my name is Georg Wittenburg I'm co-founder and CEO of in spearmint we use AI to automate data science so thanks for the kind introduction if I have to ask the obvious thing first are there any data scientists in the room one two three four a couple of them so the first section is exclusively for you we'll talk about business aspects and then maybe the more general aspects of this later on but maybe just who to begin with a very obvious yummy very yummy but but please don't be alarmed and data science has has achieved so many amazing things over the past decade we were able to classify handwriting we're able to read emotion from images right we're able to read emotion from from tweets from short text messages we're even able to recognize threats or dangerous in a moving vehicle which in itself is amazing right because you have a real-time situation in which decisions matter and can actually be deciding on almost life-and-death situations so with all of this great progress that data science
has made over the past decade I myself stumbled over one thing beyond these four examples which is this so much some of you might recognize this some of you might not when you open a file in a program like Microsoft Excel and this is a data file that you would use on a daily basis well ask well you know the state of file that you have is it delimited are the fields in the file separated by tap semi colon comma or spaces and at that point in time you I know you you start to think about this like you know yeah like state of the art like motions hand rioting threats in real-time but asking us about yeah it's this is commerce at the tap like this does not make sense right so it's really strange like you know if you ever wonder that that you're faced in your daily endeavors with these kind of like trivial decisions to make while at the same time working on data science or even artificial intelligence so so
did I about four years ago and and so I did the only thing that was kind of meaningful right and I saw okay let's build a classifier for my data files which classifies whether this data separated by tabs we call them come on and as you will undoubtedly understand this was amazingly easy right and the accuracy of that classifier was amazingly high it's not a difficult problem to build a classifier for and and that kind of got me into this rabbit hole of things because you said well you know here's the first classifier let's let's classify the fat let's specify the file format okay done not not that difficult can we can we classify the file structure like can we say whether it is like a measurement over time can we say whether its transaction maybe it's a P&L item from a business context maybe it has some location can we actually classify what the data is about and it turns out yeah like no saying whether a data set is a time series is
a classification well it's not a hard one like you can do this and we build our classifier and when you do that and you say well you know maybe maybe if I know what the data set is about maybe I can excessive phi which analysis are applicable to it like you know should I do a so I draw this on a map should I do this on a time scale should I draw this on a word cloud another classifier and if you got that and you ask well you know there's certain patterns that you can detect in the data can you build a classifier that classifies which patterns are applicable like you know which analysis methods are applicable to an analysis and then can you say you know let's can we detect the patterns so we've been doing that for about four years now and we've built two hundred and twenty six classifier and we believe that with 226 classifiers you can actually automate quite a decent part of what is generally considered what is
generally considered data science like going from raw data which might even be far messier than what I'm showing over here so actually having something that you can put on the table this includes things like cleaning of input doing a slicing and dicing of the data set doing a time series analysis doing a classification for for patterns like segmentation outliers all these sort of things that you would say yeah this is what is a data scientist spend my time on you can actually build quite a good classifier for many of these and in fact it was in September last year that for the first time at least in our company history one of these automatically generated results was put in front of the board of a company say ok so this is an automatically generated result from a automated data science process that real board members actually talk about that's interesting for for data science because it changes the role of what a dead data scientist will eventually end up doing if you as a data
scientist are putting things on a timeline and chances are that you won't be doing that two or three or four or five years from now this is a workshop setting that we had about two and a half years ago at Deutsche Bahn the guys on the left on the very right those are data scientists the two persons in the middle those are the business stakeholders that the people that actually care about what's in the data right everything that's on the table was generated by a computer and these guys are having a conversation about what they care about what could be conceivably correct or meaningful it's interesting because it changes the process all of a sudden the business people have really easy access to say well this is something I care about and the data scientists have really really easy access to say oh that's good to know right so real-world conversations that happened in this scenario include the following two data scientists on the right hand side said very interesting problem that you are describing
here I'm afraid the data cannot deliver on that and then when we were in the workshop the business guy said you know the problem I had the other day this thing over here is quite quite a good approximation can you add one more thing and then it will work for me I said oh I didn't know you meant that the other thing that happened in that very workshop is that the data scientists on the Left said oh this is something that we've known for two years and the business guy said well why didn't you ever tell me it's really relevant for me but you never told me all of a sudden it's more discussion about the results of the data science process is supposed to methods being used like no do we care about this do we know about this that's interesting right so looking beyond the mere data science process in itself so question to the audience anyone form a more like business business executive background in the audience a couple of them thank
you okay great so if you now say look we've just learned that there's a technical process that can automate great parts of data science that's that's interesting right because one of the key problems that we see many business having right now is that it's not that they don't have any data that but the problem is that is is to get value out of that data and may and data is growing it's growing about 40 percent per year so you have more and more data you still say ok so what is the one needle in the haystack that I need to know about in order to you know make a business decision and it's not that you're not trying from the business point of view right so you're investing into big infrastructure you're listening in to data lakes and say like you know this is what I want to build in order to get value out of my data and and here's here's a target rise of AI that says that basically like all of these
sort of things that are the end result of a data science process so you said these are things that you would take into a meeting and have presented to you can be automated it comes from challenges the status quo a bit so what's what's the first reaction from a business point of view that the one would have to this so here's one this is a this is a quote from from a head of strategic planning at a motive supplier about two years ago he said well if this works I can save the first two weeks of every McKinsey project right he said hey I will just automate the initial analysis have the initial results and then I have less to do for the people that are currently charging me a lot of money for this that's one possible reaction another reaction that we also get very frequently is hmm you know my data quality is really really bad but once again like we talked about the state of the art and data science or artificial
intelligence like now this is a foggy street do you really believe that for instance you're consolidated income statement has worse data quality then the image of a foggy street a self-driving car will navigate that Street right so it will work on that messy data and make decisions in real time about whether to turn left or right so I think like one of the key things for the business community is we need to raise your expectation on what an analysis tool can do even in an inside of lacking data quality so like the big the the big trend here that we're looking at it's essentially the trend of automation right so here's a process that you currently already have a plan to build and that automation has a certain automation potential Gartner says that they see an automation potential in data science for about 40 percent by 2020 what is essentially does is that it says hey right now you have a decision to either put money which is the y-axis or time as the x-axis
against the the process that you have which says okay I got an outsourced team I got an in-house team and that's basically the cost effective curve that I have in order to derive value for my data what AI does here is that it just lowers the curve it's an automation technology that allows you to run the existing process at a higher speed at a lower cost but that's not all of it so this is basically a thing that I delineated in the first part of the top it says hey you got a structured repeatable way for getting from raw data to something that you would take into a meeting but the key thing here is as with so many AI system it really is fun if you have this sort of feedback loop if you actually judge how good the result as we say yeah thumbs up thumbs down on the results that the systems providing because then you can actually teach the system or system can learn from your feedback what are the things
that you care about in your data and this is something that we actually do see in in real-world settings so this is a an example from from a case from last summer for an insurance company in which a team of team leads analysts we're looking at automatically generated results from their insurance data and they said well you know we found something that we really care about and they put these things together okay so maybe this is something that we should use force for steering our business processes but when it was eventually presented to a board member I said well here's something else that I care about which was not part of the things that you care about right so there's different pockets of knowledge experience and also business intuitions that reflect upon the data in different ways people care about different things and knowing who cares about which part in the data or in the results of the data it's very meaningful because then you can train the system say wow know if I need
a board level presentation maybe I show more of this if I need something that's more operational maybe small these results and if you have the feedback loop then there's actually some pieces of data that you can gather from your ongoing operation so finally who hasn't raised their hand so far lovely so III know that that some of you are basically here because you think AI is an exciting topic and thinking on wondering about the ramifications on society like what's the impact of that why how will it affect all of us and you've been sitting through the business side and the data science side of this talk let me tell you a joke here's a man kneeling on the street and saying I'm looking for my quarter that I dropped and the police officer says did you drop it here no I dropped the two blocks down the street then why are you looking here because the slide is better and this is actually a black illogical effect this is a effect is called a streetlight
effect which says humans intuitively search where it's easiest to search you know you don't fall under the table unless you have to you look at all the other places first and only if that's really if someone tells you to if there's no other choice then he will crawl under the table and search there and it's quite similar when you when you search for for insights in your data so this is a hypothetical landscape of of data right and we know that in that landscape of data there's always like interesting things that you could discover but you need to know where to look and there certain constraints on on you looking like you know there's a time constraint if you're doing this professionally you have maybe two days in order to find something which means that you can't look everywhere but you'll look at the places that are sort of most likely to yield a result that you have at the end of your project so you selectively look at this data space not comprehensively and
just to illustrate how how real this is like this is a typical data set that the data scientists may may analyze as part of his or her job and in fact it's one of the most valuable data sets that I've seen in the past five years because it's the data from the United States Environmental Protection Agency in which they tested personal vehicles for their mission data so this is a data set that has an inherent value if you know where the value is if you don't know then you just have a bit of data and you don't know where to start and so what the effect of automating the data science process really does it says well you know let me traverse the landscape of all the potential interesting sites and and landmarks of that data set and let me show the results for you here's one result so this is derived automatically from the prior data set this is a correlation analysis of Nyx and co2 which says that no here's a correlation here's
a bunch of outliers and the outlier up here is - fox wagon - Erick right if you don't know if you don't if you haven't looked for that one you won't find it but if in system can automatically show this to you that's good but because then all of a sudden you've gained knowledge and this is really great thing about automating the data science process is because it kind of allows you to push beyond human intuition it says yes like you know usually I have an intuition about something that I might care about and I have knowledge that I gained when when exploring this but there's actually more in the data there's more in the data that you could potentially care about if you only knew about it and now you have a structured process driven by AI that says look here things that we can raise to your intuition that you can think about and then consider whether you would like to know more about it and that's the true value add of this
particular piece of AI so the question really is what will you find in your data and these are my top 10 or top eight actually from from from the last three years of doing business in this space and well you find in a network of frauds that that are emptying bank accounts as we did for for for top 10 international bank will you will you find abnormal spending of of your external consultants that you hired as we did in the data for a government agency will you find mishandled business loans as we did for top three German bank I mentioned the outliers and vehicle emissions but then there's also damaged in German train station underpaid employees actually or that your IT cloud as is actually located in the wrong country a last one is my favorite someone discovered that their target audience of their online businesses actually women right and the last one is striking here so very poignant in in this because it's not as if the the question was difficult to ask I
can our women are our target audience of our business it's just that these guys that are driving their online shop they never asked the question and they told us about that in the dinner over dinner with also their investor present and the investor was looking at them and disbelief and saying I can't believe you never asked a question alright because it's obvious in hindsight but at the time these business people just didn't think about it so thank you very very much for your attention if you'd like to have a personal demo we are right around the corner in the dead center of the exhibition space and we'd to love to showcase how you can't find new insights and new data using artificial intelligence thank you very much okay go we have six more minutes left so are there any questions in the audience right now before the okay we have one over here in the first row threes s coming and then I saw a second one over there but we will take your first
sir hold on hold on the microphone because also the people in the back want to hear you and it might be thank you very much for the presentation and my question might be very embarrassing because I arrived late but it's very fascinating but what industries have you worked with and where you see see the the biggest or the best applicability of your product yeah so in fact because we are automating data finds there's no inherent tied to any particular industry like you know everyone is trying to get well your other data and the the process of doing so it's not tied to an industry we see most demand in industries that have certain a certain risk of not knowing things all right so that's that's finance that's audit that's just anything that's compliance related but also a factory floor data when they say hey it's a certain risk of things breaking unless we detect them early enough and that's strange enough you also have interest from the logistics area which I can't quite explain but
maybe they they have a certain innovation giving it the entire mobility discussion okay there was a second question over there in the back gentleman over there hello thank you for this great presentation I'm a Six Sigma master black belt so I know how to treat data and and and for parents but my question is how would you convince me as a customer that that you're giving me the answer I'm searching for you know you put the data there you get our answer and how do I know if it's if it's for value for me if I may answer an analogy if you're looking for present for for your wife or a loved one how do you know it's the right present all right you kind of know when seeing it right so it's kind of hard to intuitively say like I don't know what the answer is but I will recognize it when it's in front of me all right and so basically as in traditional data science you have a hypothesis driven approach we
say well we think about the business and data we have and we believe that there's something interesting here and then we go look for it we challenge that assumption what we essentially saying is have an open mind here a bunch of things that are just effectually in your data which we can raise to your attention and then we ask you to sort and prioritize pretty much the same way that you would decide for a good present you say okay so I got three alternatives and this one seems to be the most relevant one or the one with the most business impact or the one that will make my manager most happy okay the gentleman over here any crush' yeah thanks a very interesting talk I would like to have that open source of course well I didn't got a quite get this what's the AI part in the whole part because at least the first three tips you showed us I imagined could be solve over those classical business logic and just peddle and recognition
and then of course you talked about the feedback part the feedback part now extending people around around around the conference table talk about it's very laborious if you really want had a lot of feedback circuits I already planning I maybe even have a ready-made system where the data is delivered constantly people can get feedback in a constant and already made process and yes and in fact that that system exists right now and you can try it in the exhibition area I didn't focus so much on user interface issues in this talk but in fact the system that we were building provides for that mechanism of providing thumbs up thumbs on feedback to individual aspects of the analysis you have that closed feedback loop and then as for the first part of your question the the thing that that we have that we're building it's a very pragmatic way of doing it right so if there's a classifier that we can build through pure business logic right so for instance like the initial example that I
gave for classifying data files that's not a neural network right and it shouldn't be in your network because it would just be like you know too much however there are certain parts in that process especially when when classifying the semantics of data is for instance is this a postcode is this a currency value is is this a location is this a human name is this is this a city space there's all kind of like little steps that you need in the data science process in order to drive to semantically correct conclusion and that's actually something where we use say more advanced techniques from artificial intelligence to build the relevant classifiers okay one more question was in the back I saw it yeah yep I have a question so what's your prediction of population of data scientists in five years is going to decline or increase so I think they're very much in the transitionary phase right so like yes this this is new like you know going on for the past I would say between
five to ten years face on top which is how business is being digitized and data science is a great part of that and it's with any new profession once you passed the the initial phase of it people started to worry about cost they will start about efficiency and also about repeatability of the process right and from the point of view of AI we're not trying to solve data science in its entirety but we're saying data science is an 80/20 process it has a good proportion that can be automated and a good proportion that needs specialized expertise and skills that just doesn't make sense to put into an AI system so with that in mind I think that the population of data Sciences will actually be growing because there's so much data to analyze however I think that the focus of their work will change they will change from doing the let's almost a trivial things at least the things that can be automated towards communicating about these things with their business counterparts and then actually
doing the deep dives that follow from that whereas if you you know if you have a data scientist even right now that is busy doing a time series analysis consider the authorities please