Kirill: This is Episode number 327, AI Trends for 2020.
Kirill: Welcome to the SuperDataScience Podcast. My name is Kirill Eremenko, Data Science Coach and Lifestyle Entrepreneur. And each week we bring you inspiring people and ideas to help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple.
Hadelin: This podcast is brought to you by BlueLife AI. BlueLife AI is a company that empowers businesses to make massive profit by leveraging artificial intelligence at no upfront cost.
Kirill: That’s correct, you heard it right. We are so sure about artificial intelligence that we will create a customized AI solution for you and you won’t need to pay unless it actually adds massive value to your business.
Hadelin: So if you’re interested to try out artificial intelligence in your business, go to www.bluelife.ai, fill in the form, and we’ll get back to you as quick as possible.
Kirill: So once again, that’s www.bluelife.ai and Hadelin and I both look forward to working together with you.
Kirill: Welcome back to the SuperDataScience podcast ladies and gentlemen. Super excited to have you back here on this show. Happy New Year and I’m hoping you had a fantastic celebration of this new decade.
Kirill: Back in Russia we celebrate the passing of a year and the start of a year today or this year, these past few days, it’s time to celebrate the passing of a whole decade and the start of a new one. How exciting is that. Super, super pumped to be in 2020. This is the most exciting time to be alive, period.
Kirill: We’re about 12 years from when the term data scientist was coined, that was around 2008. We’re about 10 years away from the first iPad, that was back in 2010. How long ago was that? Crazy, a whole decade has passed. We’re about eight years away from the publishing of the article Data Scientist: The Sexiest Job of the 21st Century on Harvard Business Review. If you haven’t read it yet make sure to read it by D. J. Patil and Thomas Davenport, that is what gave rise to this whole explosion of data science that we’re living in. So that was eight years ago, that was in 2012.
Kirill: Then for those of us who are excited about visualization we’re about seven years away, so seven years ago the first version of Tableau was published. That was May 17th, 2013, and now Tableau Software is a multi-billion dollar company with their revenue of 1.2 billion in 2018. They’ve been acquired by Salesforce. A lot has changed. A lot has happened since May 17th, 2013. Seven years of visualization. For those of us who are excited about artificial intelligence, 9th of November 2015 was the first time that TensorFlow was published, that’s the initial release date of TensorFlow.
Kirill: So, TensorFlow has been around for almost five years imagine, and just look at all the crazy things that are happening of TensorFlow. And looking ahead, we are about 25 years away from when artificial intelligence will take over the world. I’m just kidding, I’m joking, we never know what will happen. Of course there is … As Ben Goertzel puts it, there is a irreducible uncertainty as to what will happen once we create artificial intelligence forms that are way smarter than what we are capable of. So that is just an irreducible uncertainty no matter how hard we think about it. Simply because we are going to be less intelligent, we cannot come up with what will happen.
Kirill: But according to Ray Kurzweil on the other hand, if you look at futurism.com/images/the-dawn-of-the-singularity or just look for the dawn of the singularity. He’s predicting that by 2045, which is 25 years away AI will surpass human beings as the smartest and most capable life forms on the planet. And so that is in our lifetimes, that is 25 years away my friends and that’s a point in time when we don’t know what will actually happen.
Kirill: It’s already getting pretty fast like how technology is developing, we really don’t know what to expect the next 5 years, let alone 25 years, but there we go, that’s a prediction from somebody who’s predicted things like the iPad and many other technological developments with approximately an 80% or more than an 80% accuracy. So, someone worth listening to.
Kirill: Okay. So that is 25 years away but what is happening this year? What is happening in 2020? That’s the real burning question. What should we look out for in 2020? What are the trends, the most important trends.
Kirill: So Hadelin and I were part of Udemy for Business webinar, huge shout out to Udemy for Business. If you’re a business executive and you don’t use Udemy for Business for training your team yet, make sure to jump on top of it. You can get access to plenty of high quality top-level Udemy courses, including some of our courses, including some of our exclusive courses as well and train up your team on Udemy for Business. Just go to udemy.com or search for Udemy for Business.
Kirill: But basically what we did is that we were on a webinar talking about AI trends for 2020 and beyond and we selected our top five trends which we dissected quite deeply and discussed what they are, what are these technologies and what to expect. So, that’s exactly what we’re gonna be talking about in this podcast today in our session today. So it’s actually the same audio from the webinar and you will get to experience it today in a podcast form.
Kirill: So the five trends that we’re going to be talking about are number one, robotic process automation and AI technology that’s taking over the world very silently but is very, very impactful. Natural language processing, very powerful technology. Reinforcement learning, Hadelin’s favorite technology, edge computing, something that I am learning a lot about these days and something that I’m a big fan of. And finally we’ll talk about open source AI frameworks and why that’s a trend.
Kirill: So just five technologies this year. We decided to focus on less but make it more impactful and look into some examples so you’ll hear about what these technologies are, why these are trends and also some industry case studies.
Kirill: So that’s our podcast for today, very excited to dive straight into it, so without further ado, I bring to you my good friend and business partner Hadelin de Ponteves. Let’s take it away.
Kirill: Hello everybody very excited to be here on the webinar.
Hadelin: Hello everybody very excited as well.
Kirill: And here we have our agenda for today. We’re gonna be covering off five main topics that are we should be looking out for in 2020 and they are robotic process automation, natural language processing, reinforcement learning, edge computing and open source A.I frameworks. How you do feel about this at Hadelin?
Hadelin: Very excited, very excited. Indeed these are the top trends in the artificial intelligence ecosystem and so we’re gonna describe each of them by giving you the top real-world industry examples.
Kirill: Yeah we had a look through all the technology updates with Hadelin over the past couple of weeks and especially yesterday we sat down and there is some very interesting things going on. Which one are you most excited about, Hadelin?
Hadelin: My personal favorite is reinforcement learning as it has always been. And what about yours?
Kirill: Probably edge computing, I’ve been learning more and more about it. I didn’t know that much about it earlier this year but now it’s really becoming prevalent and very excited to see what’s coming up. So without further ado, let’s get going.
Hadelin: Let’s do this.
Kirill: Robotic process automation, we’re going to take turns, I’ll start with Robotic Process Automation, and then we’ll take turns with the trends.
Kirill: So with robotic process automation or RPA, let’s do a quick overview of what it is just for those who are not up to speed. It’s a very, very simple AI technology, one of the simplest out there but also one of the most disruptive. So imagine you have a human who is doing a certain process which might be like registering a new client in system or onboarding a new client, invoicing clients, some kind of manual process that’s done on the computer, very repetitive high-volume task.
Kirill: So for instance the human might read an email, open an attachment, copy data into a CRM or customer resource management system, then get data data from a different database and then send an email reply. That might be part of the process that they have to do every single day, 10, 20, maybe a 100 per day, 100 per week in order to fulfill a certain business need such as invoicing a client. This might be relating to invoicing a client or this might be related to closing or fulfilling a sales transaction or something like that.
Kirill: So basically what RPA does is it can take over these high-volume manual repetitive tasks and in this case, we have a very nice beautiful icon there for robotic process automation. So these are not actual robots, these are software robots. So, they live on the computer but they’re able to do all the clicks and moves of the mouse and things like that that humans can. They just do them faster and more accurately.
Kirill: And so what an RPA can do is it can read the email, open the attachment, it can also copy data into CRM, it can get data from a different database and it can even send a reply. But what it’ll do is … by here, by this point in time, you can set a check, and if there’s an escalation that is required to be checked by a human … if there is no installation required, then it’ll send the email. If there is an escalation that’s required, then will handle an exception, well basically the human will come in and handle the exception. That’s in a nutshell what RPA does, so, so it can take away all these repetitive mundane, non-interesting tasks from humans and free up people for more exciting work.
Kirill: And so something to note here is that there’s about 230 million jobs worldwide that can be automated and that’s about 9% of the whole global workforce. So it’s a massive opportunity and some people see as a massive threat because it kind of means that jobs will be taken away from humans or can be taken away from humans.
Kirill: But ultimately there was a study done by the World Economic Forum in 2018 which says that for every one job that AI displaces in the coming years, 1.7 new jobs will be created, more interesting, more creative, more fun exciting jobs and that’s what we want humans to be doing ultimately. So that’s kind of like the underlying part of this whole presentation, is that we see artificial intelligence as adding value to the world. What would you comment on this, Hadelin?
Hadelin: Well, definitely robotic process automation is one of the top branches of artificial intelligence. It is more and more used in companies to automate a repetitive task. And actually on that note, being that some jobs can be automated. Well, I always relieve my students or other people with the main fear that gravitates around artificial intelligence, being that it’s going to take over the jobs.
Hadelin: Well, yes, for each job that is going to be taken, a lot more can be created. It’s just a new industrial revolution like when before a lot of the agricultural jobs were gone, well a lot of new industrial jobs were created. Well here we have exact same new phenomenon with artificial intelligence. So there is not really a fear to have on this and actually we’re just going into growth.
Kirill: Hmm, totally agree. So let’s have a look at some of the applications of RPA. This is just couple of examples to spark some ideas that you might have for your own organization. Number one would be invoicing and billing. So as the example we looked at, if you have manual processes especially in large enterprises for invoicing clients and billing and following up and things like that, all that can be automated with robotic process automation.
Kirill: Payroll processing, another one. There is a lot of checks that need to be … or check-boxes that need to be checked off at the end of the month or twice a month whenever the payroll happens. If that has been done by humans, it can be done by robotic process automation system.
Kirill: Data aggregation, RPA can… You can set up an RPA to go and extract information from Bloomberg every morning, and then from a certain email that you receive, from a database, from … I don’t know. Like some online system, from Google News, you can get data every morning or on whatever basis you want, however regular you want and put all that into a report. RPA can help with that as well.
Kirill: Shipment scheduling and tracking. Again, in logistics companies, a lot of businesses actually have logistics components to them. RPA can assist there. Finally, employee onboarding is another area where robotic process automation can help and we’ll see this in one of our examples.
Kirill: And in terms of examples, we’re going to look at … for every technology, we’re going to look at five examples in the industry of how that technology is already transforming existing businesses. And that’ll give us ideas of what we should expect in 2020. So in terms of RPA, we’ve got first example is Synergy. It’s a energy company in Western Australia and what they have done is they have been able to deploy 280 bots, I think they used the software Automation Anywhere.
Kirill: These bots were able to reduce the era by 99%, increased the speed by 40% and this is for invoicing and billing. So the billing of their customers, I’m guessing is like retail and probably commercial as well, customers that use their energy. So 90% reduction errors, 40% increase in speed, 2.3 million dollars annual savings. And of course that comes with a question, what was the price? Well, the ROI, return on investment on this project was 163% overall. So massive ROI plus annual savings that will be recurring over the years.
Kirill: Next example is Deutsche Bank. So, they’ve experimented a lot with robotic process automation and every time they’re able to automate between 30% to 70% of the task in whatever area they’re experiencing it. And actually Deutsche Bank is the example I was about earlier where they’re able to use robotic process automation for employee onboarding.
Kirill: So, what they do is they have an RPA that has the task encoded into it and when like human on boards, it probably has like the core of the task encoded into it. We don’t know the exact details but the core is encoded into this robotic process automation system and then it can train a human who is onboarded in order to discover how to do this task and then take it to the next level. So, basically they’re using RPA to help people onboard into the bank.
Kirill: AT&T are very big, very fond of robotic process automation, they’ve been it doing for years now. They have over a thousand software robots in total and in fact they’ve gotten so good at it. So, basically what the robots do is they perform task ranging from helping technicians, activate equipment for customers and then aggregating data, what we talked about to create custom reports and so on.
Kirill: And AT&T have gotten so good at RPA that they’re actually teaching other companies how to do it. They’ve trained over 2,000 employees in a hundred different organizations on how to deploy RPA.
Kirill: Vanguard is a investment company and so they have 3.2 trillion dollars in assets globally that they’re managing and they use RPA to perform certain trading task where they can be pretty straightforward. When this happens, do this, when this happens, do that. So not only they use human traders but they use robotic process automation. They use a combination of the two kinds of approaches which kind of helps them stand out in their offering.
Kirill: And finally the Walmart. Walmart a is also a big of fan of robotic process automation. They’ve deployed over 500 bots to automate anything from answering employee questions to retrieving useful information from audit documents. And this is coming from the CIO. Actually the reason for this was that humans doing those tasks were getting “a lot of those came from people who are tired of the work.” So people just getting tired of certain repetitive tasks and asked for automation and that’s where automation came in and helped. So that’s RPA in industry. Let’s move onto trend number two.
Hadelin: Natural language processing. So, a very popular branch of artificial intelligence that is growing a lot especially since the development and the creation of BERT which is a revolutionizing chatbot. So, yes, natural language processing, let’s go into it and let’s first explain what it is.
Hadelin: So first natural language processing can be divided into two sub-applications. The first one is a natural language understanding which will consist of reviewing either a text or a nodule and interpret it. For example we have the classic application of NLP which is sentiment analysis, which will analyze some text and will say whether the text is positive or negative for example.
Hadelin: And then the second application is a natural language generation, which is for example the case of a chatbot. So, basically you will get … a system will get a text as an input or even a nodule or a person speaking and saying something and the natural language generation system will generate a response to it, for example like a chatbot that talks to you for a specific task.
Hadelin: So these are the two branches of natural language processing. Now, let’s go over the different applications starting with the first one which is sentiment analysis. So, this is the example I’ve just given. For example, we can use sentiment analysis to review … to tell whether a review is positive or negative. For example we can also use sentiment analysis to say whether some Twitter tweets are positive or negative or even can be classified into some categories. Then the second application is a chatbot. So basically this is at this time a natural language generation. You have a chatbot talking with a customer about either some general conversation or some specific topic. Then a third application is machine translation. Here, well, you will have a natural language processing system that translates text from one language to another.
Hadelin: Then the next application is the automatic summarization. So, this system will this time take a text as input and it can be a full book or an article and will summarize it in a few words. And then finally the last application widely used is video captioning. So in this case while the NLP system will get a video as input and will deliver as an output the captions of that video.
Hadelin: So these are the main applications of the NLP and now let’s see how it was used in the industry. So a great industry, first industry example is Autodesk. Autodesk is a 3D, a design engineering and construction software. So they used natural language processing by downloading the spec of F1 car to design a rule hope, so to design a specific material without the need of any human involvement. So that’s pretty amazing.
Hadelin: Then another industry example is booking.com, of course a company that with which you can book your hotel and as you know, on this booking.com website, you have a lot of a reviews from customers and you also have the data from customers. Well, you can extract those reviews from the customers to predict where should be a next destination for the customer. So that’s another big company that uses NLP a lot to optimize the business.
Hadelin: Then you have YouTube of course. This uses NLP in many cases to relate to the different applications we’ve just mentioned. Well, it can be for example for a video captioning, it can take a YouTube video as an input and will return as an output, the captioning. So all this is done automatically by an NLP system. Then we have JPMorgan Chase & Co. They actually made an amazing program called COIN which stands for a Contract Intelligence and basically it does the job of interpreting commercial loan agreements which would take for example 360,000 hours done by lawyers. And actually they made an NLP system of course that did it in second. So that was really an impressive application of natural language processing.
Hadelin: And finally we can also give the funny example of an institution like the the Port of Rotterdam which also managed to use NLP by getting the port calls of the port to predict at an exact time the vessel arrivals in the port. So that’s a pretty fun example. They took as input the calls and different data, whether it is text or audio and they managed to create with a high accuracy, the vessel arrival times in the port. So that’s a pretty good industry example.
Kirill: Very, very interesting. Thanks Hadelin for walking us through that. While you’re talking, I had this question, why is NLP becoming so popular these days? What’s changed? Why are we highlighting it as a trend for 2020 when there’s so many other technologies? You know, we’re only highlighting five and one of them is NLP. What is changing these days?
Hadelin: Yes, I would say that it’s because it is the fastest growing one in terms of research and so that’s why I wanted to introduce first NLP with BERT because BERT is really a revolution. Research in artificial intelligence is very slow, it’s very continuous. But sometimes you have some big leaps in innovation and BERT was actually one of them. Also, in terms of business and economy, well, NLP is providing tremendous added value to diverse industries with the different applications we’ve just explained. So also in terms of economy well, this really is one of the top branches of artificial intelligence in terms of brought added value. So I would say yes, the two components, it has the growth and also it has the innovation.
Kirill: I would also say that just like humans are becoming lazier that we need to type into our phones and stuff, it’s much easier to talk to Google Home or Alexa or Siri. It’s much easier to say things, right? That’s probably the consumer-driven reason, that people want a new interface. Before the interface was all graphical and we are able to interact with graphics and type stuff up, which was a huge change from writing and having to always pick up the phone and talk to a human. Now we can go on a website and do things. But the next step in terms of that is, “All right, now, let’s be able to talk to our computers, why do we need to type?” Like that movie, you remember Her? Have have you seen Her?
Hadelin: Yes, yes, great movie.
Kirill: Yeah, they don’t type at all. They don’t even have keyboards in that movie.
Hadelin: That’s true. Yeah, that’s the next generation. And even the next innovation would be to not even have to type anything or to say anything. We can have a system directly connected to our brain, which will collect our thoughts. Yes, this will be I think ready within the next 10 years.
Kirill: Yeah, I would agree with you. All right, well, let’s look at the next trend then. Oh, you’re doing this one as well.
Hadelin: Yes, reinforcement learning. So, yes, this is my favorite branch of artificial intelligence.
Kirill: Why is it your favorite?
Hadelin: Well, I think this is the one with the coolest applications. For example self-driving cars or playing video games. So this is thanks to reinforcement learning, for example that AlphaGo was able to beat the world champion of Go or same for Chess and we’re going gonna show another example here in this webinar. But yes, I think this is a very, very cool branch of AI. And besides, in business it had some amazing impact and we’re going to discover what they are now.
Kirill: All right. Let’s go.
Hadelin: All right. So, what is reinforcement learning? Well, it’s a input and output-based system, meaning an AI that will take as input some data and will return as output an action to play. So here we give this example of a carrot and a wood stick. Because actually the fundamental point about reinforcement learning is that it relies on a reward. There is a reward system that will evaluate how your AI is doing. And the better it is doing, the more rewards you will give it. And the worse it is doing, the worse reward you will give it. So here, the plus one and minus one, are the two rewards.
Hadelin: Let’s say that’s we are training an agent or an AI to predict whether an object is a carrot or a wood stick. Well if it predicts a carrot well, we will give it a reward of plus one and if it predicts wood stick, we will give it a reward of minus one. So that’s the simplest example of a reinforcement learning. You have to know that it is an input and output-based system that is evaluated over a reward system and will train itself over trial and error to reach a certain goal.
Hadelin: And so now we’re gonna look at the different applications of reinforcement learning. So yes, this is another example of reinforcement learning. You can use it to train a robot to find its way in a maze and again you have a reward system that will give it plus one. Well the closer it gets to the destination, you know, getting out of the maze. And minus one if it doesn’t find its destination or gets further away from it.
Hadelin: All right, so here now we’re gonna talk about a very exciting application of a reinforcement learning which is AlphaGo Zero. So I’m sure you’ve heard of AlphaGo. Because indeed, it was an AI that beats the world champion of Go and that same AI was a made for other applications and for other games. And one of them is actually StarCraft. So there was this very recent amazing event where the team at DeepMind in London which is like a company acquired by Google well used this AlphaGo system to beat the top professionals of StarCraft.
Hadelin: And there is actually a very cool video on YouTube where you can see this AlphaGo Zero playing against the those pros. So I don’t know if you know how this game works, but it’s very complex. Actually, you have a lot of action that you can use at the same time. If you watch the professional players play this game, you will see all their fingers used in different keyboards to complete some action.
Hadelin: So basically what I’m saying is that there are many actions to play at the same time and therefore a great complexity and not only from the inputs, but also to the output. Well, the AI that played against these top professionals at StarCraft, totally beat them five to zero. So that was really a great accomplishment given the complexity of the game.
Hadelin: So yeah, that’s one of the applications I like. And now if we are looking at the different applications in industries, in business, first we have a personalized recommendations. Of course, if we take the example of Netflix, we can train a reinforcement learning system to predict whether a customer is going to like or not review. And in that case, the reward system is for example plus one if the reinforcement learning system predicted correctly, and minus one, if it was an incorrect prediction.
Hadelin: The as a next application, we have advertising budget optimization. Indeed that’s classic, but yet widely used application of reinforcement learning. Indeed some of the top companies like we’re gonna see a bit later, Alibaba are using reinforcement learning to optimize their advertising budget. Which allows them to save millions in cost.
Hadelin: Then the next application is to select the advertising content. Well for example imagine that you want to sell a product and the design team made several advertising or different ads and you don’t know which add would convert the best your customers. Well, you can make the reinforcement learning system that will quickly figure out which ad will sell the best … the product to your customers and while still optimizing the advertising budget. So you can not combine two applications at the same time.
Hadelin: Then another classic application of reinforcement learning is to ensure a customer lifetime value with same reward system that will be clearly defined. Such that a positive reward is given if indeed you are ensuring the customer lifetime value and negative reward otherwise. And last application that we can give this to predict the customer response for example for customer service. Well, you can use again a reinforcement learning algorithm for such a purpose.
Hadelin: And so now we’re gonna look at the different industry examples, the main ones. And you’re gonna see great companies like Alibaba and Google and even some other institutions that use reinforcement learning. Let’s start with the biggest one and because it was the one that had the most impact thanks to a reinforcement learning.
Hadelin: Indeed Alibaba leveraged reinforcement learning to increase its return on investment in online advertising by 240% without increasing the advertising budget. So there’s actually a research paper on this particular case study where you can see exactly how this reinforcement learning system outperformed the benchmark of the other systems. And this was the top AI that really provided the best results. So 240%, that was the incredible.
Hadelin: Then we have Google who used, again, thanks to DeepMind, one of its top AI branch, reinforcement learning that was in 2016, to reduce energy consumption in heir data centers, so that’s a very classic example. They managed to reduce energy consumption by more than 30% thanks to a reinforcement learning model which is called a DQN model based on Deep Q-learning. And that was actually incredible because this allowed them to save billions in cost also.
Hadelin: Then as a next example, we have the University of Cambridge. So they did not particularly use reinforcement learning for a particular application. What they did is that they included a new reinforcement learning program in one of their masters. And they also hired one of the top DeepMind professors to emphasize this branch of artificial intelligence. Because indeed, it has such amazing application in business that it was included now in the programs by the top AI researchers in the world, which is this DeepMind professor.
Hadelin: And then we have Tesla of course, which leverages, as we said reinforcement learning for self-driving cars. Because indeed you can train a self-driving car with reinforcement learning to navigate and to reach destinations. And of course, you will give it some positive rewards if it manages to go in the right direction and some negative rewards if it goes in the wrong direction or if for example it hits some obstacles. So we have a very, very advanced a reinforcement learning system inside a self-driving car, and Tesla is using it to build their self-driving cars.
Hadelin: And finally, Trendyol. So Trendyol, for those of you who don’t know is a leading e-commerce company. What they did is that they used reinforcement learning to create several highly personalized marketing campaigns. So that’s the applications we identified before, you know, optimizing the advertising budget and also selecting the right ads in a marketing campaign. Well, this is another one company that used reinforcement learning for advertising purposes. So yes, these are the main industry examples that connect well to the main reinforcement learning applications and so now we’re gonna move on to the next big trend in artificial intelligence.
Kirill: Thank you Hadelin. By the way, what’s your favorite application of reinforcement learning? There are so many.
Hadelin: I really like what Alibaba did and also the self-driving cars, I really like, and the online advertising budget optimization because this saves like millions of dollars to companies. So this has really, really, really good effect, amazing impact. And at the same time, it’s very easy to implement. Actually many companies could do it. It’s not that much of an advanced system. However, yes, the most advanced system would be the self-driving car, the reinforcement learning, DQN model applied within self-driving cars.
Kirill: All right, thank you. Yeah, that is very, very exciting. Also probably not just … it saves companies money, but also helps not spam people with irrelevant advertising. Sometimes we forget that tailored advertising is good for both sides like for the people and for the company.
Hadelin: Yes.
Kirill: Awesome. Okay, edge computing, this one is mine. So edge computing. Have people watching or listening to this ever seen this message saying that Siri is not available. When you try to use Siri on your iMac. So just so if you haven’t, just try using Siri when you’re offline, when you’re not connected to the internet, try clicking the Siri button or say, “Hey Siri,” on your phone. You will notice that Siri is not available when you’re not connected to the internet.
Kirill: When I see this and when I first discovered this a couple years ago, and it’s still the case. We are entering 2020 and it is still the case that you cannot call Siri or talk to Siri or ask her to make a note in your diary, on your phone, or do whatever you want. You cannot just talk to her if you’re not online. We are in the 21st Century in the cusp of 2020 and we are not able to talk to our voice assistants if we’re not connected to the internet. That is like a shocking revelation to me if you think about it.
Kirill: And so the questions for that is like why? Well the reason is that … this is where edge computing comes into play. So the reason is something that we all know which is called the cloud. It was new a decade ago or so, or was something we talked about. Now it’s like everybody is used to it. So the cloud, we’re gonna have like a diagram in the video version here. So we have a cloud and there’s like data centers out there somewhere far away. It may be in a different part of the country or maybe even in a different country where your data goes, its processed and comes back.
Kirill: That’s why Siri or Alexa or Google Home, none of those devices are actually going to be able to help. You’re not even going to be able to talk to them because the natural language processing algorithms that Hadelin was talking about, they’re quite computational heavy, and they sit and live on the cloud. So your data gets uploaded to a server, it happens all very fast, as long as you’re connected to the internet. It gets uploaded to the server, it gets processed there and then the response comes back to your device and some action is taken or whatever.
Kirill: So even though you’re talking to Siri on your phone, Siri is actually not in your phone. Siri is far, far away, thousands of kilometers sitting on server somewhere in the cloud. And what this edge computing is doing, what world we’re moving into is how about we take those microchips algorithms and we put them all onto your device. So we’ve already heard about internet of things. So your car might be connected to your phone and your phone is connected to your thermometer and to your Apple iWatch and also to your computer. And then your credit card is connected to … like as soon as you walk into a store, it knows your credit card details. Like in Amazon Go store you don’t have to be pay anything there, it all gets deducted automatically.
Kirill: Or you’re in a factory, devices might be interconnected. All this interconnectedness of devices is kind of like something that we’re seeing in the IoT space. But at the same time there’s still the algorithms that drive all of these connections and all of these interactions, they still live on the cloud. So even though your device is connected, they need to upload the data. The data has to then be processed in the cloud and then comes back making everything work together. It’s quite seamless because it happens very fast, you don’t notice. But it still happens up out there.
Kirill: What if we could put those algorithms and those microchips onto the devices so that everything could happen locally, not needing to be uploaded to the cloud? And that’s edge computing on device and this would be billions and billions of artificial intelligence algorithms happening real time on your device. So versus you could talk to Siri when you’re not connected to the internet and she could help you. Maybe you don’t need to be connected to the internet for a task you’re doing. Maybe you’re searching for something on your phone. Or maybe you want to make a note on your phone. Or maybe you just want her to access an existing library of information or play some music that you already have downloaded on your phone. So that’s edge on devices.
Kirill: Another type of edge computing is edge computing on the nodes. And nodes are somewhere in-between servers out there in the cloud and your devices that you have locally right in front of you. So these nodes are kind of like mini-servers that live very close to your local internet provider or local telecommunications station. So rather than all this information from your internet of things, devices from your mobile phone, your temperature meter in your home, your Apple iWatch, from your car, from the iPad, from your as well as healthcare devices and things like that. Rather than all that information going all the way to the cloud to be processed, it can be processed locally on an edge server which is much closer to you. Things will happen fast. It’ll be lower latencies. It much more secure.
Kirill: And these edge servers or edge nodes, why they’re necessary is because not all devices might be in the vicinity of each other. Not always you might be your car and your phone and your home temperature meter might be very close to each other, so they can connect directly with each other. They might still need some interconnectedness, but they don’t need interconnectedness across the country. They just need interconnectedness within your city or within 50 kilometers of where you live and that’s enough. So that’s where edge nodes come in. So we have if you imagine a pyramid, we have cloud computing at the very top. We have data centers which this is how everything’s happening now.
Kirill: But now what’s going to happen in 2020 and beyond is the computations are going to be pushed out onto the edge into the edge nodes. And further they’re gonna be pushed out onto your devices so that they’re actually being … these chips and these algorithms are living on your device. And for example the introduction 5G networks is really facilitating this process. 5G networks make data transferring much faster and that’s why these nodes can be set up. In fact we have an example here from Vapor IO where they’ve set up a network edge nodes around Austin. You can see they have six edge nodes all interconnected with fiber optics I believe.
Kirill: And that way if anything happens … if one of your device is near one of these edge nodes and then another device is near another one, they can connect with each other and exchange information very fast. All that information doesn’t have to go into the national cloud or the cloud server of wherever this processing is happening whether Amazon web services or your Apple cloud servers or some some Google cloud servers. It can all happen locally around Austin in this case, and all of everything you need to get done can get done with lower latency, faster execution time and more privacy.
Kirill: So here let’s sum up what is edge computing? Edge computing allows computation and data storage close to where they are needed rather than transporting all this data which is also costly across the country. And we have so much data flying around even if it’s a very little cost per unit of data, when you times that by the volume of data we have in these years and moving forward, it becomes a very, very complex exercise. It’s real time data processing, much faster, avoids network latency, allows faster responses. Cloud computing is big data, but edge computing is instant data. That’s the best way to think about it. Cloud computing, big data. Edge computing, instant data.
Kirill: And so let’s have a look at a couple of industry example. So for instance AT&T, our favorite company. So what they’re doing is they’re actually building a network of edge nodes in order to facilitate self-driving cars. So instead of your self-driving car having to communicate through the cloud with other devices around it, with maybe the smart city sensors that exists with other cars that are driving around. That can all happen locally where the car is driving through AT&T edge servers and it can get all the information faster. And of course, that’s going to be very important and crucial to the safer performance of these cars.
Kirill: Daihen, so what they did is with their edge technology, they were able to … basically Daihen, they used edge technology in order to in their factories reduce error detection. To make an error detection faster. What they do is they produce transformers I believe and with edge computing they were able to eliminate about 5,000 hours of manual entry. Because simply they are now able to detect any kind of product defect and production errors on the fly thanks to edge computing.
Kirill: Our next example is on Amazon Alexa. So as we know all these voice assistants, they’re not in your device. So the actual assistant is not in your device, the algorithm is not there. It’s just getting your audio and then it’s uploading that audio to the cloud, processing and coming back. So Amazon Alexa is in the same boat right now. The Alexa itself is not living on device, but what they have come up with is like a first step towards edge computing on the devices. And they have this new chip which only has one megabyte of RAM, is very cheap and it can be installed in pretty much any device. It can be installed in a light bulb and what it does is it listens for the wake word which is Alexa, right? As soon as you say Alexa, Alexa is supposed to wake up and start processing. Well before that requires like 100 megabytes of RAM, that was quite an expensive chip to put into your routine devices but now they really reduced it. And even though it can’t process all of the words that you’re saying, it can listen for the wake word. So you could have a chip that is listening for a wake word in your in any light bulb.
Kirill: You could have it in your cutlery, in your plates and cups and mugs and whatever else, wherever you need it, so they’re moving towards that so it’s a first step.
Kirill: Next one, Novartis, so this is an interesting example from the health care space. They’re developing, they’re testing already contact lenses that can measure the sugar content, I think it’s through the content inside … glucose levels inside tears of diabetes patient.
Kirill: So they can tell we can measure constantly how is your glucose going, do you need to have an insulin shot or not and that can be life changing or life saving for somebody with diabetes. And of course that needs to happen very fast, latency cannot be afforded there why not do that on the edge? Why not do it on their edge node or why not do it on the device if that might be somehow connected to … might have already a chip somewhere built in which is quite hard to imagine in a contact lens but maybe it’s connected to a phone or something else like that. So, edge computing would be very useful there.
Kirill: And final example, example of Pokemon Go which we all remember from 2015, 2016, went viral, millions of people walking out with the phone looking for Pokemon. Well, why does that data need to be uploaded all the way to the cloud and then processed there then come back your phone telling you, if you caught the Pokemon, if Pokemon’s there, this is augmented reality, it needs to happen fast. It needs to happen on the fly, that’s why having edge nodes in that case would be very useful because even if you have many players in the city walking all together, because they’re working through one edge nodes that’s near them or they’re working through several edge nodes that are inter connected but they’re all local anyway, they can gets the results faster. They can see all the same thing, they can … if somebody caught a Pokemon, other people cannot catch a Pokemon. So they’re all interconnected anyway and then once in a day or a couple times a day, all that data can be uploaded to the cloud to aggregate with all of the other global data that’s been uploaded from other edge nodes.
Kirill: So, basically if we localize this processing of data to edge nodes or edge devices in some cases, then it all becomes faster and we remove any redundancy where we don’t really need to be uploading stuff to the cloud.
Kirill: So those are examples for edge computing. How would you think of edge computing Hadelin?
Hadelin: Well this is definitely fascinating and it’s actually … well the one that I’m least familiar with but it’s really worth … I think it’s really worth looking into it and especially into how AI can bring added value to businesses. I think this has tremendous potential.
Kirill: And I really like, apart from the speed, I like the case that edge computing is safer, right? You’re not transferring the data to a server somewhere nationwide, it’s being processed locally. So it’s much harder to hack for somebody who is not from there. They have to hack into the nodes first and then hack the data. So, I think it’s a great step up in terms of data privacy as well.
Hadelin: That’s true it protects you from the dangers of other cloud.
Kirill: Yeah. Okay are you ready for your last one?
Hadelin: Yes, the last one last one is about open source AI frameworks. So, the first thing I want to say is that this is actually really good for the development of AI research because what the open source AI frameworks allow to do is for many people, any people on the earth to participate into the development of AI research. That’s why today we have so many GitHub pages by individuals who are really talented and do not necessarily belong to you know one of one of the top IT companies and yet participate to AI research, try to improve the current AI models thanks to the different libraries that are offered open source.
Hadelin: So, let’s have a look at them, let’s have a look at the best ones. And first, let’s quickly explain what they are, open source AI framework. So, basically there are some libraries or platforms that you can use to build artificial intelligence systems. So, for example if you want to do computer vision, you will have some frameworks that will allow you to implement a computer vision system with very few lines of codes, by using different … what we call modules, function or classes of the libraries within the AI frameworks.
Hadelin: And same we have some AI frameworks for NLP, BERT is an example of a framework. We have some AI frameworks for reinforcement learning. So, we have different AI frameworks for the different branches of artificial intelligence and now we’re gonna have a look at them or have a look at the main ones.
Hadelin: All right, so first open source AI framework application is a research and production support. So, that’s what I introduce that with. Well, the open source AI frameworks are used for research whether it is to implement some artificial intelligence for the healthcare industry or for another industry. Then we have the algorithm library, which is for example TensorFlow, PyTorch, which will contain diverse modules and diverse functions of artificial intelligence that you can use very easily with very few lines of code to implement something very complex.
Hadelin: Then we have a model creation and development so basically you can very easily create models once again with very few lines of codes for example if you want to build a CNN which is a convolutional neural network which allows you to classify some images for example. Well, without the open source AI frameworks, you would do it in hundreds of lines of code and now thanks to them, you can just do it in actually not more than 10 lines of code. With for example the Keras library.
Hadelin: Then you have as another application, prototyping and training algorithms, well, same. Thanks to those open source AI frameworks, you can train some very complex algorithms. Most of the time these frameworks are connected to a server, you know, a powerful server which is like a powerful computer connected somewhere in the world as a virtual machine to which you connect to to train your algorithms.
Hadelin: So, for example if we want to connect to what we’ve said before about reinforcement learning, some reinforcement learning modules can be very hard and time-consuming to train. Well, thanks to these AI frameworks, you can now train them much much faster.
Hadelin: And then another application, define, optimize and assess. Well thanks to these frameworks, you can have some really good pipelines that can allow to facilitate the process of building a model, defining a model and optimize the goal that you want to do. And then you have some tools which allow to asses the performance of the model you’re building.
Hadelin: So, these frameworks not only allow you to build some amazing models but also can offer you some pipelines to facilitate the process of building a model for a specific goal that you have.
Hadelin: And now let’s have a look at not really examples, but the main AI frameworks that are used today by the researchers and any individual because they’re open source. The first one and actually I think the most popular one is TensorFlow. So, TensorFlow is an AI framework developed by Google which allows today to do any branch of artificial intelligence. With TensorFlow you can do deep learning, you can build a convolutional neural network to do image classification. You can do reinforcement learning, because you can build a deep Q learning model with Tensorflow, you can also do NLP of course. You have some modules that allow you to build some NLP systems. Well you have everything in TensorFlow, this is the most popular, this is a very powerful one especially since the the development of TensorFlow 2.0 with which you can do even more advanced AI systems.
Hadelin: Then another AI framework Keras, but I shouldn’t say another one because actually TensorFlow 2.0, well, TensorFlow merged with Keras. But Keras is an AI framework that allowed to build some AI models in an even simpler way. You had even less lines of code in the AI applications you built with Keras.
Hadelin: But now Keras has merged with TensorFlow 2.0, so yes, you can use it again for any for anything, whether it is to build a convolutional neural network, a recurring neural network, even a chatbot, an auto encoder, well you can build all these systems which usually would take hundreds of lines of codes in very few lines of code thanks to Keras.
Hadelin: Then the next one that we have is Theano. So, this is a bit less used, it was used at the beginning when doing deep learning. Well, you could build a convolutional neural network once again with Theano. Today I would say that the most two widely used are TensorFlow and PyTorch. Some people are still using Theano of course, but this is not the most widely used AI framework.
Hadelin: And then we have Torch which is the base of PyTorch. So, the main AI framework that is widely used today and which is the one by Facebook is PyTorch which is based on Torch written in C++.
Hadelin: So, PyTorch is, as I’ve just said brought by Facebook not as a competition to TensorFlow because the beauty of this is that there’s not really competition. It is open source. Everyone is trying to bring its added value by you know implementing some research on these AI frameworks and PyTorch was the same as TensorFlow. It has the same applications such as building you know computer vision system or deep learning system. You can also build a chatbot, auto encoder, all the same applications at TensorFlow, but they just work slightly differently. PyTorch were the first to introduce the dynamic graphs which brought some more power, but then TensorFlow did the same, it brought on his side some more innovation and more features which each time makes the AI frameworks more and more powerful.
Hadelin: So, yes, these are the top ones.
Hadelin: And Caffe2, so, Caffe2 was a very popular one to … it stands for Convolutional Architecture for Fast Feature Embedding. This was a very popular one for deep learning. It is still used, but you will understand afterwards that it is less used because by March 2018, it was merged with the PyTorch. It was merged into PyTorch.
Hadelin: So, basically it was very very widely used for deep learning, for image classification, for image segmentation, you know, those kind of applications. But then since March 2018, it was merged into PyTorch and therefore today, I would say that the two main ones, the two main AI frameworks are used by researchers and individuals are TensorFlow 2.0, which was merged with Keras and PyTorch which was merged into Caffe2. Well, Caffe2 was merged into PyTorch.
Hadelin: So, these are the top two ones, these are extremely powerful and you can build some amazing AI application with them and you don’t even need you know powerful computer because you can connect to servers for example the AWS servers to build really powerful and train powerful AIs in a very few minutes or very few hours depending on the complexities of the system.
Hadelin: So, yes this is real revolution and that’s what allows today the research to evolve even faster. Even if as I said research you know is limited by the theory and also by the computing resource, but everyone is building new AI systems every day. That’s how BERT for example and NLP was born. It was of course thanks to all the open source AI frameworks, and so yes, these really help today in the development and the research of artificial intelligence.
Kirill: Hadelin for someone who is an executive and listening to this, what would you say about open source versus commercial stuff?
Hadelin: So open source are basically free and commercial, you have to pay to use them.
Kirill: Yeah, but like let’s say I’m an executive at a large company we have millions or billions of dollars and why would I go with TensorFlow or PyTorch or Python for that matter, if I have enough money, I have all the money in the world, I can just pay for you know other commercial tools without pointing any fingers but there’s lots of commercial alternatives. It kind of feels more secure to me that I’m paying money for something so it must be better.
Hadelin: Yes well for example let’s take the example of Apple and Samsung. So, basically Samsung chose to develop their systems using open source AI frameworks. Apple has its own system. It’s not open source. I guess it’s a business decision, well because of the open source system within Samsung, Samsung might have the chance to become number one and go over Apple because the open source allows more talent to participate into the development of the product and therefore help into the development and maybe grow faster. Apple chooses on this side to, yes not have it open source, but we’ll see, maybe they will gain a competitive edge by taking some directions that are not known and therefore might have at some an innovation that is not shared within its competitors. So, I guess it all depends on the business perspective and the business strategy that an executive would choose to go for.
Kirill: Okay, got you. All right, there we go, so that’s the open source framework examples. That brings us to the end of our webinar. We’ve got a few quick announcements. If anybody listening wants to train up their team or themselves on some of these topics, we have around like I think 60 courses in total in many different languages ranging from artificial intelligence to in this case, we’re showcasing deep learning and national language processing to deep reinforcement learning 2.0.
Kirill: You can find all these courses on Udemy for Business. In fact, there are some courses that are on Udemy for Business exclusively, so make sure to check them out. We’ve been teaching … they’re like over in total together 700,000 students, right?
Hadelin: Yes, that’s right.
Kirill: And announcement number two, you want to do this one?
Hadelin: Yeah, sure. So, Kirill and I are business partners, this company called BlueLife AI which offers two different services. The first one is consulting where we can build some customized AI solutions for you and this can be for any industry as long as we have the data.
Hadelin: And the second service is tailored onsite training or corporate trainings where we train teams and companies to stay ahead of the game in AI and teach them about the most advanced AI solutions and AI models that we have today. So, it’s just a way to stay ahead of the game in artificial intelligence and yes, we offer these two services and we would be really happy to help with BlueLife AI.
Kirill: Yeah, you can find us at bluelife.ai. On that note, I think we can wrap up. A huge thank you to the Udemy for Business team for arranging everything, very very excited and hopefully this was very helpful for everyone. Thanks Hadelin, see you later.
Hadelin: Thank you, thank you very much.
Kirill: So, there you have it ladies and gentlemen, hope you enjoyed today’s podcast thoroughly. If you’re interested in viewing the video version of this podcast with all the slides and our faces of Hadelin and me, you are welcome to check it out on Udemy for Business website. We will link to it in the show notes as usual at www.superdatascience.com/327, that’s www.superdatascience.com/327.
Kirill: Our webinar will go live on 5th of February, 2020, so, early February 2020. Make sure to look out for it, we will send an announcement as well. So, by then maybe it will be a good time to refresh on some of these technologies and check it all out again or maybe there’s a specific technology that you’re interested in and then you can just fast forward to that specific one.
Kirill: Once again thank you to Udemy for Business for arranging all of this and put all this together, record it and share it with the world, very very excited about this. If you know somebody who’s excited about technology as well, who wants to know what to learn, what to look into in 2020, then send them this podcast. Very easy to share, just send the link, www.superdatascience.com/327.
Kirill: Hope you have a have a fantastic year, stick around in the SuperDataScience podcast and with our courses, we’re going to have lots of exciting updates, lots of great things are happening in 2020 and we want to live this year together with you. Stay part of the community, stay amazing and I’ll see you next time, until then, happy analyzing.
