so thanks for the introduction we heard today already a couple of times that all the machinery all these nice products and technologies are driven by something that we know as beep learning and this is a very exciting area if you think about AI one of the hottest areas is deep learning fueling ng actually the entire eye movement so we speak today about deep learning in the future of AI how I see it or how the community sees it so first of all I always like to start with some slides politically-motivated you know this lady Chancellor Merkel she said that Internet is feelings Alan oiland it's a white spot on the map they lost 2013 not in the 90s right where the internet was invented but 2013 in particular context with respect to Snowden 2013 and then the next slide you see the former President Barack Obama where he said my successor will govern the country being transformed by AI two things right first of all he did not know that Trump would be his successor
so he was very optimistic at that point in time and secondly you know what the heck happened between the Internet is new and the society will be transformed by AI in three years so that's also a little bit of giving a pointer to the difference between Silicon Valley US and Germany so what happened was actually this so we see here we saw the slide Iranian first presentation this is a very important turning point where an machine artificial intelligence one in a very complex game called go against a human not only a human but the best the world champion in goal is a doll in five games that was in March 2016 and as already mentioned this morning many people believe that it will take decades to really conquer that and that game by a machine and suddenly it happened by deep mind and the game or the AI alphago so what was interesting here it practically did mind people Lisa doll when he came he was very confident in not losing right so he thought
he would go to this game play around and will not lose and suddenly he was losing that entire game there were five games the first three were won by the Machine the fourth game was won by the human and the fifth game again was won by the machine so I think the fourth game the Google and engineers tuned a little bit deeper images so that the human will win however there are some memes or today people would call fake news on the Internet so everyone thinks this a tall one but did you have a look at the black piece pattern at the white board so at that point in time I mean that that beautifully kind of summarizes the site guys that we have around artificial intelligence right half of the audience is like sci-fi enthusiastic embracing this great time and the other half is like oh well let's see how it develops and this is suddenly moving to this direction where people currently are very excited and this is very news thanks udit for
the pointer to this very recent news macro brain you said the first regional within artificial intelligence and the preneur if you're not getting up to speed on deep learning neural networks you lose very strong statement right so this is something that again puts an emphasis on deep learning and how important what kind of impact that has so I would like to go to a particular year 2012 where we really saw a revolution so there are two types of research right there's the slow step by step type of research reaching a boundary and then there is this kind of revolutionary impact that kind of explodes this boundary and redefines the the rules of research and we are lucky that we were in 2012 at this very moment where the rules were kind of newly defined and a revolution started in deep learning so what was that behind it's all about neural networks so neural networks is a particular part inside of the AI research area that deals with the simulation of the human brain on machines
right it's very old topic as old as we have computers people actually try to simulate the human brain that is consists of neurons being connected on a machine mathematically to simulate this activity inside of the human brain on a machine and it's very old and in 2012 it really changed the status quo of artificial intelligence so when we summarize what we actually want to do you see it here we have particular tasks for example you have a picture and then you have a human human looks at the picture and says that's a flower we have another task you know a driving situation the human and human can say drive carefully due to the current traffic situation or as we saw the game Lisa doll played and then the human who kind of moves the pieces in a way that he maximizes his game and wind finally the match so what we have here are humans in the middle of that that act as the end to end learning system right we cannot look in the
brains of the humans and that detail but somehow we can map the input a picture with pixels to an output that's a flower and that's actually a very powerful mapping mechanism and this is what your networks try to do also in an end-to-end learning scenario and then linear mapping so that's the idea you don't really know what's happening inside but you very much appreciate this mapping between an input and an output and you need this input and output pairs to train the neural networks and this is why we need this huge amount of training data to really be able to create this mapping and this is a parameter fitting this is called training and here we see what happens so when we look a little bit into detail we will see that this idea of training a neural network is actually not a new one right so we had this already as long as we have computers and there's some very important items like lunette in 1989 where particular tasks were solved by newer networks
here handwriting recognition was great it worked in that area it didn't work in other areas right so when I speak to people always people are there's always one guy who says like newer networks we're headed in the 90s it didn't work right so then suddenly Alec's net came so we see here first of all computer science guys kind of draw pictures nice pictures very complicated that's the Alex net architecture this is the reason we have to success today this is a deep neural network a particular type of deep neural network called convolution neural network and you see here in 2012 that this alex net created by Alex Cross SP was put into an image net challenge and won against all the other competitors against all the other machine learning artificial intelligence methods it did not only win it with a huge gap and everybody was surprised like what the heck is happening here right those things are not working why are they suddenly working and there's a good reason why they're working because three things
came together first of all we have a learning mechanism that has a variable capacity of learning we can increase the capacity of learning by increasing the neurons and increasing the layers in the network the more layers the more capacity do we have for learning problem we have to train that more layers means more free parameters we need the data and in 2012 we had the data with the imagenet challenge as you see here suddenly were much more data items to fit those parameters inside of the training method so the only thing that was missing was actually the machinery that is able to crunch the numbers to do the training and here one of the very ironic sarcastic things happen people misused GPUs graphic cards to train your networks right so there were people who are misusing that and a very interesting thing is that the CEO of Nvidia who invented GPUs did not suit those guys who misuse their products or his products he said oh well that's interesting let's see what we can do
let's support those people and give them more GPUs to do better work right so and then suddenly all the things change 2012 LX net was created and today Nvidia is making 1.5 billion out of 5 billion revenue by graphic cards that will never ever see a monitor in their life they don't even have a plaque right to put them on your door in their entirely dedicated to train your networks shipped all to Google and Facebook's in large scale so this is what happened the three things came together and that's the reason why 2012 is so important for deep learning with respect to that so when we see those numbers just to have an idea of the impact so we see here the results of imagenet that's the error rate so the smaller the better we see here the first traditional methods old-school right traditional four-night 28 arrow 25 percentage arrow and then suddenly Alec's net came and you see this drop so the first Network had an architecture of eight layers and I asked a
gentleman who gave the presentation this morning who is a neural specialist he said that the visual cortex has seven layers so when you compared the seven layers of the visual cortex to the original architecture of Alec's net with the eight layers you know that the eighth layer the last layers just the softmax so it actually the first people who invented the neural networks were a trying to simulate the visual cortex for this task of image net recognizing objects like flowers in images so that was the first architecture but the interesting thing happened everything suddenly was open sourced a system where you can train your networks was open source cafe and the reference model was open sourced so many many people could just download that and improve the architectures make those networks deeper which means adding more layers like vgg net and then we had Google Annette the inception architecture up to a very important moment in time 2015 whereas net with residential layers from Microsoft came into being and now we have neural networks that
have hundreds if not thousands of layers right much further than what we have in the human brain and they are much more capable so the interesting part is we still see improvements in making those networks deeper that improve where they have better performance improved error detection and this is interesting because we can now really see improvement in the perception process of that so here a very important oops five years of research right we started with eight layers already millions of connections up to billions of connections with hundreds of layers that we have currently and there's no saturation seen right so we can still there's still space to to improve respect to that so when we then have this success of neural networks and everybody is happy right suddenly we can build the self-driving car we have this problem that we have to explain when the self-driving car has an accident what happens inside of the network so it's a very important task to kind of open this black box of new networks to understand what's
happening inside and here are some slides showing what's happening so here you see and three networks or four networks where you feed the image from the bottom to the top and the first layers actually their gradient detectors H detectors like surprise me in our human brain and you see really that they are all the same for different things and when you go further you kind of zoom into the network you see there is a specialization up to the concepts that detected like faces cars elephants and chairs so that's the power of deep learning the first layers are very generic they share their representation it's internal representation share the weight and then the specialization is happening in the middle and at the end of the layers and therefore what people are currently using they're taking a very powerful neural networks throw away the concepts the cats and dogs and train their own concepts because we've suddenly with transfer learning you can use this kind of power of generalizability of those networks to train new concepts and
then suddenly scale up and recognizing your environment so when you go further in trying to understand what is happening inside of new networks a beautiful work was presented by Google deep dream was it called 2015 so you see a picture and then people try to emphasis a neural network on top of the picture people try to kind of push hard so the network was forced to detect something and indeed the network was detecting something an admiral dog a pig snail camel bird and a dogfish right inside of those cloud structures like we when we look up into the sky and see sometimes faces in the clouds so what we see here is people are kind of making they're talking about dreaming right so some some kind of human behavior that is happening now on the machine but what I actually see here is that is very nice very nice marketing gag from Google but it's a huge problem for privacy right the models the neural networks they have a compressed representation of the data they
were trained on so when you explain where you want to train a face detector it might be possible to reconstruct mr. Mueller's face privacy concerns right not that many people are aware of that that we have that and this also one point that is very important always talk to two companies right the software is not important or anymore as previously all those platforms are open-source tens of low cafe torch and others the important things are actually the models and the models are important because they compress the data and not everybody has the data so as a company be aware that your data is your IP and don't give you data away right so when I speak to German car manufacturers the self-driving is that's the thing that you develop and that you're not asking for Google to develop because in the very moment you're giving it to Google the data will be Google's so one interesting thing is also that we're moving from an app store paradigm in having apps and software to a model
store paradigm where we share models so this is very important if you think about the future we'll have trained models that we will trade and debt the power actually that we have with the standardization of deep neural network frameworks like tens of law for example and others then when we go further I was a very nice example from University of Tuebingen and colleagues over there so the task of deep neural networks until now was to have for example a picture an unstructured document and then to recognize concepts right it's an analytical task so what those guys did was very interesting because they reversed actually the function and they were creating content so their idea was we have a picture cubing em and you have a style of an artist and you can train a neural network on top of this style of annual artists and then you can fuse those two things together and create neural art so you see this picture it's being fused with the style and creates this beautiful image and I
mean I would put it on the wall it looks beautifully and it also tells us something that until now we are always thinking that you know and then future we will have machines doing all the boring work all the hard labor and we will finally be free to be artists poets and writers right and suddenly those machines are being creative and we can argue if this creativity or copy-paste right but they're moving in errors that we were not expecting that they're able to move into another example of that beautiful art that is created automatically and people are actually using this artists are really using this to create new art and so the job of an artist is moving a little bit the job of a programmer which is also very interesting to see so this is also a kind of interesting way of generating content buying neural networks instead of just analyzing that content so what we see as a next it's one of our work that we actually work we were the first team
worldwide able to extract sentiment from visual content so if something is positive or negative or has a positive or negative sentiment what we see here is the video nurses in collaboration of my former universities where I worked UC Berkeley and Columbia University what we see here is a video in real-time that is running and you see the top three so-called adjective-noun pairs concepts that we see inside of the video with the sentiment if something is positive or negative so we can use for example surveillance images to really do targeted advertising with the right context and here to build a map of Berlin for example given all the adjectives which areas are beautiful which areas are dangerous which errors are dark which areas are nice we can take this objective measurement to in subjective perception and now on top we can learn things like safety impressions or perception real estate pricing actually a big real estate company with sixty billion assets under management it's very interested in that because they have to manage their actually assets
in areas like India and China and they don't want to send people all the time there when they're looking for new objects right so we can on a large scale automatically analyze that and provide this back it's a very interesting work that was done now with deep learning and the best part because it's running on a GPU it's able to be do this in real time so what we see here that was also one of the reasons we're and via in September last year has awarded us with an envier I lab world the only two and VII labs in Europe it's us and our work and the lab in Switzerland so we see here a Jensen who won the CEO of Nvidia and he shows me this is the deep learning supercomputer what you see here a so-called DG x1 so you show set and you see a GC spining like like a child and Christmas I click but I it's great and the interesting part is there was just we had a photo shooting
like 20 people from his stuff he also has a bodyguard it's a kind of a crest action between a Navy Seal and hipster so he can kill you and still look fashionable right and what he thought was a finger print on this golden thing right so there were like 20 people and he asked for a glove and was cleaning it right they said like Damien look I'm a CEO of a multi-billion dollar company and I still have to do the work by myself all right that said something that you know a founder that is still a CEO on a multi-billion dollar company they're not that many in Germany right there's a very strong character with a very great vision so just to see what it means with respect to that so the very first Alex net that was trained in 2012 the guys needed two months of constant training of this machine two months in two years later for two weeks with new improvements of hardware last year in 2016 a new generation of GPUs
came out Pascal two hours just to get an idea of how much the hardware improves and how quicker we can train new neural networks with the idea that now you can go deeper right you can take more data to crunch and this is also something that we see that there is still an improvement in the hardware which allows us to use more data for training to fit the parameters the problem with deep neural networks is and that's the reason it took so long until they were successful it's a non convex optimization problem which means try an arrow you cannot press the button and it's guaranteed to have the best new network you try and arrow you trying already try an arrow and then suddenly it works so you have to reduce the time between starting the training and having result to get rid of all the things that are not working and kind of follow the path along the things that are working so I was actually yesterday or two days ago I was in
San Jose they have the annual GPU technology conference the so called GDC so by the talk were also part of the keynote so nice part is here yesterday all 11th of May Jenson also announced a new generation of GPUs volta within one year from 2016 to 17 we increased even the in kind of development of GPUs because in his basement he has a bunch of GPUs 1000 the classes cause a 2 and 5 and they are nothing they're doing nothing else than just improving the architecture of those chips machines building machines and that's the reason he was able to do a new GPU architecture within one year and you can see the stock price I mean the announcement was worth 10 billion of market capitalization one day change right so that's what happens two days ago and this is also a slide of the expanding universe of more than AI so you see research labs you see a technology provider cloud computing platform startups and you see established players right not that many German companies
although we have Aldi is a female somewhere so there are some companies that are understanding that AI is important right so if you think about the ducts 30 you know not everybody is understanding then on the sea level some people are understanding some people take longer to understand and then they have trouble in recruiting but it's not so evident like it is in the US and what we see here in this market it's not only AI as a new industry it's a disrupting force in every single industry starting in automotive moving quickly to retail finance this year finance is really crazy about changing due to regulation of next year medical pharmaceutical is moving from one industry to another and it's an estimated market of 500 billion US dollars and they're already 5 billion US dollars in VC money invested in companies and we have to be very careful in academia to have an expectation management here because we we are not supposed to promise too much because we know that there were a I went
us and we have to be very careful in kind of taking care of AI to not kind of over promise and under deliver so that's very important in this situation however we also have a responsibility for society so besides and beyond all this great commercial products that we can build and all these achievements in AI suddenly we can do things that were a couple of years before actually science fiction there is a responsibility in the society so we have to think very strongly about to speak to people in society about where to draw the line right where do we have to actually stop with the eye and the application of the eye and this is something that not a Mark Zuckerberg or a Sega brain should decide or a politician but society so we have to go and speak to the people to find the consensus about what should we do and what not it is very important to also take the magic away of this machine right there is no magic inside it's a
complex machinery that is working perfectly but it's no magic so we are actually with the Volkswagen system in a very strong program that will start next year about AI and society about changing that and also there is a lot of work about regulations and this is really important this market is not regulated now you can do whatever you want and there will be a demand for regulation so it's very important to first classify the RAI systems that has a direct impact of human life self-driving cars robot surgeons etc and there are I systems that classify cats and dogs right it's not that important we should focus on those a eyes that need more money and time to think about what they can do and also not focus too much on the others so at the end we have to start with a classification system for AI systems and then we have to build an AI tooth compliance so that's a great area where we can build beautiful startups that you can actually think about go
to a company like a large-scale company dime line ask them where what kind of AI systems do you have Intel's have no idea what kind of data do you have is a train which versions what the error rates right no idea right and there will be the demand to answer that questions and this is a great place actually for startups and if you think about five to ten years you know I promise the little bitty headline I just making up the companies BMW ai heck Daimler AI right and then people say oh well it's not our fault ai right you know whom right there would be a need for AI law enforcement and that comes hand-in-hand with the regulation right so we will have this kind of implications that responsibility of an AI self-acting will be not in the hand of the company or will it be in the hand of the company how did you look that right so beautiful questions that we have and last but not least we also have a machine-learning
all online course the very first German machine learning online course that we were created at the CeBIT already 5000 people it's all 338 clips not only academia but we have beautifully people from Google from the NVIDIA Amazon so that was very important when I spoke to mr. kargh Amanda mr. Volta when I designed the curriculum to have not only the kind of boring stuff but also the interesting stuff right to motivate so go for it it's all open source it's all free to watch the forum's are closed but it's still there and then with the last slide today I would like to say those are two of my heroes in Silicon Valley check the gasket and John von ocher on the founders of Adobe I truly a startup they said when they developed actually developed the desktop publishing industry with Photoshop Illustrator and PostScript so they said we were in the middle of a revolution and we were aware about our responsibilities so we are today in the middle of a revolution with the eye
and I'm wondering who is aware of the responsibility that is actually actively modifying and pushing this area so it's also very important to not only think about all those beautiful developments but also about responsibilities and maybe you can connect both and get a great startup idea for that thanks a lot for your time you