaps many more dimensions, depending on your favorite model of string theory.

  • But we've been looking at three dimensions.

  • Is scaling this up?

  • Another, Another level.

  • One of my three dimensions are interesting from us, from a computational point of view, because in a three dimensional data volume, you're looking at three dimensional matrices, three dimensional convolutions.

  • This is something that's computational, expensive, and it's something that can run really well on a super computing architectures.

  • So a team at CMU recently demonstrated for the first time that deep learning could be used to determine the physical model of the universe from three dimensional simulations of the full matter distribution.

  • So this is the full three dimensional matter rather than a two dimensional projection of the matter.

  • Density on this work showed that the network was able to make significantly better estimates of the parameters that describe the physics off the simulated universe compared to traditional methods where you might be looking at one of these statistics, like the power spectrum on dhe, So thistle is a really nice example of how the network was able to learn what's structures in this three dimensional matter falling more important rather than just looking at statistics that we in advance thought was going to be useful.

  • So we're working on scaling this up at the moment in collaboration with nurse, UC Berkeley, Intel and Crais, who are industry partners that mask we using larger simulation volumes, even more data on dhe.

  • We're using tensorflow running on thousands of CPU notes, achieving with several petabytes petr flops of performance on Quarry, which is our flagship supercomputer.

  • But perhaps the most important part of this is that we're able to predict more physical parameters with even greater accuracy by scaling up the training of this.

  • And this is something that we're really excited about on.

  • I think it's worth talking a little bit.

  • Maurin Technical Diesel about how we achieve this performance, how we're using tensorflow on our supercomputers to get this kind of performance and get this kind of insight into our data and into our science.

  • Now it's stupid.

  • Computers are fairly specialized.

  • We have a specialized hardware to allow the tens of thousands of computers we have on these supercomputers to acts together as one computer machine you want to use this machine as efficiently as possible to train our network.

  • We have a lot of performance available to us.

  • We want to take advantage of that when we're running tensorflow.

  • So the approach we take as using a fully synchronous data parallel approach where each node is training on a subset of the data on we started off as many people do.

  • You using a g r p c for this where each compute notice communicating with the parameter server to send their parameter updates and have that sent back and forward.

  • But like many other people have noted, this'll is not a very efficient way to run a scale.

  • We found that if we're running beyond 100 notes also, then we had a real communication bottleneck between the compute nodes on the parameter service.

  • So instead we use MP I, which is a message passing interface to allow our computer knows to communicate with each other directly, so removing the need for parameter servers on dhe.

  • This'd also has the advantage that can really take advantage of our high speed interconnect, the specialized hardware that connects our computer notes eso we use a fig radiant aggregation for this.

  • We use specialized MP I collective All reduce, which is designed by Crais, who are our partners with our supercomputers on this empty I already use is is pretty neat.

  • It's able to avoid imbalances in the node performance, the straggler effect, that something you might run into.

  • It's overlapping communication and computes in a way that allows very effective scaling on.

  • We've seen that we're able to run tensorflow on thousands of computing those with very little drop inefficiency on something that I've been really excited to see.

  • Here is a M P.

  • I, or reduce is coming soon in tensorflow on.

  • We're excited to see how this is gonna work in the larger community.

  • So the three things I'd like you to take away from this talk the first is that cosmology has a really cool science problems and some really cool, deep learning problems.

  • The second is that scientific data is different from natural image data.

  • Armand thes well understood statistics that we often have associate of the scientific data could be a real use.

  • I think in the deep learning community on the third thing is the m p I or reduced in our experience is the optimal strategy for scaling tensorflow up to multiple nodes.

  • And we're looking forward to seeing how the rest of the community is gonna work with this essay.

  • Now turn things back to Lawrence.

  • Thank you.

  • Thank you, Debbie.

  • Great stuff, actually.

  • Simulating universes.

  • So we're running very short on time, so I just want to share these air, like, just three great stories.

  • But there are countless more stories out there.

  • This is a map that I created of people who've starts tensorflow and get hub on who've shared their location on.

  • We have people from the Outback of Australia to the green fields of islands from the North Arctic Circle in Norway, all the way down to Deception Island in Antarctica.

  • There are countless stories being created.

  • Countless great new things being done with tensorflow with machine learning.

  • We think some of those stories are yours.

  • And if they are, please get in touch with us, would love to hear them and would love to share them so that I just want to say thank you very much for attending today.

  • Enjoy what's left of Io on.

  • Have a safe journey home.

  • Thank you.

Hi, everybody.

字幕と単語

ワンタップで英和辞典検索 単語をクリックすると、意味が表示されます

B1 中級

機械学習とTensorFlowの進歩(Google I/O '18 (Advances in machine learning and TensorFlow (Google I/O '18))

  • 2 0
    林宜悉 に公開 2021 年 01 月 14 日
動画の中の単語