[00:00:06]
>> My name's Dierdre shoemaker and I'm a professor of physics here at Georgia tack and one of the associate director for the Institute of data in engineering and science and it's my pleasure to introduce today's planetary speaker while you enjoy lunch a professor or time comes to us from the University of Illinois at Urbana champagne where he is on the faculty of astronomy and computer science engineering he also leads the Gravity Group.

[00:00:33]
At the National Center for supercomputing applications so we share a common interest in gravity and gravitational waves which you might have heard about for the last few years is new to actions came in but I lead has a different perspective he comes at this from a very strong sense of a use for less machine learning and computer science in physics so he's very much become a pioneer in the world of gravitational waves and bring in these new techniques to our more classical type of research and you'll see today that he's going to talk about not only multi messenger astrophysics as a research topic but also how deep learning and high performance computing are changing the way we do our work so I hope you enjoy his talk let's welcome.

[00:01:21]
Thank you. Thank you for being introduction and the invitation to come to Atlanta it is my 1st time here and it is really a great place. So before I dive into the topics I want to percent to you today I want you to understand the context of the research that I am doing with my group so to give you a glimpse of the group I have created here you see students faculty and some of their high school students at that for my group and you can see that they come from different departments and units and I prefer to have this group at this you say National Center for supercomputing applications because the type of research that I do does not feeding in a single department or unit so if you're in astronomy it is rare that people talk about high performance computing if you're in physics it is you know it's still in progress to accept AI as signal processing tool but N.C.S. a place where diversity where inclusion and richness of ideas that come from all these different people have lost some son have led to really important accomplishments in recent years.

[00:02:36]
And in addition to having connections to national agencies that fund my research we have also realized that these critical last many people will know he didn't Georgia Tech that making connections within histories critical to the U.S. They are different here in computing in software and the connection of the 2.

[00:02:56]
Obviously you also want to be different here using the best supercomputers that we have in the country and right now do you know he's leading the way with supercomputers like summit and come in in a couple of years say 21 and this is a snapshot of my research group some of them are missing 4 new members have become Gravity Group students disarm or sought these will be updated next time you see it took from me.

[00:03:25]
OK now diving into one of the drivers for the research I do use is Gravitational Wave astrophysics is what I hope us about 4 years ago this was a dream for many of us those that have been in the field had you know the 6 side meant that when the advance like that would finally start collecting data we would open up finally to gravitational wave spectrum and he just happens that if you're following the news.

[00:03:54]
You have been hearing about major discoveries for high frequency gravitational waves in a band that is you know sensible to our ears but there are other missions like Closer time in arrays that don't use a gravitational wave detector on earth but rather to use a collection of processing the galaxy as their detector and there are plans to have interferometer sing the space so that we can get to frequencies around 1000000 Hertz we cannot do that here on Earth because we have earthquakes right so to isolate they did just from these type of noise you need to put them in the space and what is it that we have learned from gravitational wave us through physics Well just like that do try to follow this you use you have usefully sation of the 1st served as you know waves are detected by Lego now when you're seeing this we still a sentient think about everything that is behind us we are solving AI and since equations using super computers we're also using super computers to turn the day trying to these beautiful be smiley sation and one thing also that you can get from these whistle a station is that then you make a solutions we got from my sense equations and compared those to the data where basically the same a number lab in the time series data presented a gravitational waves larger than 99.9 percent.

[00:05:18]
So I think about is this beautiful theory that I use and developed more than a 100 years ago when you get to make a solutions they do describe nature with outstanding accuracy this was a triumph of many different fields computer science high performance computing theoretical physics to interpret what we're looking at and just within 4 years we now have a slew of different black hole mergers that like one view have detected and you can see that all of these are the ones that you see here were observed by X. ray a predatory is but only as new ones provide insights into the gravitational wave spectrum and it is worth mentioning these all of these have been observed without electromagnetic counterparts so we are looking into something entirely new it's a completely new spectrum and no good we have a catalogue you know Dick sidemen about seen the 1st gravitational waves sober but now we're thinking about what do these directions tell us about the distribution of masses and spins of these black holes because then discuss a natural connection to stellar evolution processes that lead to different mission of these black holes so we are now entering from the 1st pictures era into the best of P.-Six era and this is obviously much more exciting and with neutral stars you know that about 2 years ago like when we're going to take to the 1st neutron star collision and this was observed not only in gravitation of waves but also across the electromagnetic spectrum.

[00:06:55]
And here things are more exciting for you see the 2 new term stars they're orbiting around each other now you see title interactions here and what happens is that they finally merge and then everything that you wish to wear jewelry is producing these cosmic laps that then spread throughout the universe so interesting things to note about these observation.

[00:07:20]
And this is going back to the capabilities we have with with interferometers we only get to see these type of evolution using live one here drone before they merge but when we actually get to know exactly date entity of these sources is when they are colliding which companies have to kill or hurt.

[00:07:41]
In that domain of frequencies like who's not sensitive to those it is here where the traumatic observations provide additional insights What is it type of light that we see emanating from these sources and this can tell you what are you looking at to neutron stars when you dream star on a long massive black hole it cetera so only these I.D.'s.

[00:08:04]
Can be brought together to go up to new insights about the sources and that is one more you know people say that we did are doing gravitational waves to physics or support fortunate because we have seen a ton of the black hole mergers. Right now to neutron star collisions this is public knowledge but there is another source.

[00:08:25]
That we really want to see anytime soon and this is called collapse a part of if we just happens to be the case that these even have been seen that Aleksey then we will see waves originating from this explosion we will see electromagnetic emission that cross examine a tick band and we will also field neutrinos saw these could be you know dear I could typical multi messenger source you get to see the source using all these disparate cosmic messengers and this is something that we are trying to find and what you see here basically escort matter what is radius hot What is blue is cold so roughly speaking and I need to say it is because sometimes astronomers get upset.

[00:09:16]
We are not creating new commissioner astrophysics electromagnetic observations or neutrinos have been used in the past but it is for the 1st time that we can use gravitational waves to see events or to listen to events that were beyond a reach in the past. When you have systems that involve neutron stars then you also get to observe the light or gravitational aura trauma going to where is there any money to it from them and then we have events like collapse loop or Nova which are to give you an idea of what is happening here 20 solar must start that collapses because he has exhausted its nuclear fuel then we will listen to gravitational waves scenic summing up the waves and feel extra particles.

[00:10:06]
Now there is one thing that these cosmic messengers share. They have the following properties when you see gravitational waves you need to have an observatory already to capture the waves electromagnetic waves emanating from it and also a neutrino detector to capture these and all of this happens in real time.

[00:10:29]
OK now I am guiding doing day separation and to understand why we need to use AI to maximize discovery here. Gravitational waves describe a higher dimensional parameter space. Electromagnetic observations are a big data set of service ships same as here so we have this pretty challenges but they are big in a sense everything the person of these data which are a genius data incomplete very noisy need to be processed in real time and at this scale to make sure that we can let some mice discovery and this is the theme of these talk so to pierce the 1st elements of their gravitational wave challenge what we have right now he's a couple of detectors here in the U.S. that are the state of the art like go is collection data that has explicit quality from which we can extract gravitational waves.

[00:11:25]
We are using numerical relativity to so find since equations to obtain insights into the type of sources that we can detect and this information guides the development of algorithms to find gravitational waves in data and we also need a deep understanding of the sources so that as a mission before we know exactly what we are looking at when we say hi here have a gravitational wave signal OK Now tell me was he produced by school I didn't do new term stars something unexpected it's something that is beyond the police use of general itty bitty when we need theory to figure that out.

[00:12:02]
Uncritical we need high performance computing to obtain the simulations to process the data and to interpret the results so you see how all of these come together sometimes you can think of silos offices just working on theory or on numerical solutions or just gathering data but it is in gravitational wave astrophysics and broadly speaking multi-mission or astrophysics that for the 1st time we have deeper to need to do work with a very broad cross-section of the scientific community.

[00:12:36]
And here comes a challenge a slave one will continue to increase their sensitivity now have to put you to see signals for longer periods of time in the date. This also means that if you increase the sensitivity you get to see wider volume of the universe Naturally this means then that you will see more sources and the frequency will increase OK in the last 2 observing runs with Lego we were able to identify about one signal for every 15 days of search data this is not every 15 days of observation every 15 days of search data meaning when the data is good or if you want to translate this into several periods about one event every other month now when Lego increases their sensitivity to reach the target we're talking about signals detected every 15 minutes.

[00:13:34]
Now think about it if searching for signals can be done more or less in real time getting crude estimate for sources and we are providing for souls in the latency that takes several months for a single source how do we go about doing decent real time on the scale at some point we're going to be you know with a massive backlog of searches that we need to characterize and write papers about so these approaches not scalable even though right now Lego is using resources from up in science create that is known for kind of the physicists we have been using Lou Waters who have been using exceeds and other supercomputers in Europe this is still is not enough to cope with the sources that we have discovered in the past.

[00:14:20]
In different words we need to rethink the computing paradigm we're following here. And I used to say very clearly what the challenges are right now signal processing techniques used for repetition a way of detection the most sensitive ones target deaf or dimensional parameter space now black hole mergers define a 9 dimensional parameter space if you try to extend the signal many Follett you want to cover with these searches but yes become fully scalable saw covering these parameters pretty strong feasible now the other turns is that the scene is that we are looking for a week and they are imbedded in no negotiation on stationary ones OK so some of these searches have been studied that exhaustion like mites returning when we understand very well the performance and behavior of these algorithms in Gulshan ones and they have been working really well in a stationary house but we also have trouble when you have leeches contaminates into signals or glitches that most are normally right so the question is how do we go about this how do we develop better algorithms that are resilient to these type of noise anomalies and the following one multi-mission your searches are time sensitive we cannot allow to know some of these to delay our searches as it has happened in the past.

[00:15:52]
Now good or to be challenges sociological. People have been developing this algorithm for decades so we need to go smoothly about how we talk about after processing techniques because sometimes people take to heart when you present something I'll turn to what the community has been using for years he says to legal problem it's not a problem with you right but people sometimes take that personally now what are the opportunities well look at what is happening around you world the big data revolution has transformed everything even human interactions these started in 20 trends in 2010 when companies like Facebook a mushroom another's started producing torrents of data that had to be processed and then computer scientists that have been working in computer vision for a long time realized that if they combined these ideas of AI that were developed in the fifty's and combined those with G.P. use that was the big thing then it was possible to do image analysis in that way that was completely unheard of by 2015 neural networks that we now know us rest met for the 1st time surpassed human performance classifying images in a very famous they decide known as image Snit and after that the deluge now we have neural nets that outperform humans in many different tasks now for some people that is great for someone or see this is scary maybe out of lack of information or training on the field but the question really is based on these major breakthroughs seen technology and industry is there any way sign signs domains can learn from these experiences and try to address the competition I'll grant challenges that we are facing.

[00:17:43]
So I think it is just natural to ask can we use these I.D.'s to try to solve problems that require low latency and the processing of big data sets. And this is one clear distinction between these 2 most of the algorithms that we use that are traditionally in machine learning random forest you name it for much filtering these are for the most part well established you can treat one here one thing here on Earth thing over here but there is not a lot that you can change in the week era that we're living in there is plenty of room for innovation and I'm going to talk about some of completion and accomplishments that we have.

[00:18:30]
Presented to the community in the last few years and even though they look astonishing it must be emphasized that it is plenty of room for improvement so this is not a complete product and here is where your creativity and innovation are critical and then again social logical aspects.

[00:18:49]
People are afraid of machines taking over the world but that's you know streaming out of ignorance and we need to all to be dictated on these and not only always but also policymakers and people that are leading countries so let me tell you how these things are happening.

[00:19:07]
Used to also show that I have been part of the problem. Since I am at the National Center for supercomputer negations one of my contributions has been to produce competition of frameworks to use supercomputers to process gravitational wave data to model these using numerical relativity and to produce a result he says you'll see that the public understands what is happening and we have published these results here but what you can also learn from this is that there are way too many resources that are in China to do this type of things OK Huge You can see these supercomputers not only in the U.S. but also in Europe.

[00:19:46]
And so huge you have an opportunity to either continue pushing this pardon for work or to step back and rethink what you're doing and so I'm I'm sorry for it is slight It has a lot of works but this is the motivation for the rest of the talk and the idea here is that according to these philosopher he's always an opportunity to improve what we are doing what any decent science or any order human enterprise and usually there are 2 groups those that are looking for innovative ideas and others that prefer to maintain the status quo and at some point these.

[00:20:30]
You know rate of ideas get some traction. A given paradigm may looking complete the conservatives may attack it but if their new paradigm emerges and you know addresses the comments or concerns that come from the conservative scientists apart I'm sure it will happen and saw a little me talk about a few things that we have planted in the community to start looking forward to implementing this paradigm at full scale.

[00:21:02]
And the 1st driver is going to be gravitational wave us to physics and the Q motivation these days right we now see many Gravitational Wave sources I don't increase in pace. And we have depression needs to come up with algorithms that can bring together disparate data sets and we need to process them in real time with a scale forget about the sociological things that come with us working with different communities they just focus about the algorithm things and this is what I was telling you a few minutes ago.

[00:21:37]
We had some ideas about back in the fifty's machine learning has been going on for a long time traditional machine learning but it is right at this time when D.H. B.C. community the high performance computing community realized that the supercomputers we have right now are not getting any faster just why the I need you said this time around 20092010 that communities from Europe the U.S. and Japan got together to figure out a roadmap for axis scale computing I did was right around the time that the big data revolution emerged now high performance computing you know this is prestigious community they have been in the game for a long time they did not anticipate this but yes a few years later 2015 here in the US there was a pretty presidential strategy that consisted of fusing data revolution and its P.C. unheard of if you talk to people that worked with blue waters just 5 years ago and you had been there talking to them about ways my I thought library in blue what has little to be laughing at you by thing is super information in a high performance computing environment will have by then nothing to be greater evolution we have more deals that you can load in Blue Waters and other supercomputers for by Thornton for float setter no people no longer talk about H P C They talk about it's B. C. D. were these for date and you have heard about supercomputers in the US like summit and have over 20000 G.P.S. right state of the art for this type of analysis so if the high performance computing community was capable of innovating.

[00:23:21]
And into change the path for the creation of the nation or ration of excess killed supercomputers surely we should be able to do it and I'm talking about astronomy physics can we kill engineering you name it we have the opportunity to do that so let me tell you what I have been doing with the students in my group and they're the panel under the left has one of the key components high performance computing we needed for everything to obtain America relative to solutions trying to answer questions to understand the physics of gravitational wave sources but we also used H.B.C. to create the data sets with which we will train these neural networks.

[00:24:01]
And then since the models we're developing are super deep we're no longer using one G.P.U. one G.P.U. is when you're a newbie interest writing into fuel we now use kinder severe abuse so that we can train these neural nets at the scale. And obviously you have here 2 alternatives when you're training you can use accelerators as that once you get in there fate a supercomputer now going to.

[00:24:24]
Or you can also use G.P.S. depends on their architecture that you're playing with and what happens here is that when you develop the state of the our neural nets then you can accelerate processing of data and inference OK And what we have found for the neural nets that we have trained is that the final product is just a few megabytes insights so you can take that neural net as many of the ones that you have on your phone and process could have additional wave data in real time and so we played around with these with a variety of time series data like speech and then applied these to gravitational waves and for those that are physicists in the audience when I need talk about classification I'm talking about gravitational wave detection rate and when I talk about regression this is what I'm there is to Mission feeling now department is that produce a signal with a given shape.

[00:25:20]
And discuss the advantage that it is faster and deeper than any other algorithm we're using right now for gravitational wave detection and so this is the history of how this came about around $27000.00 we presented the 1st generation of these neural nets applied in gumption owns and then several months later in real life on noise you just happens that as I was mentioning before this has some social logical impacts So this paper was only published after these one was published because then the referee had no other human opposing the publication of this one you know this is really odd because if you had a person that developed much filtering centerpiece Agung and he had these papers and waited for publication with Matrix race and gushing on's and then there are 3 come comes back and says Sorry you 1st need to demonstrate the scene then gushing in on a stationary noise come back later when you have demonstrated that but this is the sociology of the field right you put barriers just hoping that the thing falls apart but anyways it is also really good that it's a completely independent team that is in the U.K. the other only information that we presented in these 2 articles and they were produced our claims completely independent which is really good this is reproducibility now for these signal processing algorithms that we presented here you know when you are probing an idea you start with something small it was just a 2 dimensional signal many fault now we are presenting algorithms that are at the scale and I'll walk you through that in a sec So let me tell you what are the main findings that we've got here and this is my student Daniel George that is now a scientists in Google X. what we've demonstrated here is the following.

[00:27:11]
We take a 2 dimensional team know many faults basically the masses of black holes and train a neural net to identify these times use data in realistic scenarios basically relied on us and we are using small training data sets we're talking about 45000 weight forms now the way we trained and you know it and this is important to understand why don't you only be so fast.

[00:27:36]
We have a training proceeding which we take for example a 2nd of data and then we plays that way from anywhere on that segment that is called Time invariance OK so it doesn't matter what a signal is going to be in the beta the algorithm will be able to find it if you try to name the position of a way form in the training data set is going to fail the other aspect is scale invariance a signal can come with any signal to noise ratio so we training earlier today he was able to see a very loud signal all the way to really quiet signals and he was really able to see all of these I'm not only the training big to say that we used but also new types of sources that are right now going to be detected with traditional algorithms and these are the papers that follow up on these so let me show you a recently session of how this is working.

[00:28:31]
On the left and you see a simulation of how did to block a program to merge and this is the 1st addiction by like huge you see the different layers of the neural net and this is laid in Windows person like a beta you're going to see that all these neurons have been fired up as we're processing the data and these are really busy because they act as the noises so basically the net is telling young Nothing interesting but as displayed in windows getting closer to Signa you will see that the neurons in the deepest layer said now active which is expected because here is where gets traction happens in the neuron it so let's go on a song we are still processing just noise.

[00:29:11]
The nuance here I really active output in basically a flood line because there's nothing of interest. And then as you get close to durian signal you're going to see how these neurons get really active now declassify or says all gravitational wave detected and by the way the masses that produce disable signal are D's ones.

[00:29:32]
So this is really important as well for development of AI we want to understand how the algorithms are working for that people don't believe that these are black boxes now that was you know the 1st approach I say said We use about 45000 templates to train those minutes no we want to use something more realistic black holes that spin now this is a 4 dimensional signal many fold and we'll use an about 10000000 templates for the training now we no longer just one G.P.U. that takes about 10 hours to train to nearly it now we're using hundreds of G.P. use to do this in a couple of hours and then again we can normally not just do detection but also regression and these are not only important aspect here most of the times the models that we use for AI are deterministic you don't get a lot of statistical information about the parameters that you measure but for these more complex ignominy formed we have just finished doing this analysis song.

[00:30:34]
To explain to you what is happening here this is entire catalog that Lego recently published of all the 10 black holes that have been detected in the 2 previous observing runs and here you see different parameters beside the masses of the 2 black holes here you see demean the standard deviation and disk units for each of these measurements the masses you can think of the physics before demerge or before the collision of the black holes and these 3 parameters are the ones I determine day dentity of the black hole that is formed after a merger so this is the final spin of the black hole and this real and imaginary frequencies basically tell you how fast the waves are saluting around merger and he says that dump in time of the gravitational wave.

[00:31:19]
Now this is the 1st time neural nets are used to infer destry properties directly from time series data and you know what is very exciting about is that this we can tell you right away what are the signals predicted by General itty bitty So the neural net can do under flight detection parameters to mission and tests of general activity each of these signals can be processed into milliseconds I suppose to from weeks to months using the standard algorithms that we have unlike So this is certainly you know a very optimal way to do parameters to mation.

[00:31:55]
And now we have a new dimension here we continue to use the tools that have been developed for many years that are critical for new developments but we are now trying to leverage what is happening around to be data revolution and this is one of the key contributions of my group to the field now let me just tell you another exciting developments in the last few minutes I have and this again comes from the convergence of gravitational waves and letter scalar ceramic of service so it just happens that when you see a gravitational wave source emanating from the whole collision as I said some of times some of the times they do not I made light unless they are in a really weird astronomical environment like surrounded by plasma we can now measured the expansion of the universe by correlating the location of the source to nearby potential hosts galaxies.

[00:32:51]
OK So the challenge here is just see a gravitational wave and you're able to identify the cause galaxy now just turns out that for to plug holes that Lego can detect there almost had traces of one or when the universe was half its current size and we don't have really complete deigns galaxy catalogs to do these now dispersant you see here burner suits is the one who predicted that you could use this you know waves and Galaxy catalogs to measure the Hubble Constant and he was just a word that did into metal you know a 100 years later it in turn was critical to confirm and since future of General activity but we have a few electromagnetic surveys that can help us do that and this is the dark energy surgery so let me tell you how we can leverage what we have done in the past and push forward these really know of 80 if I would say groundbreaking approach to label data this is the galaxies do projects that collected data from this lone GITTOES case for a very a nice astronomers to label data about a 1000000 galaxies and then you come to probability for the labeling which if you're confident that you're looking at a spiral or elliptical galaxy so these were certainly a 1st in the field now huge you have the dark energy survey that just completed up Servatius up to 6 years here we have 300000000 galaxies so who's going to volunteer to label all discoveries.

[00:34:22]
We need to think differently again so when I had been doing with my students is the following You remember today talk to you about image night when we have here about 1300000 images divided into about a 1000 groups labeled by humans. Now a neural nets. That Burroughs's to see midges are much better than humans I didn't define the objects here.

[00:34:46]
And the State of the units for human specific ation least exception now what we are doing is the following we have this images from the Sloan Digital Sky Survey we don't have many images labeled from the US to classify galaxies but we don't have to so what we do is we take some of the best images from as D.S.S. and do transfer learning to these nearly it so no need to have a galaxy images in mass we just start on Friesen delay years the deepest layers of the net and then gradually tweak the neural net so that now than here on it it's not used to classify the real object images but to classify galaxies and what we find says that we can obtain state of their accuracy to label galaxies and there is even something more important that we can do let me tell you whether these you see these 2 clusters here they represent galaxy midges from desks that are on labeled We feed them into the new unit and we're considering 2 classes spiral and elliptical and then the neural net tells us hey from all those images that you're feeding into me I realize that they belong to different classes based on their morphology looking at 3 different filters and so we go and look at their 2nd to last a year of the new meds that has 1024 neurons and we look at the activation values you can think of this as the fingerprint of these images and then we divide them according to their morphology as you can see here so with this approach we are now certain the training data sets from this that will allow us to go on go through the entire map and construct a complete galaxy catalog that we can use for grip this in a way for smaller.

[00:36:36]
And C S I was talking about it's P.C. as well I newspaper we achieve for the 1st time to convergence of distributed training and also a deep thirst for learning so the process that usually would take about 5 hours we can do it in 8 minutes a huge you can see one of the new classifications we get from are nearly 8 so you can see on your own it can be used not only to classify to do regression but also to start developing your own training the distance which is pretty good and now let me show you how they're activation values of the new I'm spurious.

[00:37:09]
You're doing the unfreezing So here you start with the 1st weeks you see that activation values are not that big the yellow represent the spiral that has more structure and a blue that is not very feature rich and you see that you had to can do nearly that very rapidly did you know lead feet or south that what you're what you are presented to it are spiral galaxies and then later on dictation of their leap to get Alex's pickup and then joined up with a new Only that is super a bust to get into fight these 2 classes of images saw then again doing this type of utilization helps you understand what a neuron it is doing and you know when we did these with a bunch of images and we we found these it was a bit counter-intuitive it was like I see that in you know it is just figuring out faster to spiral down to leave to go and so what you can do is you can extract the different parts of the image that in your lead to sex strikes in its layer and we indeed found that because a spiral galaxies are a feature it's nearly This able to grasp those features faster.

[00:38:17]
So it's critical and. Just to try to summarize this is the type of things that many institutions even here at Georgia Tech are pursuing interdisciplinary research that has strong component not only on some specific domains but don't also try to bring corporations seem to this game and this is this is scary because you don't want to be creating a ball from scratch every single time I own a thing in video because they have been heavily supporting my research and I think many people may also benefit from in interactions of decent nature so let me finish with a picture of my group and begin the slate about being bold being creative and understanding how we can leverage this outstanding developments to address some of the competition or challenges that we're facing in your respective fields Thank you.

[00:39:15]
Clay. And Mike. And. Yeah so you know fields that is very similar to what you're describing is kinder geophysics so they produce a ton of data they will not process everything so they are using machine learning right now to determine what data is worth and ally soon and what data will be discarded so even at that level of machine learning has been use really useful now their order challenge is related to real time analysis where they mention in connection to Lego there we don't have a big data problem it is just about to make a bytes per 2nd you can use your phone for that but for illicit Steve which will be districts this sort of D.S. they rescinded a big problem we will have then a bytes of data per night a 1000000 subtree years and a few of them will be related to gravitational waves so then the question again is how can I rank these 3 years and make sure that I'm going to follow up the ones that are of interest to me so the power of again machine learning can be used to talk I'm pleased that I mean is right now the time to play these games we are doing Dead with this we are creating synthetic data to simulate delicious Jeus in areas so instead of receiving one image every 6 hours a day this is currently doing we are producing the 7 minutes to try to understand where the bottlenecks and how we can fix them not be shy.

[00:42:15]
Some way approach now is to partner with any site or and and so we're going to be deploying days for real time analysis I think once people see how it is working then all these preconceptions can be addressed now are the more technical level for example when people say OK you're telling me that your neural net can estimate the promise of the source and this is the a statistical error did you get out of it how can I trust in your model so our you know there are some procedures that people follow now for example you trained in unit since you are.

[00:42:53]
The weight so you want to see where it. Differs Solti is robust and you did use the 100 times or Jews who promoted you layer says once and how you do training with prayer surrounded wits and now you also get an error for the parameter So you're predicting for there are multiple ways drivers discontents we only need you know on a thoughtful pointed question that we can't answer.

[00:43:40]
Thwack.