[00:00:05]
>> All right good afternoon everybody welcome to our final plenary session of M.L.S.E. 2019. It's a pleasure to introduce Ross Thompson from Google Ross trained as a computational physicist and has worked in a broad range of academic and industry fields for microgravity fluid simulation for NASA as computation to computation advertising Google currently He works as a solutions architect for scientific computing and the Google plot cloud platform which I'm sure he'll tell us more about thanks for that introduction and thanks to all of you for staying to the last session of the last day of the meeting I was afraid we're going to be like 3 people in the audience so.

[00:00:54]
Thank you give yourselves a round of applause for your good dedication. Solutions architect Google is. Not working right now let me just. See if I can find my. Group. About that. But that. Seems to last count today there we go all right solutions architect to Google has 3 responsibilities the 1st of all we work with customers 2nd of all we work we go to converse conferences like this we do evangelism we speak to people in public and 3rd of all we write up what we call solutions so after we solve interesting problems for customers we write up those solutions and publish them on the web if you go out to Google's cloud platform and you want to figure out how to do pretty much anything on the Google cloud platform should be a solution out there to be fair they aren't all done yet so if you find something you really want to do get in contact with us and we'll try and figure out the right way to do it.

[00:02:09]
So I guess I as you know I was trained as a theoretical computational physicist through many winding route I ended up at Google I got to say that I think I have the best job at Google Don't worry about the C.E.O. also don't tell my boss because he finds out he might make me work harder to get to work on a daily basis with scientists like all of you and it's a real privilege to be here so what we're talking about today is machine learning because you guys have been here all week turning about machine learning and to be full disclosure I am not a daily practitioner of machine learning I've done in the past in many different flavors but day to day I do not keep up with the hard core tensor flow new models that are coming out however I do know fairly broadly what's going on with the cloud platform and I'm also going to talk about that and a bit more about the industry and where things stand so without further complication so if you look at.

[00:03:06]
All the hype about AI It's kind of like the weather everybody's talking about it but nobody's doing anything about it you would expect today I am machine learning were everywhere the truth of the matter is that despite the number of research papers going up and according to this chart almost as fast if not faster than Moore's law you would expect that pretty much every industry had achieved some level of AI capability but according to studies we're probably at about 20 percent now of companies that actually have a working production AI model.

[00:03:40]
Google is probably up there but I'm going to say even arc we're not at 100 percent of all of our tools are not a base which kind of makes sense. So one of the real problems in the industry right now is that there is a lack of expertise it's machine learning as you know is a very hard subject takes years to master.

[00:04:03]
Of that we estimate 21000000 programmers that are out in the world probably a few 1000 of those are actually capable of producing a custom machine learning model so Google's response to this is to do what we're calling democratize ation of machine learning and of course the way to democratize is to use complicated words like democratize which is what I like about it so why is machine learning hired as an earlier presentation with Dr Chu this after smarting it was talking about all the things you need to do.

[00:04:39]
In order to get a model that works. One of our developer advocates highly Crosby has a great diagram where she takes all these circles and puts them in a bubble and all the and I don't turn out the actual training of models as a tiny little bubble at the center because you spend most of your time as you saw on the last presentation actually cleaning up the data and getting it ready.

[00:05:04]
It's no different on Google but so what I'm going to take you through is some of the tools that Google has in order to enable this kind of work. Clearly you guys are at the forefront of a lot of machine learning so maybe a lot of this work isn't going to be relevant to you but I hope some of it will and some of it will potentially add value to your research projects.

[00:05:24]
So this is one of our favorite statements Google is an AI company you know I'm realistic about it Google from the outside is an advertising company we make 90 percent of our revenue by selling ads and we do it really well but from the outside end Google really is an AI company how do we make our ads better we use artificial intelligence when you're doing a $1000000000.00 worth of work you know point one percent increase that still turns out to be real money so every every step of the way Google is using machine learning and I to improve our technology there's not pretty much any project product that Google exposes externally that doesn't have some kind of machine learning in it and you know it's we've gone from using classical machine learning Basine approaches to now tensor flow is propagated everywhere inside Google So let's go back to democratizing A I N M L What do we mean when we say that somebody figured out that Google has a really cool mission in life which is to get out of the world's data and make it universally accessible and useful so our AI people said let's gather and make how we phrase here democratize AI by making it accessible sensible fast useful for enterprises and developers I work with some astronomers and they say well sort of just organizing the world's data why don't we organize the universe this data so it comes to that to machine learning is going to be everywhere Google wants to help people get their.

[00:06:57]
So another great way to democratize things is to make a really complicated diagram like this one. We just try to say that people tend to confuse AI and M.L. I don't know if there's a really strong boundary between them anymore there's a lot of different things going on some of them you can categorize as M.L. some of them are AI.

[00:07:17]
At this point I don't worry about it I think the cool kids say M.L. and the newspaper say I think this may be the way to think about it. So Google plowed platform. In our effort to democratize AI is has basically produced 3 different flavors of machine learning it's kind of like Goldilocks and the 3 Bears there's the 2 cold the just right and the 2 hot and I'll take you through these so what does that mean to the far.

[00:07:48]
When you're looking at right the diagram is our set of A.P.I. is that you can use these are basically precut and A.P.I. as with all of Google's technology built in so it's Google models and Google's data and you can call these A.P.I. and get back answers in the middle is the easier approach to M.L. which we call auto M L which uses Google's models and your data to then augment those models and to the far left we have you know this is the hard core tensor flow custom model development.

[00:08:26]
As you can see the graph on the bottom says for the left is easier for the right is easier and to the left is harder. Let's go through the A.P.I. eyes I think a lot of these A.P.I. eyes are targeted toward enterprise businesses not necessarily toward research but some of the talks I've seen point to there being a strong possibility that some of these guys may actually be useful in certain contexts for some research programs so one of the things we can do is we have something called the detail D.L.P. A.P.I. The data loss prevention it allows you to identify dataset so it'll run through a text file and depending on what you tell it to remove it will remove social security numbers names telephone numbers emails so for people who are doing it give me a logical research or social research of some kind and they have personalized personally identifiable information this A.P.I. could be useful to remove it all right so the vision A.P.I. is a cool one.

[00:09:29]
You can throw a picture at. The A.P.I. and Google will tell you is there a dog or a cat or a house if there's text and it will extract the text and if there is a logo in there it'll tell you with a logo Additionally it'll do things like find.

[00:09:47]
Objectionable content for instance the natural language A.P.I. which we presentations this morning also talk a little bit about we'll take text analyze it and you could teach your children how to do that structuring of sentences Graham pulling apart the sentences or magically. It's got some interesting. Applications for that so if you were actually trying to access access large amounts of data do sentiment analysis on it figure out what was being said and you could actually use this A.P.I..

[00:10:22]
The speech A.P.I. is basically speech to text text to speech it does recognize and it does speech recognition so to pull down data there's a cool demo that online where you can just talk and it'll basically take dictation I could run that for you now but I think I'm actually cursed by the demo gods so I never actually attempt those in public.

[00:10:44]
And as the translation A.P.I. clearly if you've ever used Google you've probably messed around the translation A.P.I. Just in case you're taking a French class don't try to use the translation A.P.I. because as my wife is an Italian teacher she will tell you you can always recognize her right away when it's been done but still it's pretty good.

[00:11:03]
In addition to analyzing images we can also analyze video through the video intelligence A.P.I. is also a great one so they come back to the middle which is start to get a little more interesting potentially for researchers is the auto mail platform where now you have your data trying the models and you going to retrain some of Google's models what does that look like so one of the cool I'll skip this one so the goals of cloud auto M.L. are to make it easier to train models to you don't have to be a full level M.L. practitioner you don't know anything about tens or flow you just need to have your data and it'll simple to train and then you can actually serve these models up he said that So 2 of the things that are pretty cool here is the learn to learn ideas that you've got a neural network which will actually teach other neural networks how to well actually do the training internally so this is the stuff that's under under the hood and auto M.L. is the learn to learn where neural networks train each other there's the transfer learning where there's train models that then take your model your data and get retrained efficiently on their.

[00:12:13]
And thirdly and one thing I find the most interesting is hyper parameter tuning and I think that type of parameter tuning could potentially be of use for researchers who have models with high levels of parameters and a very broad parameters base that they need to explore in my days as a physicist you know we used to just iterate through the parameters and do a full grid it was pretty painful the really cool application of this is.

[00:12:39]
We try to figure out how to make a better dessert there's a couple crazy guys who sit about 2 chairs over from me in Pittsburgh Greg chance he and Benjamin Saul Nick who among the other crazy things they do they have a product internally that is hyper parameter tuning is called busier and when you join the busier team you get a 0 hat which looks like medieval Turkish busier.

[00:13:02]
The goal of this was to define a recipe run it through the hyper parameter tuning A.P.I. And just to see what they could get out so naturally they put certain constraints on it so that you didn't put a half a pound of salt into your cookie recipe but they did some amazing work in collaboration with our cafe chefs made the cookies on a weekly basis they put them down there you got to vote on a one to 10 scale and they would take that data back and then go.

[00:13:31]
Another pass through the hyper parameter tuning it was pretty cool at the end they tasted interesting they would like to have a teaspoon of cayenne pepper in the recipe so that might not have been for me best. Well more about the hyper parameter turning if you've got a complex parameter space and the objective function is you know somewhat textured sometimes the standard optimization tools can get stuck in and minima The nice thing about piece this parameter tuning tool is that it's been designed to deal with these complex surfaces and navigate through them normally this is used in building a large scale machine learning model using tensor flow whether to convolutional network for example many layers and lots of interconnect you can use the hyper parameter tuning model to decide how many layers and how big they should be and try and choose an optimal model but it doesn't just have to be applied to machine learning one of.

[00:14:33]
My very clever colleagues to lock who used to be chief scientist that no one has come to work on my team doesn't work for me even if we were colleagues on the team and he does lots of amazing stuff if you've ever done any of the Google Quick labs code labs or training sessions LOC figures very heavily in those so what he did was take a very simple model of traffic congestion with 2 externally tunable parameters submitted it to the hyper parameter tuning model and A.P.I. and was able to you know find an optimal solution to the problem it could be extended to a lot of different optimization problems I think it has possibilities.

[00:15:23]
So the auto mail vision tool is one that I think is really very interesting at least we've done some interesting things with it for scientists in other contexts so what does it do it's basically you have a bunch of images that are potentially label so in this case.

[00:15:41]
One of our code labs that for training will take you through how to build a flower recognition tool where you can get a bunch of flowers and images that have been labeled you send it up to the vision in A.P.I. and then you can get a model that will then subsequently record recognize flowers it's pretty cool and actually pretty easy to do.

[00:16:02]
In fact one of our projects from the lab set up at our cloud next in April out in San Francisco was to set up a learning model to distinguish between a hex bolt of 10 millimeters and another one of 7 sixteenth's and it actually going to the I can't tell them apart but frankly the machine was able to figure it out and get with like 99 percent accuracy more interestingly is this work that was done by my colleague Jamie Kenney brilliant guy he was at it astronomy conference back in December I guess and he was somebody said you know we have a all these images from Hubble and from all these there's the base telescopes that have images of these objects and it's very laborious for us to go through and label them so he said well the vision ab I will do that for you and by the time he says turn to present around noon he had this model built he took a bunch of images categorized them by the different types of galactic shape and he was able to do a decent predictive model at another conference he went to he's always showing off he took plankton of various forms and again built a labeling model around that so if you know my boy Jamie can he's pretty smart but I think if you can do it in a half a day most of us could probably do it in a day or 2 so it's really interesting work.

[00:17:28]
So. Tables is another interesting one I think this could have an interesting when you have a data set that you just have a set of parameters and you have a set of yes no responses for example you can build very quickly and all model that can actually be served up so I think people who have tabular data doesn't have to be just pure images you could also run it off tabular data and get a server model.

[00:18:01]
That's one of the things that we talk about I'm not sure why we use this term but. Being able to handle data in the wild so the goal is to not to have tons of pre-processing that you need to do with the tables. Or with Mel but that you can actually take the data pretty much raw as it comes and get a cervical model out of it and again so here's another awe thing that might be useful for people in certain kinds of language research where you can actually train use our neural machine translation model to then put in your own translations and be able to translate things that are very technical and potentially get decent answers and.

[00:18:45]
So the takeaways from auto Mel are that we want to make it easy we don't want you to have to be. An M.L. specialist to be able to run these There's a lot of potential for use in other research fields aside from just M.L. so and here's the last and the most.

[00:19:09]
Challenging part of getting to M.L. is where you're actually generating custom and male models using tensor flow or other applications like it Larry. So what is our what we call the Google AI platform it's basically a platform as a service if you're going to run. M.L. models you can generate models directly here.

[00:19:35]
It's industry scale you can scale it up and have very large models if you need to go beyond what a standard C.P.U. can do I'll talk a bit about more about what that is. And has built in version management and because it's a pure tensor flow model being developed you can take it any platform and run it so there's technically not any vendor lock in so as you know tens or flow is Google Open Source this and has been supporting it in developing it we just released version 2.0 of tensor flow I understood there was a presentation this on Sunday about this and one of my colleagues did it's it's a challenging piece of code to learn and to work with there's a lot of things they've done to make it easier internally but it still takes somebody who's got skills to be able to run it it's not not somebody off the street so again.

[00:20:26]
Internally in Google as you can see this is the kind of the ramp of the number of projects we have internally they're using tensor flow it's very fast. So in order to. Be able to run some of the models that we run a Google regular C.P. using regular infrastructure we're not sufficient to handle the load and you couldn't train things fast enough so we started an internal project to develop basics the patient specific integrated circuits to support our work so that we have released what we call T P U's they are now generally available for people to sign up to Google Cloud platform or run on a one of they do they're just massive matrix multipliers So you have tensors you can add we call them chancers but they're really just major Cs of different sizes you can multiply them together you can do various algebraic.

[00:21:23]
Algebraic operations on them. The top row is where you have a single T.P. you tensor processing unit and there's 2 different flavors There's the version 2 which has been out for several years and a version 3 which is only recently just come out of beta so they're now both available on the bottom row is what happens when we take G P U's and we stack them up and connect them up to make thousands of them and then we get you know correspondingly higher performance as a result they're pretty cool devices I've personally never run on one but I have seen them in the wild and you know there is a thing of beauty the one of the top right you see all the copper on it's water cooled so that you can stack them up very tight if it was air cooled you'd need to have a lot of air flowing through there so the water cooling allows you to stack them up tight and have much shorter interconnect as a result it's pretty interesting engineering.

[00:22:17]
To go back this is what they look like when you see a pod all stacked together with all the cables I don't know who the guy is who does all the cables but he's very very tidy. So what kind of projects can you do with tensor floor there's a ton of stuff out there I'm going to go through a few of them one of my favorites is Christian Lew and Andrew Africans in last name from you.

[00:22:41]
Put together this model to detect exoplanet transits there are a lot of telescopes out there that are measuring the intensity of the the brightness of stars and some of them are variable for various reasons but some of them very because there's a planet that's going in front of them and that signature can be detected in a variety of ways with a lot of algorithms out there to do it but a lot of them were still not being detected and this work they found a new way to do it using tensor flow I think it's pretty cool they discovered a few new planets so it's Andrew Vandenberg So you asked me Ross I don't know how to get started on all this.

[00:23:21]
It's hard not to lie it's hard but Google's trying to make it so that you can find all the information you need so just like the all the solutions I was talking about earlier we have a place called Ai hub which solutions architects like myself are contributing to we have people from our developer relations teams our product teams are all contributing explanations of just how do you get things done around in learning on Google Cloud platform some of them are specific to Google Cloud platform some of them are you know more generic and just how do you do things with tens or flow sample models tons of code out on get hub.

[00:23:57]
Even that can be a bit of a burden because it's so much to go through and learn. And so some of the work that's been done it's really cool is Deckard lab has they are mission is to understand sort of the geography and atmosphere and how all that affects earth and human society so one of the projects they did with us was to you know calculate crop yields before the before the government actually released the data we also worked with the international group to try and who are trying to stop illegal fishing by using all of the satellite data they were able to detect illegal fishing off of various countries so it's it's really interesting work you've probably seen this it was very popular on the news for a while Google's been working on detecting.

[00:24:53]
How to pronounce that and it diabetic retinopathy it just by taking machine learning models and looking at the images. Additionally we've been working with but ologists to detect. Cancer in. Slides So there's a lot of a lot of work going on like this every week you'll see new publications coming out where people are using machine learning to do medical work Google is very much in the forefront of working on these kind of things we have a whole team dedicated to.

[00:25:27]
Machine learning and applications to the health field. I was working with a woman from Katherine ruff from our research department was working on trying to do machine learning on health records health records are a big mass nobody ever enters you saw earlier today that how many different variants of the word vomiting there could be this is the kind of thing they deal with and she's been working on trying to apply machine learning to trying to clean up this data and understand a little better it's a huge challenge and machine learning can't do it and maybe it's not doable and so that's sort of the machine learning is cool part.

[00:26:07]
How do you actually do a whole pipeline a doctor to this morning talked about. You know again this whole pipeline if you want to get your data clean it up feature analysis and then train it serves So Google has a whole set of technologies around this so embarrassing to put these up at any time you go to a Google Talk for Google cloud there's a lot of blue hexagon as well I'm not going to be any different today each of the little blue hexagons has a certain meaning about a product on Google if anybody can name one of them the weight they're all labelled never mind is going to say so how many people here have actually logged into Google Cloud platform before all right that's good prize for you well prize for you prize for you.

[00:26:52]
Baseball is not my forte. We actually have somebody internally built a game to do. Dean and before I gave up. Our marketing people are amazing but sometimes they make me laugh so. You can see so the idea here is that similar to what Dr Chu talked about is that you've got to get the data in you've got to manipulate it you've got to serve it up you've got to train it and then hopefully at the end you're going to put together something that serves it up to your public you guys doing research probably building a serving side of this problem possibly less important I don't know for sure but.

[00:27:51]
In industry the whole idea is that you want to build a model that people can then use so this is kind of. The canonical and away end pipeline for machine learning that we talk about so what are the tools that we have there's a lot of them and they're some of them are pretty cool This is called data prep data Prep is a collaboration with a company called trifecta and Google has been using letting people use dry fact on the cloud platform you can take very large files trifecta will grab the 1st 1000 rows or so and analyze it and show you the top row they're showing you distributions you can build.

[00:28:43]
Transformations what they call wrangling and then you can then apply that those transformations you built directly to your data and the entire dataset it's it's a pretty cool package. It's not directly integrated into all I mean it works talks to Google but it's still a separate package run by a separate company.

[00:29:07]
One of our core tools that is really the workhorse of our data flow our data pipelines is we call data flow externally it's known as Apache beam we open source that it's a. We now the buzzword now in cloud technologies is server lists that data flow is a service platform the idea of service is that you know personally the thing I hate the most about doing any project is the configuration I hate spinning up machines I hate connecting them together installing the libraries making sure everything's working it's always a huge pain so service in theory takes all that away so data flow is one of the tools that we use it's basically a Map Reduce but written in a language you can write it in Python or job but written in such a way that it's actually much more intuitive the kind of transformations you're doing you can also integrate to a lot of different tools.

[00:30:03]
We can integrate the big query on the other side it's a pretty flexible tool and again if you're interested in doing this there's a ton of documentation out there we have all kinds probably more code samples and you'll ever want to run your life one of the really cool features about.

[00:30:23]
Data Flow is that when you write a data pipeline with it you can take a whole batch meaning a whole block of data and run it at once or you can do a streaming model so if you've got something coming into it from a bunch of devices or from a bunch of external websites you can use the same transformations but you use it and using it in the same same code so it's a very effective tool for doing those kind of things some one of the other tools we have is Google.

[00:30:55]
Has been hosting Jupiter notebooks we've tried a couple of iterations of this. Run original product was called Data lab I wouldn't use that if I were you I didn't like it so now we have a couple of new ones out there there's the cloud AI platform notebooks which I have to say I'm a big fan of you just go into a cloud platform and click a button and then wait 3 minutes and then it pops up a link where you can go and get your notebook and it will make available G.P. use to you it comes pre-installed with tensor flow or Python or a bunch of other analytical tools really very easy to use our other tool is called colab which is also publicly available.

[00:31:43]
The AI platform you'll pay for the usage of the machine that it's running on but colab is 100 percent free you can get G.P.U. T.P. use you don't can't get as much processing but if you're developing new code you can play with it and it's very convenient because you can store things on your Google Drive very easy to use.

[00:32:07]
So it's got to be at least one person out there who is who loves sequel who loves people. Is. Some people want to do everything it's equal I'm not really one of those but big worry is possibly my favorite Google Cloud platform tool. What is Big where it's.

[00:32:36]
Used to be an internal tool the Google use called Dremel basically was designed to handle all the logs of all the people who were doing all the queries all the other interactions of people with Google are collected and logs in the logs were analyzed to make sure that you know we were doing things optimally So again we could serve you better ads so that you could buy more stuff but it was used internally for that so the same technology has now been externalized in Form a big query and it's really a great product.

[00:33:04]
I've been using it with I did some work with our friends at the L S S T The large synoptic Survey Telescope which always gives me a tongue twister with them doing a proof of concept to see could we store some of their star catalog data in big query and for certain applications I think it's going to be far exceed what they're currently using and for others you know it's kind of a toss up so big is an amazing tool I stored 18 terabytes of star catalog data in there for the for their tool to do a full table scan on that was taking them on the order of you know hours to get through it if not more not days and I was able to take that same query and get it run initially in 60 seconds but then after some optimizations and tweaks I was able to get it down to like 4 seconds to run so big has a lot of power potentially fast a lot of the work that people are doing on a dupe and Sparc could probably be done on big query.

[00:34:08]
Big query is very affordable you don't pay. You pay to store the data but it's roughly the same price you pay to store it on our cloud storage in general so pretty cheap and you pay by query and even that's not very expensive in fact our sales people kind of hate this product because they can't make any money selling it people so it's always a good thing to worry about so.

[00:34:33]
What else can you do aside from just do regular sequence database we started introducing a query and now where you can actually do machine learning directly inside of. Big query right now it's only we have some relatively simple models. Linear fitting. Like my brain is failing me now but some relatively simple fitting we're moving into You'll soon see some announcements if you follow big very on Twitter for example but you'll see some announcements about the new kind of optimizations that we're offering in fact big where you're going to become kind of a center of an ecosystem of tools and if you love sequel which apparently a couple of you do you will be able to do pretty much your entire data pipeline inside a big wary which is I think going to be really interesting.

[00:35:26]
So what else can you do if. You like to answer flow sure there's somebody out there who likes to answer plenty but they're like Spencer flow. It's not just it's not just for breakfast anymore it's not just. I really think it's. A lot of the simulation work you can do out there is a fact of li.

[00:36:00]
Just matrix multiplying you can actually do sort of fluid dynamics simulations using to answer flow and there's a very simple example out on our training site to do a simple separation of water droplets falling on a pond. But what if you don't love tensor flow there are other options.

[00:36:22]
We have a there's a language called Julio which I imagine some of you are familiar with and I'm out of little things like and ask you who love Julian's anybody you Julia All right so Julia has a new up and coming language. There's a lot of simulation work going on there it's becoming a fairly broad community a couple of guys I work with are huge Julia fans but people have written.

[00:36:47]
Pure fluid simulations in pure Julia now instead of pulling things from other libraries and then really cool thing about this that I forgot to mention is that if you're doing large scale simulations if you use tensor flow you can run it on your local C.P.U. without doing any significant modifications to it you can then run it on a G.P.U. or also on R T P U's getting amazing acceleration out of it the same thing is true of Julien Leyre which if you're running a pure Giulia you can run it on R T P You systems and as I'm going to tell you in a 2nd or 2 you can actually get access to some of these for free and get pending on your project if your project is entry interesting you can apply for a access to us like a 1000 T P U's and run your simulation and get some interesting work done so.

[00:37:39]
What do I do now. If you're interested in trying out some stuff with Google cloud there's a variety of ways to get access to do that you can get anybody who wants to get free credits on Google right now. You can apply. Really polite about it you might be out to get $1010.00 K. to do your project this goes by project so if you do a 2nd project you can potentially renew the grant you can't do the same grant the same year so be careful how you structure things but you can get a fair number of times renewed.

[00:38:23]
So that's that's my public service announcement number one. Public service announcement number 2 is something people should look at is our public data set I think our competitor Amazon also has a public data set but these are pretty cool is that we take some of the more important data sets out there we.

[00:38:42]
Put them on big queries so that you can access them we also have some that are some of these projects from no we have some massive amounts of satellite data up there for the GOES satellite so there's a lot of interesting things up there for you to look at and again if you don't query the entire terabyte data set it can be very cheap so to structure your career he's carefully.

[00:39:09]
This is one we're working on right now which is the guy. Data set from the European Space Agency massive amount of data about astronomy astronomical objects I also published one recently called the wise data set which is the which is an infrared survey of the U.S. So there's a lot out there and I spoke really quickly and we're at the end of it and the topic.

[00:39:38]
I do also have some more information up here if you'd like that to. Come and see it after. Any questions you want to. One of my projects is a full employment for physicists. Yes I had a really hard question planned but unfortunately I can't remember it now I just want to say for those of you that I could not get around to let me give you a U.R.L. that you can go to to access information about the research credits program that is cloud Doc google dot com slash edu pretty easy to remember cloud a google dot com slash edu And then the other thing that everyone can do is go to the cloud google dot com We have a free tier where you can sign up and receive $300.00 of free Google Cloud credit and access the free to here for a year thank you very much.

[00:40:45]
So my full employment for physicists was have physicists do more work on tensor flow and not everybody's going to be professors so maybe you could be a machine learning guy at the end or gal. Hello so I know I'm working reinforcement learning and I know a couple of people here have been working also in that area and I know Google participate with Google did mine and I was so kind after they opened up Roach OK So we have been creating our own him by domain mile an hour own agency and I see that you're releasing some for 2 also saw no sign of the agency and raid revealed I don't know if that thing where we continue growing and you will operate as the same way you you showed a machine learning techniques for image processing or something like that where you can embed your own by domain and write all day agents or something it's possible I don't know anything about that roadmap but they hold their cards pretty close they don't share a ton of that information but I would expect to see a lot more things our new head of cloud is actually based in Pittsburgh His name is Andrew Moore he was one of the founders of the Pittsburgh office and he went back to Carnegie Mellon University dean of computer science but somebody offered him a big bag of money and so he's come back to work for us to head up our.

[00:42:16]
Efforts and Andrew is a true genius Spiers people I've ever had the pleasure to work with looking for the kind of work he's going to produce. Thank you for a great talk. I'm glad to see. There are some applications of science engineering simulation even the very Also the real really hard problem for for for the beyond is image processing so of course the biggest question for machine learning for solving those scientists or engineering problems are lack of data in a sense that a lot of complex engineering problems see high dimensional and in other words there are so many prime to hundreds of prime parameters probably needed to in this case are you just don't have enough data not there this is just too sparse is very expensive to get a data submission to that war experimental data.

[00:43:19]
So you know from from Google's perspective have you thought about this problem. Going beyond existing approach in the solving say image processing in vision type of problem really some engineering simulation complex high dimensional problem you have very. Basic optimization and all the hubbub you know how to have a good process model to find the optimal process printer to have the kooky say you probably don't don't have hundreds of thousands of prime interest and.

[00:43:56]
Bat for high dimension problem because. The expensive competition expense for the data point because you have so many prime to 2 and you have to collect so many data points in that case and then just doesn't work you know I don't I'm not. Any in any way an expert.

[00:44:18]
Optimization tools don't know where the limitations on what we can actually put into our. The I prefer. But you know if. It's Google's always interested in collaboration with people if you were interested in trying something with that tool it's used internally with models that are fairly complex so I don't know what the limitations are but there's certainly an opportunity to just try to push the envelope and use it for as many as we can but yeah I don't know if there's any easy fix in that sense when you've got high dimensionality you've got.

[00:44:55]
It's not really a great options yeah but we have lots of computing power so if you want to run more and more simulations That's another option. Thank you very much appreciate you.