How Machine Learning is turning the Automotive Industry upside down | Jan Zawadzki

How Machine Learning is turning the Automotive Industry upside down | Jan Zawadzki


[Music] all right dance break is over once upon a time that was a giant a giant so big and powerful that everybody would know about the giant people would love the giant people would work for the giant people would obsess about the giant now the giant would have a funny smell a smell that wouldn’t be too good for the environment but that’s for a different story this giant has encountered a little fairy that’s giving a giant a pill with tiny letters written machine-learning on the pill the fairy says if you take this pill you will grow much bigger and bigger but it wouldn’t cause also severe pain how will the story unfold we shall see next hi everybody as you can tell by the introduction this is not going to be a very technical talk all right rather I’m going to try to give you my perspective on a meta-level about how the automotive industry will change due to machine learning my name is Nancy Varkey as Andy said I’m the project lead for karmak we’re 100% owned by Volkswagen Group and we’re basically a software company or Volkswagen Group so I might be biased a little bit towards folks wagon but I’m really trying to paint a bigger picture here this story this presentation is divided into three topics first topic is why should you care about the automotive industry why is the automotive industry in this analogy a giant and not maybe a turtle second the impact of machine learning on the automotive industry and lastly major challenges that cause Peter caused pain to the Giant in the automotive industry so let’s start with the very first topic when I talk about the impact of the autonomous inner of the autumn of the automotive industry and where it is today I want to focus again on three topics I want to focus on the economy unemployment and on society so why should you care that then automotive industry exists and that it basically is in severe pain right now well if you look what you see here in the background is basically a division of sectors in the United States okay so what fidelity investments is done is they have taken the market capitalization or market value of different companies and put them in sectors what is market capitalization if you’re a public company and you have a thousand shares of them worth $10 then your market capitalization is $10,000 and it explains how much the market thinks how valuable certain industries the companies are so if you look do I have a pointer here I guess I can just go front so in the US the automotive industry is important it’s it’s somewhere in the middle it’s a select overview it’s more important than the tobacco or the airline’s industry but it’s also far less important than the software industry if you were to be a CEO of a company today would you rather be a software or automotive company second let’s take a step closer and look at Germany and the impact of the automotive industry of the German economy we always say Germany is a car country and if you look at the market value of our automotive heroes and diamond lab BMW and Volkswagen amongst others we see that the automotive sector is one of the most important sectors in Germany okay I can go on in terms of revenue but I think you get the picture automotive industry it’s important in the u.s. in Germany and globally let’s take it a little bit more personal let’s see which impact automotive industry has in our lives if we evaluate how many people work globally in the economy in the automotive industry we have 9 million people who are directly employed by the automotive industry 9 million people if you take that a step further you can easily increase that number to 50 million people if you take into account indirect employees like people who work in car dealerships people in the car insurance business car wash tops the whole ecosystem around it automotives there will be 50 million people if you look at Germany we find very similar numbers we have about 80 to 800,000 people who work directly in the automotive industry and indirectly between 1.8 to 2.4 million people so that’s one in 50 people in Germany that’s indirectly employed by the automotive industry lastly why does the automotive industry exist right it exists before because it provides freedom through mobility to people let’s look at the numbers how the society has is changing and in general when the automotive industry what we see on this graph here is global car sales and what we can see is that global car sales have increased nicely up until 2015 but in the recent years it least seems that global car sales are somewhere plateauing obviously you can attack those numbers and say once other countries become a wealthy they can get more cars so that would drive that will increase again but for now it looks like global car sales are stagnating while on the other hand what you see on this graph is not too important what you exactly see here but what you have to take with you is that in demand for mobility will increase you see on the graph here different options like ride hailing trains and buses flights the whole mobility ecosystem and it’s predicted that the forecast will increase in the future I think I’m starting here if you study from the global market Research Institute but there are many other story as research papers that say demand for mobility will increase so sales are flat demand for mobility increases ok if we go back to our giant analogy we can see why it’s a giant and maybe not a turtle it’s important it’s a large player it’s a juggernaut basically of the economy now let’s talk about this little pill that was given to the giant the little pill that says machine learning on it I’m not going to go into detail about machine learning do not worry it’s basically a set of algorithms or tool to find patterns and data and it has been around for a long time some people trace back its route until the 1820s when Gauss laid the foundation first with the least squares formula use it to calculate or predict the future movement of planets if I believe correctly but again its mass how far do you want to go back why is why has the Perry given the pill right now to the giant in or 10 years ago or 10 years in the future because of opportunities to machine learning are tangible right now we have more data we have a better infrastructure we can use those algorithms that have been around for a long time on a massive scale so when I talk about the impact of machine learning I want to again focus on three things I want to focus on the products in autonomous driving personal assistants and last but not least on the entrance of new compare it’s our autonomous driving what is autonomous driving it’s basically bringing somebody or something from A to B without a driver there are many forms of autonomous driving can take place one that comes to mind easily is Robo taxis where you can get in an urban environment from A to B without a cab driver research estimates that that could bring down the cost per mile to down to 50 Cent’s where I think it’s right now somewhere between a dollar and fifty so they can lower the cost significantly you can also have autonomous driving for your personal consumption right you enter the highway you press a button and you don’t have to worry about driving on a highway anymore and then another very big part that is sometimes forgotten about autonomous driving is logistics the whole trucking business the whole long-haul business is also an a giant itself basically and yeah another story is you also have autonomous tractors like Agriculture’s huge feeder autonomous mining trucks John Deere had for a decade already semi autonomous tractor around that topic is especially prevalent in the US while you have fields that are so long that it just takes more than a day to just float through one line of the field so autonomous driving touches really anything that transports somebody or something from A to B I want why it’s a game changer I’m hoping I can show that to you on this little project management triangle if you develop a product you want to offer it to the customer often in a short amount of time right we don’t want to wait long you want to produce it at a lost cause low low cost and eventually also sell it at a lot low cost probably and you want to provide a high quality to the user most products are somewhere in the middle here driving today is also everywhere in the middle it can get you from A to B in a reasonable short of time and the reasonable cost and at a reasonable quality what will autonomous driving do project management slight obviously everything will go up so why low costs if we go again back to the example I said with Robo taxis bringing down the costs for driving an urban environment from 150 dollars per mile down to 50 cents that’s very low I’m not sure how much how much lower it can get unless we have a puppy to a movie later just runs forever without energy high quality if you drive yourself and you don’t have to be stuck in a traffic jam anymore or on the highway that increases the quality of driving to a lot of people because it frees you up to do whatever you want and why short time at first autonomous driving will be slower because it has to be more careful they’ll avoid really accidents but short time really relates to if you’re in a logistics business transporting goods from A to B your truck drivers I’m not forced to take a break anymore your computer doesn’t need a break it can basically go all night thus bringing goods from A to B in a shorter amount of time and if a product would just increase either one of those parts that it already would be a unique selling point autonomous driving can increase all three of those faults and it looks super easy right is anybody of you know the story of George Hotz George has cool so I’m going to let this video run in the background George shot she was a famous hacker with the hacker name of Geo hot who hacked the PlayStation 3 at the age of 14 and the age of 26 he hacked a Honda Civic and what he did he put a video in the front in the front window a camera that will record the front window applied a machine learning model to understand where the where the lane is and then bought a $30.00 signal in trajector that what gives signals to the Honda Civic on how to steal and how far to go as you can see here somebody there’s no intervention in the in the wheel the car basically drives itself look super easy he founded the company karma da day I was valued at billions was the net was supposed to be the next Silicon Valley unicorn it was set to launch for 999 dollars a self-driving starter kit so that everybody the autonomous driving would be ready for everybody that was in 2016 how will the story unfold again I shall tell you later and then we have those online degrees this Udacity has an online course for self-driving cars since 2011 since eight years all the company basically has to do is take his whole entire IT department make them watch this online course and they will know how to develop self-driving cars super easy right so something proclaim that’s about topic of autonomous driving if you think it a little bit further some people claim that autonomous driving eventually becomes utility if you look years into the future it won’t be a differentiator anymore to develop machine learning algorithms I like to talk about the Holy Trinity of AI where you have data infrastructure and algorithms to make machine learning work and each of those three parts of the Holy Trinity of AI can be out sourced eventually market forces should force the prices down that will be coming utility and available to everybody what will then be the key differentiator that could be personal assistance what’s a personal assistant we know personal assistance from our Alexa at home Siri on the phone Google home whatever those devices are it’s basically a machine learning algorithm that knows you and tailors the services for you and where I would love to go with personal assistance is getting away from all those tedious knobs and turns and twists that you have to do I still can’t figure out which which knob actually heats up the front window and the rear window they all look so similar to simply go into this part here right where you can joy you’re right in the car without having to interact with the car anymore you can interact with it through voice I’m lucky enough to go to the Silicon Valley at least twice or three times a year because we work together with the future Innovation Lab and Volkswagen and when you talk to investors and VC’s in Silicon Valley there right now concerned about they are finding the next platform the platform to build the next billion-dollar unicorns internet was a great platform that gave rise to dragon house like Google and Facebook for instance then came mobile as a platform mobile gave rise to Airbnb and uber so shortened response times that they can really make a billion dollar valuation worth it what’s the next platform ai is not a platform it’s a set of algorithms some could argue that the car could be the next platform if you have all the time why you’re why you’re transitioning to to work if you have all the time in the world that could be used to purchase digital services in the area of gaming virtual reality entertainment you name it so if we go back to the analogy the fairy has given the giant disappear and the giant understands now what the what the pool can do but should the suit the giant really take the pill I mean everything is going fine we’re selling millions of millions of cars globally everything is going smoothly but then a giant sees on a far horizon a little baby giant a little baby giant that has been feasting on those machine-learning pills and feasting feasting feasting and the giant c-czar oh this this baby on the horizon it can’t walk yet it can’t even crawl but if it can walk and crawl it was smashers so let’s talk about why this is important right now and I have just focused on the products and ranjeev famous deep learning luminary likes to say is a new electricity I haven’t even touched on topics that Stefan from BMW and Stefan Stefan from diamond I’ve mentioned before machine learning will touch so many other parts like smart production and visual inspection why is it important now or not just a PowerPoint slide if you take anything from from this presentation with you I want you to think about the next two graphs what do we see here we see again market valuations market capitalizations of traditional automotive companies and of new entrants of the babies that we’re seeing on the horizon that have been feasting on machine learning on the very top you see way more uber and Tesla let me say a few words about those companies way more spinoff from Google the famous German engineers Abbas and tune together with some Stanford engineers won the DARPA challenge self-driving car DARPA change in 2006 and since 2009 Google has invested more than a billion of dollars in its self-driving car unit it was incorporated I think in 2015 and launched his first commercial service at the end of 28 2018 last year it focuses on creating level 5 high-level autonomy that’s his focus it’s focusing on machine learning to make self-driving cars a reality it’s it hasn’t really made a lot of money but investors are valuing it anywhere between 175 to 220 million dollars where does it come from it comes 80 billion is accounted for the Robo taxi business 90 billion the potential for logistics that I said before and 10 billion towards outsourcing the stack is a software stack that’s more than the combined market value of traditional OMS more than folks BMW and Daimler combined what’s with uber uber is the right Hayling company it’s set to IP all become public this year and its values also ranging somewhere between 90 to 120 billion dollars on the contrast you see here the number of employees that it takes to generate this market value okay Weimer has just barely cracked a thousand employees last year uber has a thousand 500 employees on its self-driving car unit 16,000 employees in general Volkswagen on the other hand has more than 600,000 now I don’t want to be unfair this is obviously this could all become a huge bubble right if self-driving cars don’t check out if machine learning doesn’t work this bubble will burst and Volkswagen BMW and diamond are there creating cars right now I think last year they shipped more than 16 million cars while Tesla for instance has focused on just as has problems producing more than 5,000 cars a year but if this checks out that’s what I mean the giant baby will be ready to smash the real giant okay so the giant has understood now okay I have to take this but I have to focus on machine learning and it’s causing severe pain and I want to tell you about those challenges and then at the end give you a little bit something to think about so again I’m focusing on the Holy Trinity of AI data infrastructure and algorithms and I want to take this next 10 minutes to talk with you about those challenges it’s our data that’s my favorite topic we can talk about data all day I think it’s the right place also to talk with people about data if you say that data is in your oil you could argue that the automotive company is geographically the new Norway Venezuela or Middle East because it produces such a plethora of data of the most valuable resource in the world we will see how it’s managed but it’s a different topic so let’s say that the self-driving car produces up to 10 terabytes of data per day what’s 10 terabytes of data we have megabyte gigabyte terabyte if you have a phone and it has let’s say 256 megabytes of space there’s 40 phones 40 phones I can stack them up probably I’m up until here produced by one car and one day that introduces and new challenges challenges that really haven’t been seen before in other industries let’s say you want to see how many selfies you took three days ago you have to go through all of those 40 phones check all the pictures you took three days ago and see how many of those four selfies it produces large challenges in terms of data management data search ability data visualization the whole groundwork and then what do you do when you you need those sensors to collect data you need this data to create the machine learning algorithms to make self-driving cars or eventually personal systems a reality what do we see here we see the amount of miles that have been collected by Tesla and by way more wayne was correct more than ten million miles last year Tesla is more than 1 billion miles already driven on their autopilot and what you see here is that they have an ice hockey stick curve I couldn’t find detailed numbers on OMS how many miles they have collected but they are collecting it and it’s real and it’s not just about collecting any type of data if you stand for five minutes at a red light there’s limited amount of value that you get there what you really want is to find the data and the spots where your machine learning algorithm fails that’s why Weimer is doing a funny thing their CEO sometimes dresses up as Big Bird or as a skater or as goofy and walks in front of the car and obviously in the real world goofy would very rarely or big bird would very rarely walk in front of a car but it’s still checking if the car would stop because hypothetically if Big Bird were to drive in front of the car it should stuff so thank you way more CEO for doing your job next we’ll talk about infrastructure you don’t have to look at the number see too closely what I want to give you is creating the sensors the eyes the the hands the feeling of the giant the hence the sensors of the car equipping the car with sensors is expensive the most expensive part is lighter the market leader Lida favela dine lighter cost anywhere between seventy to ninety thousand dollars what slider it’s laser emitting thing that usually turns on top of cars and it’s very good at measuring the distance okay some people Tesla says they don’t need lidar because they say which humans have only two eyes we can drive perfectly why should the car need lidar so they basically taking light out of the equation but the consensus is right now that you need lighter to estimate the distance if you think about it lala i don’t know how if you’re like eight meters away from me or nine meters away from me for self-driving car to make it really safe we should know as far as good as possible how far an object is really away from us Weimer has gone away that they can reduce the costs by lighter by ninety percent is what they claim I think they also want to sell their light equipment soon but the thing that you can say here equiping cars with sensors for self-driving capabilities it’s very expensive still experts say that you if you want to buy your car the consumers would be willing to pay between eight and fifteen thousand dollars more for self-driving capabilities we’re not there yet but you will see on the next slide is that traditional companies as we’ve heard before come from a traditional background it’s a historically going background and what you will see next here are different electronical components and how they developed over the time I think over 40 years span so this is not an org chart that’s a different topic I’m not going to talk about that but this just goes to show you how the electrical component landscape has evolved in the car and even if this would all be beautiful microservices orchestrated with I don’t know whatever you want it would still be complex this folk song CEO has recently said that there are more than a hundred independent electric components in a car at work right now those components are programmed in embedded software and then car companies basically stick them together and make the car work if you want to make use of the data and if you want to make the services talk to each other independently having those legacy systems poses significant challenges and lastly let’s talk about the algorithms let’s talk about my favorite part anybody of you know the story about the junk science claiming that you can tell if somebody’s the criminal based on their face as I was published in 2016 or 2017 by Chinese researchers and apart from the saw issues which I will not get into is from the fizzy fizzy ignominy standpoint that the Nazis Triton eighty years ago and it didn’t work I want to show you why it’s rubbish what you see on the top is images of convicted criminals in the Chinese system and what you see there are driver’s licenses like like official documents like driver’s license images or passport photos on the bottom you see images of people who are not convicted criminally those pictures were put on websites like linkedin or on your professional website where you want to make a good impression on others the photos on top are just government IDs who of you would put your driver’s license image on your LinkedIn page who would do it nobody would do it because you don’t it’s a totally different purpose so what the machine learning model has actually learned is to fail is to check okay what’s the smile detector what’s the facial relaxation and expressing detector it hasn’t learned who’s actually criminal or not a similar thing goes for the second example a student at the University of California Berkeley trained in algorithm he went to ebay he wanted to buy a husky he wasn’t sure if somebody was saying it on eBay is selling her wolf or a husky so he trained the classifier to give ya separate between Huskies and wolves it presented it to his professor was 90% accurate everything was great and professors how can we be sure that the model is really learning was supposed to learn they did some investigation work came out of it is that they had Huskies as input images where there was snow in the background while while the Wolves had no snow in the background so this model solely learned to recognize snow in the background which is not what we wanted to do and what I want to give you or what I want you to take with you from this presentation is there are two ways to approach this one we have to make sure data sets are not biased that’s common knowledge but it’s really important that the data sets are not biased and second there’s a whole different process of verifying neural networks in machine learning machine learning is not deterministic anymore it’s not if then if then you be 100% sure that it does exactly what’s supposed to do it’s statistic so the process has to change to make sure that what we’re putting in the cars is really doing what it’s supposed to do so let’s go back to the George Hotz story to finish the story George Hodge had as common a I company was ready to sell his $999 devices to make every car a self-driving car and then all of a sudden the national traffic and highway Security Administration called hey your program is overriding secure signals please don’t do that so his basically his whole startup blew up from one day to another until he had a mind changing idea to open-source the entire stack and to rebrand it not as autonomous driving as semi autonomous driving they’re selling devices now for semi autonomous driving and what they ask the users is to upload their data because the open sourcing mistake in return people who use the software have to give them the data and they have collected more than five million miles of self-driving cars all right so where are we right now I think we all understand that the automotive industry is large it’s a giant it’s important part of our lives machine learning poses new challenges but also tremendous opportunities and what I want you to understand is that in my humble opinion this is right now the most exciting time to be in the automotive industry so much is at stake so much is changing and I really don’t care if you join the Daimler BMW of fuck’s wrong or anybody else but if you think any of those challenges are interesting extremely large data management verifying neural networks and making the infrastructure ready for machine learning then I think this is the right place to go thank you for your attention [Applause] [Music]


Leave a Reply

Your email address will not be published. Required fields are marked *