[00:00:05]
>> 1st of all thanks to the fund David for hosting us and giving us the opportunity also to share. A little bit about what we are doing what I'm talking about today is a topic that is very dear to us all the very practical considerations that comes although for a long time experience of things not working in the real Of All right so.

[00:00:30]
Just very briefly I got on my side so I have. Long experience and applied R. and D. I've worked in a variety of roles from and of is a contributor to more on the people management in R. and D. side all more on the Expo side. We've been able to work in various mostly industrial domains from automation of a mobility to some of us and security.

[00:00:59]
Before a thought with the content just very briefly about a logon a zation team and comforting them to the female 13 to being a large. Tronics digitization conglomerate hosted in Germany. The month in fiscal year 2016 so these numbers are already outdated but they're growing spent on the Auto $5000000000.00 in R. and D. across all of the industry industries that we are active and.

[00:01:31]
We are filing many many patents to the tune of between 15 and 20 patents that they globally. We're working with strategic partners and academia we call these the C K I universities there are 9 of them at the moment and I'll just take certain is one of them we have large strategic alignment on the various We saw divinities we are a global organization and not only the research but even as a whole distributed to well nobody this is only the research side.

[00:02:13]
Then topic wise we have been focusing recently a call for the R. and D. to the tune of half a 1500000000 euros on the set off key technologies that we feel. Important for the future starting from cybersecurity of additive manufacturing and certainly artificial intelligence and if you talk future for automation one of the benefits that we have an organization is that they can always take across technological and cross business look at these these topics so we can easily pull in expertise from type of security for example and see can we use generators adversarial networking and type of security to make impact there.

[00:02:59]
How we do how we do our innovation is to start a low maturity levels collaborating with a DIMIA to solve the hard questions where we need help questions that we can add on saw a lot of those then we do the applied research up to the prototyping and then we help our businesses to bring that innovation into product innovation or sometimes business and process model innovation but it doesn't have to be something tangible sometimes the optimize business model.

[00:03:30]
All right finally finally we are actively engaged with many government agencies and Department of Energy the Department of Defense the H S The whole stuff smaller more focused government entities to do basic research primarily So let me start with a couple of motivating examples here the perception that the most leverage in particular the most of the stuff will be focused on business but.

[00:04:02]
Without loss of generality many of you certainly very very well of way off the impact that automated perception has made in the past couple of years of. Thought one of the driving drones and lots of applications in the future. As an example. You can take a low resolution input image and generate the high resolution.

[00:04:29]
And it's kind of Fiesta I style. Image generation last year of a song that you can do this from scratch with just truly astounding so you can from 0 generate novel data that has not been seen but this consistent with reality by any means you can come up with.

[00:04:51]
Certainly if you look close it doesn't always work so. If you look here right there is the wrong something something really wrong with this maybe a guy doing the tennis racket and the system picks up on the wrong thing and then you get something that doesn't really doesn't really work so it's not a failure safe final example of something that the that is from Siemens was just the nonsense of term of the 1st autonomous train really a train that goes between cities fully autonomous of the.

[00:05:22]
First and I go back to this example later on to illustrate some of the some of the aspects All right so I mean I want to talk about scholar state of what is the what is the problem so when you think about automating these things and never getting in the real world.

[00:05:41]
There's typically uncertain conditions why this is very well known to you and this is the whole of problematic modeling and so on so you never know what you are what you're going to get. A corollary is that there is like you know 100 percent performance so we can only strive to improve as much as possible and there is the question What is the up a ceiling of where we can go no matter what technology.

[00:06:03]
Basically comes down to you know we can we can do it but how well can we do it this is crucially important in the real budget doesn't suffer us to show OK I've done it on this but how. Can I go. So we need to talk about performance evaluation this is a topic that important data in general can help in all of these and these topics so also from a historical perspective particularly for vision if you look back how we do things has changed dramatically about 10 years back and 20 years back we were modeling separately LOL mid to high level vision handcrafting a lot of things in very getting good performance but typically not better than the new and far removed from human in most cases so this is changed dramatically where we are now starting to get better individual Toth than the average human not quite as the best human but as the average human consistently and the way that we are building these some systems has changed right we're talking more about modeling data and modeling as a group of them's and then throwing it into a solve and getting a solution out that means again data becomes of primary importance OK so now let's let's buy the premise that machine running a state the hungry all right so what is that if you can discuss that is almost almost trivial right maybe I don't have data at all and you cannot attain it even though a couple of examples of practical examples where this is the case and.

[00:07:30]
Practice. Then even if you have to pay to Labor Day to May be Scott's OK So quality data and you may talk about unsupervised a week supervised learning in this case certainly but even then the important data may be scarred So even if you have an abundance of data in terms of the raw numbers you still may not be any closer to the objective function that you're trying to optimize so in order to understand this let's briefly look at where you need data for right everybody from area well that you train your model.

[00:08:10]
And then during runtime you feed and data. To use the model so typically what happens is that you have some underlying latent target distribution you get some data and then you try to find a process that matches your estimate of the situation for the target distribution what can go wrong tool to pick the things that go wrong that you have a systematic bias so you're estimating this systematic systematically wrong model.

[00:08:36]
Or you have rare events basically where there is something at the tail end of the distribution that you're just not capturing very well in your data regularizing well in the optimization to capture this aspect of this brings me back to the example that I brought up earlier in the one of the press releases on this presentation there was this nice passage where they say that half way into the on a windy day of the tracks and imagine people staged where an employee pushed the pram onto the walkway across the path of the truck so what there was a guy pushing the pram right and that is the long tail of the decision they had to do it that way because it's hard to even create an account of the situation in real life so you had to create it and put a panda just to show that what they're doing is verkaik.

[00:09:23]
All right So to summarize the main message and if there is anything that you take away from this talk of the slide not that all data is created equal and certainly the aspect of Scott City you have to understand in terms of your objective function from a base in perspective this is very clear you're doing a cost optimization or cost minimization depending on how you pose your model for example minimizing the likelihood of person into an ultimate autonomous vehicle or right even if you have a lot of data it can be scarred from the Basie instance so as a concrete.

[00:09:59]
Lead up to a concrete example if they want to detect a critical event let's say we have a roughly estimate of natural prior incidence of the millions $1.00 and $1.00 in the millions so it's very unlikely to appeal so you need millions of examples just to have a few examples of HIV positive event that's already a problem there is nothing you can train you can barely barely even test and there is many more things to be said about the situations when it comes to false positives.

[00:10:25]
Let's have a concrete example that is now a product that you can that you can buy from Siemens that's deployed in multiple locations world wide to solar insolation prediction variable to predict solar insolation meaning the amount of radiation that you get from the sun and the time rising of minutes but 5 years back that was a largely unsolved problem so we thought let's take a fish eye camera put it on a building let a scan the horizon.

[00:10:52]
The sky from horizon to horizon one image every 2nd and then we see if we can predict when these situations are happening so we have to solve way out of the clouds where the cult moving to. Where is the center of respect to the to the sun. Impact us that we can vastly improve the performance of for example hybrid Pollak plants that use a mix of solar and fossil fuel sources now the challenge as it turned out very quickly in terms of the cost function but.

[00:11:22]
Most of the time to the tune of 99.99 percent nothing of interest happens only a very few of the time something interesting happens but when something happens it's very very costly All right so that is a practical example of where the situation happened so we have easily tens of millions of state and we can get hundreds of millions of data but we still only have $10100.00 at most of critical events and how do you train a model and how do you evaluate a model so that you can problem is general as a show in the in the read about.

[00:12:00]
Is is an issue so this is how it looks like this is an easy example of prediction. At the bottom you see the prediction this is a much hotter example it's a very challenging physical process and we have very limited tops of ability because most things happen and infrared and we only observe R.G.B. color but it's good enough and we're meeting the objectives finally another example of that I'm going to talk a little bit more in a minute about.

[00:12:32]
Recognition of thing that happened in a 2nd so now what kind of approach of have really been following these this is not to say that this is the only way to go about it but these are motivated by a mix of practical considerations technical feasibility and opportunity Also sometimes you have an opportunity there with some nice idea that you can test and then you just do it.

[00:12:57]
First of all using a different domain these are topics like the result learning domain transfer. We can use priors maybe we have some data that is not directly related to the task but there is knowledge to be learned from data that is not directly related and we want to understand how can we use this knowledge to prime the system or regularize it so that it works better.

[00:13:25]
We want to use human knowledge can we involve the human in the difficult decisions but without all the burdening the human so we have to be able to ask the right questions at the right time without asking stupid questions all the time basically and certainly using incomplete data as I mentioned earlier week and unsupervised learning and I'm going to hold 2 examples concrete of what we've been doing here.

[00:13:53]
Just to. Illustrate that we are taking this very seriously for some of the topics here the stuff the of our recent publications and you see the and I see the and these are all driven by this desire to address problems with. Essentially and they all highlight different aspects of that.

[00:14:16]
I will skip through this for the sake of time but these are 2 examples of. The work that we've been that we've been doing. Domain adaptation and. Attention modeling for the human in the loop. And the 3rd idea as I mentioned learning from humans in an interactive way basically the artificial intelligence system understands what to us the human At what point in time to make headway into a drawing objective but I won't go into much detail about the Shaw case studies and the rest of the talk.

[00:15:02]
To highlight how we are approaching these problems if you have any questions by the way during the talk we'll have time at the at the end of the discussion but please feel free to interrupt me at any point in time so the 1st case study this object recognition for.

[00:15:25]
Spare Parts of us the way it looks like. There is that. Siemens biz trains OK we have a couple of dozen main train modern each model have various variations depending on the market in Dubai you need I'll reflect the last glass of Stockholm you don't need that and so on each of them that has various instances it's all what it comes down to is that the spare parts catalog for mobility from the author of hundreds of thousands of pods that you can all the the problem that they had when a train comes in facades us which they do on a semi or annually basis that it takes between an hour or 2 a couple of days to just identify what is broken and what the stock auto number of the POTUS which is a significant loss of sufficiency in the supply chain management and we said OK we can let's listen can we automate this.

[00:16:27]
The idea would be that you take some device you go into any train take a snapshot point to something can you tell me about pot that is from wells many pots. Which will to have a pretty good impact so all. We ask him how many parts can we possibly see the 10s to hundreds of thousands OK it's a challenge we asked how much data do we have all much data can we get on someone that's no way that you can that you can data it's not that you can't get into a train and to do a reconstruction and scan everything it's that it doesn't scale you couldn't do that for all the trains it would not it's not possible to scale is that it would be too much effort OK So if no data at all it's always said OK we actually have some data we have had data All right so we went to the to the engineering department and told them give us a cut data these are huge right that's the whole train and then we started building perception models.

[00:17:29]
From the cat data to solve the problem and. This is about the modeling for example. You. Have to think about what does a sensor do. What kind of viewpoints can you take It's not an easy problem but it's fundamentally doable there is a couple of challenges anonymous catheter that are not obvious one of them is You Don't Know What You're Missing Young just a prior it's sometimes in a dataset you have a prior That's a lot of this but it's implicitly there such as elimination distribution and vision.

[00:18:08]
And viewpoint distribution so we're missing these so we have to supplement by human knowledge for example or other ways of doing that certainly you have to be smart about modeling the sense of the opportunity certainly is that you may have an infinite amount of data which means we can very nice things when it comes to bootstrapping and other methods around larning better models.

[00:18:30]
Now go ahead. Right. OK Excellent question. So the question is on the on the kind of representation that is being used at that other could be self-less models parametric models of walks of life moderates and how do you go about choosing the right way so he a. First of all these models typically exist in various sensations depending if you take a geometric or a function of you for example typically these are Sufis models so we started with the surface of presentation from the thought and the engineering side whether built the trains they have parametric models those were a little bit hard for us to.

[00:19:27]
Use in the sense that we needed large scale data so we want to tool. Be able to generate millions of examples very quickly and the way we did that both to describe ties. Seaman's of leading portfolio and doing this with Siemens and X. and how to take a pro metric models into and optimize the script as a nation for rendering purposes for example and then we've built a handcrafted special made renderer that is able to render very quickly a large data so that we can generate them so.

[00:20:04]
By the objective of generating large data very quickly because this eventually was a sample based approach that we needed to generate that will generate a lot of sample and that would have been nothing gained with the voxel representation All right so this is all about how it looks now nice device to take a snapshot and then it recognizes that there was a button that the thought of that.

[00:20:30]
Part and it brings you to this humans they call it racism all of which is basically the. Operating Web site for the Siemens. This was revealed at a global trade center the vaccine product launch in 18 we have a bunch of publications all the uncertainty many many patent applications and now we have an opportunity to bring this technology into other domains because this is a very generic set of tools certainly.

[00:21:04]
All right so this was the 1st case study practically what you what you can do if you have no real data and there would be another couple of lectures on the practical issues of really making this work from the train model into the into the real evolved. But the outside the gold a little bit because I also want to talk about another case study.

[00:21:30]
Which is about compositional Visual Concepts so a state of the art already. Ostia that's been you can take an input sample and then apply a transformation of that sample for example you can take a daytime image of a female area of an outdoor scenery and you can transform it into a nighttime image so they have this idea of switching concepts between day and night making it a image and I'd imagine item it's at the image.

[00:22:12]
This is then how it how it looks like now and so these are the the output you're synthetically generates a condition based on the on the input and the training process you can do this and make sunny days rainy and rainy days Sunny you can switch between summer and winter and so on right you get the.

[00:22:33]
Idea and particularly interesting that's been real and rendered image of the certainly of a future interests of the entertainment industry. So this is already pretty nice and pretty impressive so now can we use the for the data Scotty problem and say we take one sample and apply all the concepts.

[00:22:56]
That would be great because then we can generate for free Almost a lot of meaningful new data that isn't fien and up to the limits of how we can optimize. Drawn from a natural prior so it's not too biased let's assume that this is the case so now let's look at how that would Burke right let's assume we can we can do this for the visit concept and here we have.

[00:23:23]
That there be a few father not domains winter nights winter days some one night some days we can land the concept between day to night and weekend on the concept between summer and winter OK Now what that basically means for. Concepts we can generate. New example so how about applying multiple concepts at the same time so we can generate an exponential number of new samples so if you think about what's adding concepts would mean which is also grades excitements promise or free free data.

[00:24:04]
Certainly there has to be a catch right there is no there is no free is free lunch and very very clearly the catch is that you also need an exponential number of data to train such a thing OK and we started asking OK is there not. A middle ground where you don't need that much data but this can achieve some of the the compositional.

[00:24:29]
Concept space. It's all the problem is that we need exponential data so we wanted to look for can we learn the concepts individually but so that there composable you know composable in the sense that I can apply one concept and then apply another concept and what I get is a natural example not something that looks off.

[00:24:50]
And I can do this by a different pots possibly. But give me the ability to express and learn concepts with the potentially sub exponential number of data certainly the economists that I don't know made of brick or it may not work well if even if it works sold before I go into into a hobby we've done that a very brief excursion into a general to this area and that works unless this is entirely clear to everybody.

[00:25:22]
Do it very so that I. I did very briefly though the basic idea generated a fair networking is a very nice optimization process that allows you to build data generation schemes centrally the way it's being used the underlying idea is that you have an interplay between a generator and a discriminator whereas the generators roll essentially is to create fake data and that's a company supposed to create realistic thoughts money and the discriminator supposed to tell that false money from real money a pot discriminator has some real money as a reference that he can use to along his own model of what real money looks like and the generators privy to understanding how to discriminate does the determination if the data is real or not and it turns out that you can pose this and in the larning and optimization context and build generators that are able to generate jobs for really good money basically to the point where the discriminator cannot tell them apart anymore if optimization succeeds that may not very much an open topic where it succeeds and where it doesn't about this is one of the one of the very visually appealing places but this has been used as a form before image translation way to take an image of on a main input is another domain there is a couple of examples here again where you take an artistic style from money and make a full to over the fit in the same attack or you can draw the brush to hospice and so on so let's look at very briefly how this image to image translation is done with journals of authorial networks so you have 2 domains and by the.

[00:27:09]
Generator and of 2 minutes into play basically you build a model that is able to generate from the domain of X. of all images with the bra such on the right into the domain of Y. into the set of images of horses. And the same way back and these other discriminator saying yes this is a real image of a zebra and this is a discriminator thing yes this is the real image of a horse the way that this is done this via cyclic consistency basically essentially saying OK if I'm able to do this and I apply the generator here from X. to the limited Vi I should get an image of a horse that looks natural to the discriminant But then if I apply the hypothesized inverse generation I should end up back into the space of zebra so you close the loop and you don't have to worry too much about the middle ground because all you have of the broth so here this is about the the cycle consistency last basically where you start here move over here in this domain and then move back and then you would require that these are close to each other meaning you successfully traversed the concept and inverse while ending up at the same sample and certainly you can do this the other way around as well as start in the other domain follow them move back what and it turns out that this works fairly well on a on a wide set of problems.

[00:28:33]
Now coming back to or to our point of learning more concepts so this is one concept right if there's one concept of the brand translating the protocols and back let's say we have one concept which is color images of handbags and line drawings of handbags this is the Phoebe and color and line drawings of bags and then we have another concept which is color images of hand back to the color images of shoes so we would like to create different apparel like shoes but the same texture characteristics that the underlying idea now there is only data available before these 2 so I only have to say to that looks like the pads or unpaired data what I don't have this line drawings of shoes.

[00:29:19]
I don't have any data that would allow me to goal from any of the states to the to the state so now if you leave leave apply existing tools basically which you but you can do in principle you won't end up with very convincing results all this is going from color color facts to edge bags and then to edge shoes and you'll see it on get very convincing results and what we've done basically is to get more convincing results and I'm going to show minute.

[00:29:54]
How we're doing this with cyclic consistency so let's look at a slightly different example here the 2 concepts are. Having eye glasses on or not having eye glasses on and having bangs or not having bangs. Front and now let's say we start at no eyeglasses No no bangs so we can train and we have data with eyeglasses no bangs.

[00:30:21]
Which is this the main here and then we can do the state of the art thing and train a model to traverse here and then we can train a model to traverse he up adding adding banks so the states with banks but with no eyeglasses So the question then is we don't have training data for eyeglasses having eyeglasses and banks here you can also see the concept of the compass and compositionality would like to learn to put on eyeglasses and banks so that it's a meaningful.

[00:30:49]
So how do we how do we do this. First of all we impose a similar cycle consistency constrained in the early of our only that is a LONG AS I CAN all over now we cycle through a domain where we don't have any data now there is no data available for transitioning between those and those yet.

[00:31:12]
So this is essentially an unseen domain which tells your body from the start this is a very very challenging optimization problem because things can go high haywire and why would we think that this is meaningful a good question. There is reason to believe that this is meaningful because we are doing the cycle of consistency so we are moving here moving here and then moving back and moving back here so this is a regular a zation that you get of the space that propagates through these through the through the States and its entirely reasonable to assume that the short version OK this is a cyclical this is a constraint you have a fall phosphates and.

[00:31:53]
Then we introduce another constraint that was not there in the earlier work which is the community commutative constraint basically if I start with a sample here traverse through the concept this way or if I start here and traverse the concept this way I should end up with about something very similar it would not matter in which all the I apply the concept which implies Congress had to finesse and this is the this is the other constraint that we have that we impose certainly if you look closer that things may not become interested in a child for certain concepts may not be committed meaning we would not be able to do it that way but to an approximation many things turn out to be competitive.

[00:32:41]
So if you do this again you basically get suddenly better at the sampling this is an entirely unseen domain so we're able to predict how that looks like right and you can close the cycle as well so that this corresponds to the to the States now this is the state handbag color handbag line line sure color sure and then back to the 1st base was the fold around the space here is the same as here but in the naive way you though you have no way of enforcing fact that consistency so you get something vastly different at the end whereas if you impose cyclic consistency you get something that looks like the original domain now as with many of these methods.

[00:33:26]
Generative and. It's not horrifically difficult to. Analyze if they're doing something meaningful there are a few good direct measure of the performance of such a model and. Basically what we. Use them to evaluate is a sorrow get problems or we chose a particular problem. Of face identification to see if we're doing something meaningful So basically we took up database of face images.

[00:34:00]
Augmented them with a learned concepts. Some of them are the culmination of concepts unseen. As you can see here and then. We have a Later on a on the sorrow get tough basically and we could show that we can improve the performance of faith identification or one shot faith identification.

[00:34:29]
Significantly Now the idea being involved with faith identification you have a sort of a face maybe without glasses and then you want to identify them half a year later they start to spare and so on and if we augment meaningfully then you should be able to improve. The performance so it's.

[00:34:51]
Most of what I've shown is on 2 concepts. The way that this is formulated it's natural to extend this to multiple concepts you basically have a graph structure of the concept space so what I showed earlier the graph this is one of the subgraph here followed subgraph these are the green the old Saabs are just the non-green the ones.

[00:35:21]
Observed and then what we have done basically is the tools to infer these molds or to be able to generate samples from these from these domains for further use but if you think about that there is a natural graph underlying this which is the hypercube depending on the ups of observability distribution in that graph which is the green olds here and you can infer the other nodes as well and this is how multi concept.

[00:35:48]
Learning would work. Certainly you can't escape the matter that you're trying to solve an exponential problem with such exponential data so there is something to be lost and it's not entirely clear what the. That certainly has to do with how far you are removed from the observable nodes the modes that are immediately adjacent to observe some of the more confident once you go farther away of the secondary in Florence becomes harder and harder because it's father removed from the data in the Field Poll cyclic consistency here to one of the notes so there was this whole question of how do you optimize this and it's right to question and thought practically we've gotten to end it with $33.00 concepts and it's worked reasonably reasonably well but if you add more concept and at the exponential more data it won't work anymore there is an Certainly an interesting perspective from our side we typically don't care too much because we can often design the kinds of concepts around the particular problem that we have and we only have to be as good as the application requires to still have impact in the business we have to know I'll be Treasury performance that we have to achieve sometimes that sufficient to be.

[00:36:59]
To be 90 percent and then we're done and we don't have to be too strict about information theory information theoretic balance there. All right so. Just finally this this yes if you are we are going to virtual. On the on the topic with a good set of speakers and we are aiming to have a 2nd and sensation of this workshop at a conference latest 2019 to also get the community more engaged and structure.

[00:37:35]
Views and serious undertakings in in this program because it's a very very practical and pressing issue for us the number of problems that we can solve with AI and the impact that we can make in practice in all the different businesses not only in seem uncertain about all across the globe it's growing fast at the moment in the data it's growing unfortunately that's a huge issue and we have to catch up much more and we hope that we can do our part in the community to foster this kind of thinking a little bit so finally.

[00:38:09]
We are located in beautiful Princeton New Jersey. Shots 2 and a half hours from here and we have openings typically year round. At any level food and food times and as I showed earlier in the thick nations are not limited to perception vision perception and their P.C. but across a wide range of domains including.

[00:38:38]
Automation and autonomous systems and Siemens has a strategic partnership with jobs at Tech which makes it easier for us to collaborate. It's father talk. Thank you for your attention I'd be happy to answer questions but 50 feet free to contact me at any point in time as well.

[00:39:18]
So. So much. He. Went learing. Could you just for the students in. The context of. The selection of your product because OK intuitively you think it feel. So much you think. That's right more than we really aren't very. Bright you may be right it's quick and. You may not be sadness it might be.

[00:40:08]
So kind of curious why these particular because that OK you know. Using that base is perhaps what we usually see when it really looks. Like. You could definitely was an only. Son but could you explain how you selected OK so. So there are a multi The thank you for the question there are multiple effect for that the 1st of all how we select the problem is generally always wore work from the customer back what's All right so all of the threatened by some very specific demand now or in the future so we do a certain amount of projecting what will happen in the future OK Now to your question about the type of gossipy it's very important to understand that we do not have abundant data we may have a lot of data but that's not good enough.

[00:41:08]
We actually around 2012 we won on the planning hype both and the very fast upswing we've been back into the past 1020 years and. The main problems that we had been working on and had we had enough data to a train machine learning model beyond the as the EMS and so on back then there was very little we wouldn't have had enough data on any of the problems it's really really really hard to get to get they talk.

[00:41:38]
And these examples feel well select that. Also highlight the point if you recall the spare public admission problem it's a very few minutes problem right trains and they reduce obvious and service is very important trains have a very long life cycle they have decades of life full service is where we make a lot of business.

[00:41:58]
But there is no way that we can even get data even if you wanted to perhaps in a number of Yes Basically once this is deployed we can secretly kept data but even then there are legal issues you can't just do that now and there is no way that you can scale data acquisition to go all out in the sense of a few study.

[00:42:18]
And get that data sometimes it's need to constrain sometimes it's technical constraint sometimes it's just seeing funding constraints sometimes it's physically impossible we have placation swear it's physically hard to get the sensor in if you think about a gas turbine for example this is like. $1700.00 Kelvin regular hardware will just melt so if you you can't do that you can scale that arbitrarily.

[00:42:49]
I think one of the one of the aspects in there is that as an organization we have data from various the mains and this is certainly of interest for us basically to see can we aggregate data from different domains that may not be related immediately but can we abstract the knowledge in the service of knowledge representation and knowledge graph to transfer it to other domains and learn better there this is certainly a strength that we can play and this is something that we're actively pursuing.

[00:43:25]
This. Now we certainly do have a competitive advantage in terms of data but it may be slightly different than what you are but you would expect So for example with the spare part recognition again here we have the models of the trains and you're not going to get thoughts.

[00:44:08]
It's only that it's not directly the data that you would need to dust off the shelf train train a system if you want and that and that goes beyond that right we have all of this engineering data and knowledge wealth throughout the organization at the moment we don't have the tools to transform that into our production model now.

[00:44:26]
But it's not that we have secretly data bases lying around of a particular thing that would make the task easy I mean it certainly makes this is certainly the case what I'm talking here more specifically about to pick a vision of locations. Let's say you want you want to on the you want to do condition monitoring and gas turbine that's a very different.

[00:44:49]
Way of gas turbines these days a modern gas turbine this instrument of a couple of 121-0000 fans or each measuring a particular quantity and that data we certainly have and to capitalize on that this is mine FIA which is the law too for integrating the small it's from from the fields and then we can call it in mind a top to improve the performance of a gas turbine based on conditions on the top data or predict outages this is huge value.

[00:45:19]
And there we certainly have a lot of data that strew for a set of other domains as well but not in vision so much. So we can take this offline Perhaps I can I can give a very general or the other question both on the human. Interaction the things that our products naturally interact with humans and for entirely non-technical reasons things like trust and acceptance we need to understand what role the human plays because often we have a perfect technical solution but nobody's going to use it because they fear that it's taking the jobs away or that somehow less useful to them and we have to understand that and we would like to do this and I must achieve chick fashion.

[00:46:26]
Because in the future it seems. I will essentially be able to predict and understand also about what the human wants and does and that should be beneficial overall that is the the overall perspective and there are many many subjects challenges that you can that you can put there and opportunities.

[00:46:48]
To. Sure so. There is a couple of failure patterns in this in this case right one this one is really. More from a from a from a social perspective right where people feel threatened and they're not going to that are going to accept that they are going to silently supple Tajh whatever photo you do or to bring it into the feud.

[00:47:31]
When automation is seen as a threat to the well known. Behavior. Starting with the industrial OF IT revolution but this is nothing nothing different if you think about what happened in that industrial revolution people are afraid of AI and we have to communicate properly and communicate the benefits and also also focus on the benefits that this that this brings that is one common failure pattern which has nothing to do with technology the other very common failure pattern is.

[00:48:07]
Such if you cation and legal. Sometimes you have a liability issue and very specifically if you're if you look at the people. That's a look at a gas turbine that has to be serviced regularly gas turbine failed it does all very very very spectacular to the order of reducing a building like the store rubble in microseconds it's very painful so this is very important to do that and the people that service the gas turbines.

[00:48:44]
Fundamentally responsible to sign off at the Cato continue running and there are many things that can go wrong so they have a whole established set of certifications for these people level one level to level 3 right where these people go through years and years of training to. Achieve a level of certification that they can find out yes the scales haven't worked now you come in with your automation and you show OK I achieved a level of accuracy compared to the to the expert it's at the moment very very hard just based on the amount of data that you have to even show that you are statistically certain compared to an expert that you are spot on with that result you you know you have a hard time showing at the moment the quantification even if you have a perfect algorithm it is hard to show all in a convincing way that you're at least as good as the trained and certified engineer now that's 7.

[00:49:54]
Yeah so you don't have nobody has sufficient data because you would need you would need so 1st you would need. Audits of magnitude more data than a human has to show for because you don't have that trust in automation at the moment generically basically that the 3 liable as you have in the human so you need all of that in.

[00:50:24]
I know what you're asking about things that are hard to deploy this is a question that if this is an example that is how to deploy typically what happens there how we deploy this is certainly that we aim at a lower level right that we don't take the X. I'm talking about taking the X. but all to complete that is hard Well what we do then is to make the life of the expert easier right we instead of having to look at thousands of samples he only has to look at the most critical 10 samples that already that is a compromise that helps not waste of time and we still have confidence in the in the resort so we find a middle ground on the situation but you were specifically asking for situations where non technical on strictly technical issues keep us from deploying the.

[00:51:11]
2 of. You had a question future. Future. Lots of them right. Now. I think. Particularly here in the space when we're talking about generating data there is the there is the issue how do we evaluate that we generating meaningful data that is unsolved and if we had better solutions for for this would be huge impactful of easing surrogate problems there are there other tools that means to.

[00:51:55]
Follow it such models but that's not well done at the moment and that's keeping us from from moving forward as a bit. Right yeah.