Azure AI: Making AI real for your business - GS009
Our. Conversational. Di Agra. Lets, us put people first. Agitate. Breaks is the best way to do Spock face another takes period. The. Promise, of stack. Overflows, bot built with Azure is, to keep developers, in the zone while they're working on their code by, making these tools readily available through, Azure machine learning I think that Microsoft, is doing the gada science community a great service you. Can. Predict. Potential, failures, and then do the work to, avoid, outages. Before, something disasters, happen a, Shabbat. Service is speeding up the development time, and giving us the flexibility, to skeletally on different channels whilst, maintaining a, unified cognitive, foundation, based on Lewis a speech API. The. Analytics engine that's, been built on Azure cloud, allows. Us to click and analyze the data needed, to deliver the insights our customers, require to stay competitive. Hey. I can help us in social, media and analytics, desktop, stores. Also, internally, marketed, merchandise, we're, very happy that we're one of the first to, do this with Microsoft, and to be ahead of the curve. Please. Welcome corporate, vice president, as you're a I Eric. Boyd. Thank. You and welcome really appreciate everyone coming my, name is Eric Boyd and I lead the azure AI platform. Team and I. Want to talk to you a bit about how, you can use AI really. To energize your businesses. Ai. Has really been transforming. Each and every business around the world we, see this at Microsoft, where we see AI transforming. Each of our businesses. Products like Outlook and, office. And things that you guys know and use every day and when, we talk to our customers they, come to us with new and different ideas every, day about the way AI is impacting, their lives and, we're bringing that all together in, a really fantastic platform, the azure AI platform. To enable you to make AI available, to your customers, to change the way your business interacts, I, want. To start by talking through what AI is and. And really what's advanced, it over the past several years, AI. Really, has come from you, know it's been around 20-25. Years what's, what's why is it so hot all of a sudden it's driven by three primary, changes, the, first is the availability of big data now, with things like you. Know Hadoop, and spark and, data bricks you, can manage large amounts of data very easily, in ways that you never could before, the. Second, is the dramatic increases in processing, that's available both, because of hardware advances, the, ability to use GPUs, and and FPGAs. And new hardware devices to really accelerate as well, as the massive, computation, that's available through the cloud through Azure and things like that and the. Third has really been the advances, in the algorithms, and the models themselves when, you put the three of them together you've, got a tremendous amount of data a tremendous. Amount of computation, and new, algorithms, that have really unlocked, potential, that's never really been seen. Before and it's really changing, everything. In the industry about how you can use AI. At. Microsoft, we've been at this for quite a long time our. Microsoft, Research team founded over 25 years ago has, been at the forefront of, a number of areas of AI really, having breakthroughs, across, vision. Across speech, across, language, we. Were the first to reach human parity in all of those categories, the first for object, recognition to, have human, already the, first for, speech, recognition. The, first for translation, the, first for, you. No question understanding, machine comprehension, and so, these are really tremendous, advances, that we've seen from our research team and then, what we do is we work to take those advances, and bring them directly into the products that we make available in the Azure AI platform.
And We've. Seen tremendous uptake, from that we have over 1.2, million users, using, our Azure, cognitive, services we. Have three hundred and forty thousand, people who are creating BOTS using, the bot framework and, we've seen tremendous adoption. Of these services, really over the last couple of years and the momentum is continuing to grow and, as, I mentioned we see this really moving. To every, corner of Microsoft, ai is really, changing every, business including, ours and so, you see things like Windows, which you don't think of as AI products. But, when you sign into Windows using Windows hello that's, face recognition that's helping, you sign in right there when, you use Skype and it has translation, built right into it when, you use Bing Bing as a product, that cannot exist without machine, learning everything from the ranking algorithms, to how the ads are displayed all, comes, from machine learning to, the microsoft, hololens a, fantastic, product that uses AI to help, compose, the scene figure out where I should put various, objects and how to draw the scene for the users all very. Very powerful, but. Even bringing it a little bit more home I want to go through a couple of examples in our office product, where you can see exactly how, this is touching the lives of our users every day now. We've all sat through really boring powerpoints. This one of course is not one of those but we've all sat through really boring powerpoints. And, by. Using PowerPoint designer you can take your very, ordinary looking slides, and get great recommendations. Powered by IAI to, make them much, more interesting, we look at the text, we look at the images. You upload we. Look for dates and things like that and you can figure out from there oh this is a process. Or this is a timeline. That we're trying to show and really, help you communicate your message in, a much clearer, way and really, have slides that pop from those boring bullets. That you started with if, you, look at word we've, taken word and now you can use diction, to talk to word and have it actually composed the sentences, right there for you or the other side back you can have it read aloud back, to you so, if you're a more, oral hear or you want to sort of hear the words read back you can find the awkward phrases, when you hear someone else say them in a way that you can't when you read them if. You look at Excel the new ideas, tab that's been added to excel looks, at your data analyzes. It and shows anomalies. Interesting. Patterns, interesting. Trends, and really opens up all the things that you can do in it so taking three products, that have been around for a long time at Microsoft, and seeing, how AI is really unlocking, tremendous, new capabilities. For them and we've, learned a ton in the process, in doing this and this, is what we're doing is we're bringing all the learnings that we've had taken from all of the different areas across Microsoft. Of how do you build AI models, how do you have a thousand, developers, coordinating, on a model how do you deploy them in production how do you update them all of that coming together in the Azure AI platform. That, really will enable all of our businesses, and all of our customers to unlock, the power of AI for their applications. For their businesses, for connecting to their customers, and so, I want to talk to you about the, ways that you can do that we see three primary, solutions, that, we can help enable for you the, first is around AI apps and, BOTS the, second is around knowledge mining, and the third is around machine learning, and so I'm gonna spend some time talking about each of these interns, let's start with AI apps and BOTS so.
The Simplest way to get started is to take your existing applications. And add, AI into, them and the, way that you can do this is by adding cognitive, services cognitive, services are rich. Powerful AI models, that we've developed based. On our own first party applications, and are making available envision. In speech, and in language and just injecting, those right into your existing applications. And. So we see this work really seamlessly together cognitive. Services and bot services, and so there are a lot of applications that you can light up even, without being a sophisticated, AI developer. You can take your existing app and then add speech to text in it and so now you've changed the interface that people can interact with you by using a really powerful speech, model, and trying. To train your own speech model is tremendously, expensive and very hard to acquire the data and so using pre-trained, really powerful, ones really, sort of helps accelerate where you want, to go with that and then, the bot service makes it really simple to create interactive, BOTS to interact with your users without, having to think through the complexities, of language, understanding. Of dialogue, flow all of that's abstracted. Out for you to let you focus on your. BOTS experience. What are the things that you're really trying to enable and how can you do that in the simplest, and most seamless, way possible. But. It's one thing to talk about these services, it's something else entirely to just sort of see them in action so I'd like to invite noelle out on stage to come in and walk through it a little bit more detail exactly how these things work noelle. Hey. Thank you very much Eric, all right good. Hi. Everybody thank. You so much for being here I am. Really. Excited to share with you some of the things you can do with cognitive services and, some pre-built models pre-built. Machine learning models the, first thing I wanted to draw your attention to and. Actually I've been a developer and enterprise developer for about 20 years and it's just it's, incredible what you could do today, that you couldn't do even just a few years ago so. If you take a look at our website you can actually go out to a. Kms. Slash cognitive, services and I'll take you right here and you can test some of these pre-built models right, inside the browser so. I just wanted to show it to you really. Quickly if you take, a look you can just scroll down and, see, right. Now it's looking at an image and it provides, these tags this is a web-based service so you don't have to be a data scientist to get access to this you just pass, it in, the URL you pass it in image and you can see I just grabbed, another image and it provides, a different set of tags, and information. About that image now. What. Happens though when you, want to actually do something even more, with this information and leverage, this model by customizing. It well, I thought I would do just that, so, I started. I actually thought about this project it's a you, know we're in Orlando like, Mickey lives here I thought it'd be cool if I created, something that I struggled, with right how do you tell the difference between Chip. And Dale, right, the Chipmunks how do you even know anyone, know. By the way no, all right good you'll see so. This I decided. To build an object detection, model. Based. On yeah. Their noses it turns out chip. Has, a chocolate chip black nose and Dale, has, a brown nose so. Why, not build a model that can detect the two so, I added about twenty images to my model different. Tags for, each you know a tag for chip, and a tag for Dale I then, trained the model and got, some performance, data on how it works and, then. Once, I got that performance, data right, then I could go and actually just do a quick test, I'm. Gonna pass in an image, and. I, just wanted to make sure I grabbed an image I'd use Bing Search just grabbed an image that wasn't already in my data set and you can see right it pulls up now, again as a developer.
I'm. Not a you know data scientist, in any way it, was really cool that I could do this in a couple of minutes I literally built this in less than an hour then, of course I got addicted, to it and started messing, around with the model which you will too I hope now. This, is incredibly, interesting, to be able to do as a developer but there's another use case that, every, business on the planet needs to take advantage of and that's the ability to build a conversational. Bot and you as a developer can do it in just a couple minutes so I'm. What I'm gonna do is actually leverage, and an, application. You can go here as well there's a link. Called aka MS, slash, intelligent. Kiosk demo, all one word and you, can actually pull up this demo and build this yourself if you want to but I'm going to leverage a QA. Explorer. To build a bot in just a couple minutes okay. One, minute, so. I'm going to type in I thank. You oh, you, know what I'm actually going to use ignite. And. Then. It's. Gonna ask me what landing page do you want to demo this on right. So I'm gonna go over and grab the ignite page. And. Grab. The ignite page and put it in and, then I'm going to pick a knowledge base where is the QA gonna come from so I'm gonna actually pick a URL because it's easy but I could also choose a, viii file or. Something. Else you know a QA. Maker project so, I'm just gonna grab this and, paste. It in and hit create now. What's happening, right, now is, that it's going through the process of grabbing. All of the questions and answers and creating. A dialogue for it that you can then edit we, also have a new, utility. Called, chit, chat that allows you to add a personality, to this bot but, most importantly, it's creating a natural language model, so not, only do, I get, to type, in of course exactly what's in the FAQ but I can type in variations. To, what's in the FAQ as well, you can imagine the types of things we can do right HR, finance, we, can create QA, BOTS. In minutes for departments, that need it the most so, you can see here right I've pulled up the site here's my little bot right here at the bottom right just in a couple minutes I can ask it things. Like Oh, what. Do we always ask like what's the, agenda. Right. And if I say what's the agenda it pops right up but, I could also say. Other things like where's, my agenda how. About that agenda right, so if I just type in right, where. Where's. My agenda that's not specifically. In. The. FAQ on the website but however I say it it'll, pop up and provide, the answer. Right. So you can imagine right with object detection and the ability to check between good produce and bad produce or good silicon, and bad silicon, right good hardware and bad object, detection can help you solve this problem as a developer, in minutes, and then with QA maker you can you, can solve for top of mind questions. For your customers, and your employees so. I hope you see in just in a couple minutes how much fun you can have and that you'll go out and take a look at it soon thanks, for having me and I'll give it back to Eric. Thanks. So well. Really. Great to see just how easy it is to put that together and really, to see it in action and you. Know we've been hard at work in this area too there's a bunch of things that we're announcing so. I want to start by talking about the some, of the things we're announcing around our speech stack I'm. Happy to say that today we are announcing the general availability of, our speech service, the, speech service, enables you to do. Speech to text to text to speech and speech translation, and, one of the really cool things with it too is that you, can add your own custom speech, to it and so if you're a doctor. If you're in some particularly, particular. Field that has a lot of challenging, vocabulary, record. A few phrases and upload, them and train a custom, model just for yourself so, that will recognize exactly the type of speech that you want to be able to recognize really. Exciting, stuff we're glad to see this out and you. Know as I said training, a speech model is tremendously, hard and expensive we have one of the best on the planet we have taken this speech model and we've consolidated, all, of the systems internally, at Microsoft, and we're using this same speech model anywhere, at Microsoft, you see speech to text we're, using the same system and so that gives you the confidence that this is a hardened.
High, Quality, high SLA, high availability system. That, you can count on. The. Next thing I want to announce is on our bot framework again. I'm happy to announce that today we are announcing the general availability of, bot framework v4 the, bot framework really makes it simple for you to create, your own BOTS and it all showed you the simple Q&A maker and how how easily you can create a Q&A bot bot. Framework makes it easier to do richer. BOTS to have cards, to have good dialogue, flow in them and, all the full components, of things that you can do to really simplify your BOTS a really, powerful way to connect with your users and to connect with all these cognitive services underneath, them to, provide a really compelling. Experience, for the users that you're trying to connect with and so, let's look at a video of how one of our key customers is using this to see this some of this in action. Adobe. Is committed, to help our customers be successful one, of the biggest things that we hear from our customers is it's hard for them to find the right skills, to use the latest technology. With. Experience, league we can create a single place for all of our customers to go to get access to the best enablement, resources, that oh he has off it uses chat BOTS to really reduce friction. Points as customers, move along their learning journey Adobe. Sees conversational. AI as, foundational. To providing, the personalized, experiences. For each of our users the bot, framework is extremely, flexible it, lets us use our own back-end services, we can typically go from, an idea for a bot to, having a working prototype in just a matter of days cognitive. Services handles. The natural language processing and. The azure bas service handles, sending the messages to and from the user so that lets us focus on our user experience as, I, think about our vision for the future I see the ability to integrate chat, pods not only in experience, lis but to connect that to other surfaces, that were engaging with our customers, we just see tremendous potential for conversational. AI to impact all parts of our business. And. Now I'd like to invite Marta from AI to come and join us to talk a little bit from Adobe, to come and talk a little bit more about the things they're doing what they I hi. Eric thanks Laura. So. Why don't we start by having you tell us a little bit you, know what do you do at Adobe well. I'm excited to be here I'm the vice president of customer success in, Adobe's, digital experience, business well. That's great so customer, success you, know that ties, a lot to one of our core values around, how our success really is only when our customer success you, know tell me more about how that's important at Adobe, yeah. So for us customer, success is all about helping our customers to be able to make experience, their business and to do, that we have two kind of key things that we're offering one. Is the Adobe experience, platform, which provides the basis for content, and data services, for our customers to truly, deliver personalized, experiences. To their customers, and then. The other thing is experience, League which you saw in the video and, experienced League is our guided or intelligent, learning platform, and it helps our customers access, this assess, the skills in their, team's and then provide a recommended training path to help them learn about the new features and capabilities well. That sounds great. You know we're here today to talk about AI, maybe, you could tell us a bit more about how Adobe's, using AI yeah.
So AI is super important, especially for being able to do personalization, at scale and, so, we have Adobe sensei, which. Is a set of intelligence services that we build into all of our products from Creative Cloud to experience, cloud and that, really helps our customers deliver, personalized, experiences. The, other thing that we're investing in is conversational. AI and, so, we're trying to look at more natural, ways that we can interface with our customers, and to use things like predictive, recommendations, based on machine learning and, conversational. AI we spent a lot of time here focusing, on that can, you tell us about how you're using Azure AI services, in your in your BOTS, and then conversational, yeah yeah, so first of all all of our experience, cloud is hosted on Azure which is fantastic but, we're also using Azure, AI. And, the cognitive services in our content moderation, platform, and specifically. With text, and image moderation it helps us call. Out offensive content, the. Next thing that we're doing is that we're integrating, intelligent, agents, into things like experience, league, and those intelligent, agents, use the AI, bot. Service, to, be able to provide more natural, ways to interface with our users and that sounds interesting so tell me more about this this intelligent, agent that uses the bots there's so many how that's having impact across Adobe, yeah. So the Azure, AI, bot service, is kind of the basis for this intelligent, agent and there's a couple of use cases that I want to call out so. What is being able to build, out the profile of the user but not to just ask them a bunch of questions all at once but to pop up questions, in a more natural intuitive, way to, build that profile, so we can provide more. Specific. Learning paths for them and then, the other thing that we're doing is, integrating. Natural language processing from, things like Luis de I to. Be able to do things like and, Q&A maker to be able to at when, people ask a question in the community then they can search, all the past questions like Mike was just shown in your demo and then, serve up the best answer for the user in real-time versus, waiting for an answer from the community that's great so, talking, about it's great can we see it in action yeah, I'd love to invite Ben to the stage to, show us a demo that sounds great thanks Sarah. Well. Thank you Marta and Eric. Guys. At Adobe, we are changing the world through digital experiences. That. Means creating tools that, and empower, everyone, to create as well, as enabling, them on how to use those products so today, I'm excited to show you the Adobe experience, league it's, the go-to place for users, of the Adobe experience, cloud to find everything, that they need to know to, be successful in their jobs in the. Experience league we have intelligent, guidance this is where you can find hundreds of on-demand videos, and guides and webinars we, have access to the community where, you can connect with your peers all over the world we.
Can Also connect you with experts, inside Adobe for service and support now. As you can imagine between. All these categories we're talking about a ton of content, and so, one of the things that we're doing is we're using Azure AI to, help us look through all of that and personalize. The journey for each one of our users and we're, doing that in the form of an intelligent agent today. I'm going to share with you three examples of, how we're doing that the. First. Thing I'll do is dive into intelligent, guidance and in, here I have a learning, guide a step by look at what's next in my learning curriculum as I use Adobe products, I can see all of my guides events, and broken, down by products, as well and the, experience League is built around content, for every type of user whether you're experienced, or a beginning. User, now. As I navigate to look about at Adobe campaign I'm. Learning more about campaign, in the experience, League is learning more about my preferences, as well so. We click on the intelligent, agent what we see is that based on the interaction, I did on the website the, intelligent, agent asks, if I want Adobe campaign, to be part of my learning curriculum, and sure, enough I do and, at. This point the Asian is also going to look at if my profile, is complete and it's, not and it's gonna ask what my role is in the organization, and I'll say I'm a user and, just. Like that through this agent, I'm able to update my profile, and we, did this with the azure bot service alongside, our own custom, Adobe data, so, this is a great partnership but this is really just the beginning of what I want to talk about. Moving. Forward as I continue, to learn about Adobe campaign, I may have questions, they want to ask the community we've, used the intelligent agent to make it really easy for me to engage in this for. Instance I've been running an email campaign lately, and I noticed that my email open rates rates, were not as high as I would have liked them to be so, I'm gonna check with the community I'm gonna say what. Causes low open rates. And. We'll see what we find and sure. Enough were able to pull from the Adobe forums, a piece of data related, to this question now, the, scenario I told you was about email but I didn't say email in the question right, the context. Was able to be picked up by the intelligent, agent and provide, that relevant answer and that's natural language, understanding now. I'm gonna take it a step further and say. Does. It depend, on age. Now. Again we're gonna use the context, to understand, what, we're talking about here we're looking at using, Q&A, maker here to understand, the secondary, context, here this is really an exciting way to have a really human-like conversation. With one of these agents and. Finally. Thinking, about the next step after you complete a piece of learning what happens next how do you continue, to engage and. To do that we're providing, recommendations, in this case we're doing it in the forum of events, here. I'm recommended, through the agents who attend a recorded, webinar now. This is cool but I actually more, of a face to face kind, of person so I want to see if there's anything near me. And. We'll see what happens and, so. When I sent the search in here we're gonna look at all of the data about me and all of the data that we have in the experience, League and understand, that we're providing the context, of an event, that's happening in San Francisco, even though I'm not there right now so think about those two types of data they're bringing together this is really, contextual, and this is really powerful and we're really excited about how we're using the bot service to power this here and I. Just want to take a quick a moment to talk a little bit more about the architecture, underneath, what, we're building here so as, you heard Marta say the entire experience, League are experienced cloud and is hosted in Azure along with our API is we're using Azure CDN. To deploy the content, we, talked a lot about Azure about service, and how excited, we are to be implementing, that and then finally as your cognitive services with, language understanding. And QA maker in the, future we're looking at more in product, around product uses. Of conversational, technology, and when you think about what we're doing for, each user of each, organization. That's. It's really personal experience thank, you.
Thank. You Ben, I'll. Tell you it's it's, really exciting to hear customers, talk about your products it's one thing you know we build them and we envision, the ways that people are gonna go use them but then to have a customer, come and tell you the way it's having impact on their business it's really fulfilling so, very. Excited to have Adobe come and talk about that I, want, to talk next about our, next solution arias or knowledge mining, so. The first an obvious question is what, is knowledge mining, when, we talk to customers and we talk to companies they, have massive, amounts of data that's. Virtually. Unusable, to them they have information that's locked up in you, know scanned. In faxes, in PDF. Files in, you, know old powerpoints. Manuals. Images. Of their product, presentations. That they used to give. Have all this rich trove, of data but they can't do anything with it because it's hard to search you'd have to go through each individual, document it's, not organized, and so, what if we could use AI to unlock, the information, of those documents, and so, that's what we do with knowledge mining, we take these cognitive services that we've been talking about things, that can do optical. Character recognition that. Can turn handwritten, you know lines into, text that, can you know recognize, images find, the entities in it if you had a sales conference in Boston will find everything talking, about Boston, and pull all of those documents together, and then, you create this really rich index, that then you can use Azure search to search across it now we've had a bunch of companies who've worked with us who can just completely, unlock, the information, that they have in there and find. Really rich value, out of this documents, it was literally just consuming. Disk space in their in their enterprises, and. To talk about this a little bit more I'd like to invite out mohnish, from I started us to come and talk some more about this. Thank. You thank you Eric great, to be here. You. Know you, could say I saw this was born in Asher we, have been working with the platform for a long time and it's. Interesting to see where. We have come with the platform. We. Have, if. You look at our customer, list we've. Actually, been. Trusted, by the most iconic brands. In, the world, there's, Microsoft, as well and we. We, have taken. Contract. Management to the next level we manage any and all contracts. At. Enterprise. Scale, but. Look at some of the numbers at the bottom we. Manage more than five million, contracts. We. Manage more than five hundred billion dollars, in, contract, value it's. True hyper, scale all running, on Microsoft Azure, and, we. Believe that I certain, that contract management is so critical, and so transformational. Because, for, the first time in history. Contracts. Are being digitized, and it's not just taking paper contracts, scanning them and making them converting, them them into PDF or. Images. But, actually, taking the unstructured, data that, is in the contract, and converting. It, structured, data that is actionable. Contract. Management is being reimagined, using. AI and what. Is interesting to see is that, when. We do this transformation, when we do this conversion. We. Can actually take, the. Intelligence. That is built into, the contract surface. It out and, for. That we, use as your cognitive search because. It helps us bring, that inside back to the customers. So. Let's take a real scenario, if. I were a contract, manager in AI service and I. Got, a contract, from a customer an email maybe a PDF or an image and now. I have to look into the contract, and I, have to see what, is inside the contract, if I have to understand, risk everybody's talking about the gdpr for example the. General, data protection regulation, if, I have the cop if I have the contract, and the contract has, a clause that is actually, risky, I need. To surface it up but look at what I have to do I have to send it to Finance I have to send it to legal, I have to send it to the risk person, who's an expert at looking at that language and, it's. Just so manual, and so, labor-intensive. But. If I could take that and cognitive. Search helps, us do that if I could take that contract that blob of text and bring. That insight. Out, of that contract, identify. The clauses, that, actually, reflect. My risk in the contract, and risk. Is just one example it's compliance, it is finance, your risk its, contractual. Risk all kinds, of risk then, all of that can bubble up into, the contract management space so, let's make this, real. Let's. Go to the isalus app. I'll. Go to agreement, management this, is where the repository, of all contracts, in isit is actually. Comes in you'll find all of the contracts, in an enterprise in. A one searchable, common.
Global, Repository, I've. Already uploaded the. Document. That I got from my customer, and. This. Is where I land up on the details page the, details page is where the entire history of the contract who are the people who work on that contra what, are the associations, what, are the risks what are the obligations what. Are the clauses inside the contract what does it actually mean so. Let me go to class discovery, and let's see what is't is did with this contract, and remember. A contract, is a document, it's. A blob of text we, took that text and we broke, it down into. A list of clauses that. Are easily managed, on an individual, level, so. Let me click on one of these clauses. Let's. Say there's an amendment representation. Whatever that is and now, that Clause actually, tells me what. Risk, I have it. Can say it can tell me what, are the things that are inside that class for example cognitive. Search, allows me using, named entity recognition to, bubble up locations. Places. Organizations. And I. Know what, is in there you'll. Also see that using. Microsoft. Translate, services, I also. Have a translation, of that clause in French I also. Have that in German but, it's not just translation when it comes to contract management now, I have a very powerful tool, in my hands because if. I'm an English speaking person and I type in let's say I'm looking at risk and I say show. Me all the contracts, which have data, privacy, as a term. In them and I do this search across my global contract, repository, not, only do the English contracts, that contain that term come. Up but, the. Contracts, in French in Russian, in German, a dozen, under languages. Actually. Get searched and I bobble up all of those contracts, from a global depository it's, really really powerful. But. You know what most people, in the contract management world actually. Spend a lot of time, in Microsoft, Word and so. Let's go to word let's download, this document and open it in Word. Now. Here's, how the word document. Looks like when we have an add-in let. Me go in here, and. If. You look at this. Zoom. In a little bit you, see the same structure, that you saw in the browser actually, surface, inside. Word so. The insights. That you get are now right where you need it, this is an example of pervasive AI and this is where I can actually make this service call from, word, into. Cognitive search to get all of this information you'll, see the clauses that are actually. Something. That I have to really look deep into which. Are flagged, in red that could potentially be, my GDP, our risk in that contract, and, it makes it super super. Easy to, do this now, the, other thing remember we, also use named, entity recognition so. We have people locations. Places. Right, inside. Organizations. Right inside our, clauses. So, let, me do one quick thing let me click on locations, it actually goes and filters, all, of my clauses to show me only the clauses that actually, have location, information in them, and. Let's see what this. Can do to help a contract. Manager make. Its make it sly make their life easier so, for example. When. I look at this I say here I found a place bellevue, which is okay because, that's our headquarters. So safe enough but, when I go up top I interestingly. I now see Germany and Russia now, what is Germany, and Russia are doing in a contract, which is between two US companies, because. That's where my customer, was so, that makes it a little bit more interesting, so if I go in there and then I click on the class text, right.
In Microsoft, Word what. Is interesting now I see is hmm, this, is interesting there's. Actually, an obligation. For me that, now, cognitive, search has detected, that says that. You have to be compliant, and do a third party audit to. Make sure that you comply, with, data privacy. And security laws in Germany, and Russia now. That came from left side actually. This this was interesting I would have not, been looking, for this if I had not bubbled, up all this, information, that came. In right in Word but, what is interesting is now I can, take care of my habla of my compliance because, I can be more careful, I know that, I have to be compliant or I can negotiate or, at, the other side of it is of course there is a financial, impact as well for this because, if. I have to comply, with these laws in two different countries after. Think about storage, I have to think about operations, I have to think about the, compliance requirements that the third party audit can take together so there's a whole lot of stuff that, actually goes beyond, that. I certa. CI an azure cognitive. Search actually surface, very quickly so, that's the power of. Of. Kind. Of combining UI AI as well as. Surfacing. That AI into. An obligation, into, an. Application. That actually gets, in right. Where you need it that kind of information right, where you need it. We. Use about a dozen. Assured. Services, across the board more than a dozen actually, it, makes it really really powerful for us very easy for us to use and we, solve. The most complex, contract. Management problems, for some of the largest companies in the world that. Makes it really interesting, and really powerful it actually, is. Changing. The face of contract. Management as we, see it that's. It that was the tip of the iceberg I just wanted to make sure that you actually got. A good feel. Of how, we, do it this way thank, you. Thank. You Manish. This. Is one of our fastest, growing product areas where we have companies, really all over super, interested, in how they can unlock the power of their data and. It really ranges all over from you know the fast-paced world of compliance contract, management, all the way to the NBA and so we're really excited to see that the traction we're seeing with this I want. To move on now to our next, solution. Area really, focusing, on machine, learning and, so the previous time sections, talked. About a whole bunch of things that are built on top of this platform the, applications. The cognitive, services but. In the end of the day if you need to go and build a custom AI model, we have a really powerful platform. That we've built to enable that and so I want to go through the. Way that we set this platform of how we think about the different layers and how they all compose together. So. When we think about our machine, learning platform. We think about it in five layers and so there's the top layer where we produce really. Sophisticated models. Either internal, models that we're going to run customers, third-party, models or a rich, set of sophisticated. Pre-trained, models the cognitive services that we make available to you the. Next layer down is the frameworks, where, you can use tensorflow or PI torch or any framework, that you want and we really have worked to create onyx, to help tie all of those together and, then. The next layer down is, the, powerful services, and how you can use Azure. Data bricks or deep, learning VMs, or you. Know Azure, machine learning to, really manage, your workflow, and and really help build these models and, then under that we have the infrastructure, lots, of really powerful hardware, from. CPUs, and GPUs to, FPGAs. To really accelerate either. The training, or the inferencing, of these models and then, the base layer is the deployment layer making it simple to deploy anywhere. You need to and. So we're gonna talk about each of these layers in turn obviously, we've already talked about the top layer with the cognitive services section so. Let's talk a little bit more about the the next layer down actually, we have a whole bunch of investments, that we're making I forgot to highlight that we've been making a whole bunch of progress on these so we have new things to announce in, all of these different areas so we'll definitely be showing, that, but. Moving into the frameworks lay and talking about that one.
Of The things that's really accelerated. Machine learning, over the last few years has been this openness. How everyone, is sharing the, software, is open source the models are getting shared the everything's, getting published and this has had a tremendous accelerant. Effect on the industry where machine learning is taking off way faster, than otherwise, might have been expected, and at, Microsoft, we're really committed to this we want we hear from customers all the time that, they want to be able to use whichever tools, whichever, framework, they want to use they don't to be tied down to a particular way of doing things a particular platform, a particular, piece of hardware and so, this has been a big commitment of ours to support really the ecosystem, of frameworks, that are out there but. There's some challenges that come with that because supporting, multiple frameworks now, means if I want to optimize each of those I have to do individual, work with each framework, to go and make it work well with the hardware and, so, we, worked internally, and we started talking to others in the community and, worked with Facebook, to create onyx. And then invited a bunch of others Amazon, and Nvidia and Intel. And Qualcomm into to cooperate, on this and it's been a really fantastic partnership. With all these companies, onyx. Is a exchange. Format, where you can take a model, from one format from one framework, and easily. Convert it into the other framework, and it dramatically, simplifies, things for hardware manufacturers, because they can optimize for onyx and then sort of see the acceleration, from there so. We see a lot of great things that are going on with that to, talk about in a little bit more detail I'd, like to invite Jerome out from Facebook to come in and talk through some of the things we've done in our partnerships. Thank. You Jerome thanks for Cowen's. So you, know I think people are probably pretty familiar with Facebook but, maybe you could talk a little bit about how, you guys are using AI and Facebook. Go. To the main Facebook app and, when you go to Instagram, what, you see is actually determined by a pretty sophisticated, machine. Learning algorithm, that learns what, your preferences are and try to match that to the content, is posted on the platform, more. Than that you know every ad that you see on the platform all our revenue actually come from a similar algorithm. Driven. By machine learning that learns what, are the most relevant ads for you and then there are a bunch of experiences on the platform, driven by it you know six billion translation. A day are done using, automated, machine translation, if. You can see very well the, system will actually look, at the pictures and describe them for you our videos it's. Also able to identify when, people need help or, people need and blood donation only no we actually do 200, trillion inference. Really, write trillion. Inferences, a day on the platform Wow that isn't a very impressive number, it's. Interestingly. Here, how you know much as we're seeing at Microsoft, a I sort of in going into each pocket, at, Facebook you guys have sort of the same thing you. Know I was talking about our partnership, with onyx you know tell me a little bit more about how honest has been useful to you at Facebook onyx.
Actually Came from a real world problem internally we had a platform, that. We use for production workload, called cafe 2 and then, we had another platform used for experimentation, exploration. In kanpai torch and the prime was how do you go from one to the other and, that's what we create actually this format, that lets us you know have, a common set of operation, between the two system more. Than that we needed to also interact with all the hardware vendors, and the device makers, you know in the cloud and in. Mobile, to, basically get them to optimize a set of operation, so we actually create an ecosystem around, it partner, with Microsoft and I believe 15 other companies to, develop optimized system for monix yeah, it really mirrors a lot of our experience, at Microsoft, where we. Had you know different teams choosing, different frameworks, that they wanted to work on and then needing to figure out how to make them interchangeable. Or how to operate, against hardware really, was begging, for this, this interchange format, the basically. Begging for onyx and so it's. Been it's, great to sort of see across the industry how everyone's sort of rallied to this cause you. Mentioned hi torch a minute ago pi torch is also something we've been collaborating a lot on together you. Know making PI torch really available across the Microsoft, stack tell, us a bit more about what you guys are doing with pie charts pie charts is an interesting project it was a skunk, project, that Facebook started, just two years ago five. Engineers we wanted to make deep learning really much more usable for developers, and researchers and, just. In two years that kind, of to fire and now. You have lots of researchers one of the most used framework, for research, now. What we found out and then the past few. Months is that we need it also from where I can allow it to go from research to production and, we decided to combine cafe, to invite or together with onyx and create, a one unified platform to, allow actually more, self and every company to go very very quickly from this deep exploration of deep learning to. Production, workloads, yeah, we're seeing similar things pie torches. Certainly. Becoming the most popular framework, internally at Microsoft, is people are really adopting it because it's so easy to get started with and really to move from there all the way into production, and so, you, know we're really embracing, it then as a platform and making sure that it works across. Our platform, so, tell, me if people are interested in pie torch how can they get started with it today why they can actually go to PI, tours at all and there's actually preview release if you're an expert users of. OneNote or today that just came out actually last week and then, obviously we have a partnership together so actually PI touch is integrated, into Microsoft. Products so the data science VM as. Your ml as. You collab actually, as, your notebooks are actually you, know using a PI today and supporting it yeah no it's great so, you, know really looking to wrap, up I mean thank you for coming but I'd love to hear you talk about you know where do you see AI going in the future what's what's coming next what's interesting you know everybody, talks about AI but, at places like Facebook you know five years ago we, didn't, have any I team five. Years on actually, AI is pretty much everywhere, I mentioned, you know 200 trillion influences, so, this evolution, from being just a skunk. Project to begin court. All the to. The whole company I think we believe is something is gonna happen not just for us but for Microsoft, but for the whole industry no it's great we see exactly that internally, at Microsoft, and we see it with so many of our customers where, many. Of them are not quite as sophisticated as, Facebook is yet but, they're starting that journey and and more and more of their workload is moving to AI and so it's it's really exciting to see so, well thank you so much for joining us today and look forward to continue to collaborate lo with you all right thank you thank you.
So. As I continue through the the sort of five sections, that we think about in our AI, platform, the next one I want to talk about is the machine, learning services, and so, you, know if you're an AI developer, and you've been using your. You. Know building your model or creating, your data and now I actually want to go and train the model you, know there are a couple different ways that we can work with this on the one hand we see fairly, sophisticated users. That have large data sets, that they want to use Azure data bricks that gives a great spark environment, to go and train their models and, then Azure machine learning to, go and manage you know the training, the iteration, the deploying. To a lot of compute and the like on the, other hand we see a number of customers who have you know much smaller models, that they're looking at and maybe not as big a data set and so they're gonna operate mostly, just in Python and, so there we have the data science VM that they can go and spin up and, use directly, you know brings. All of the frameworks. And and environments, that you would need to be really productive right, from the start but. We have a lot of things we want to announce around, Azure, machine learning that. We're announcing an in preview, today one. Of the things I'm most excited about is the automated machine learning and hyper parameter, tuning this, is a service that's going to make our data. Scientists, way more productive it, really helps them select. A model and find the right parameters to, it to really make it work much, much more quickly and so we'll talk a little bit more about that in a minute another. Thing that's very exciting is the distributed, to training. That you can do now. You can take a model, and distribute, it to a cluster, of hardware. Of CPUs, of GPUs, of what have you to really accelerate your training, and get it done in a much more efficient, manner, we've. Talked about FPGAs, FPGAs, really, help accelerate the inferencing. The the serving, of models and we've. Now, added a number of capabilities, for more additional models that you can now run on the, FPGAs, in production. As. Well as we, have a deployment. System, that really makes it simple to deploy anywhere. You need to one of the really challenging things is I've, built my model I've trained it and now I need to deploy it maybe, in hybrid, maybe I'm Prem on the edge in, the cloud and and, managing, the differences, that I need in those is a lot of overhead and we've, made it really smooth, and simple to deploy to a docker container, use, our kubernetes, and deploy, it anywhere you need to so, a lot of really exciting things that we're talking about they're all available in a Python SDK we. Heard from a lot of our customers that. They want to use the, tools that they already are using if they're using a jupiter notebook if they're using data. Bricks notebook if they're using. Vs. Code if they're using PyCharm, any environment, that they're in they, don't want to have to switch to manage their model training and so by having an, app I thought based SDK makes it really simple to go and tap into all the power that we bring together in the azure platform. So. I talked about automated. Machine learning and so I want to talk a little bit more about that you, know building from Scott's example, this morning Scott was talking about the work that a data scientist, would go through to predict the price of a car and the.
Things That Scott's talked about are you, know the way that you start is you you go and you gathers you prepare some data and then you build a model and then you train it and, so if you think about the work that a data scientist has to do it's, fairly time-consuming you, start and you collect all your data and the first thing you need to do is identify which. Set of features do I think are going to be important, if. I'm training a new model for a car do I think it's going to be the year of the car maybe, the mileage, maybe some assessment, of condition, the color can affect the price and. So figuring. Out those and then now there are a whole bunch of algorithms. That I can use and they, all work differently, some of them are better with certain data sets some of them have you know different characteristics. To them and so what the data scientist has to do is use their expertise, to figure out which algorithm that I think is gonna work best in this particular environment, and then the next thing they have to do is each algorithm, has a whole host of tuning, parameters, that you need to set ranging. From the learning rate a number of levels a whole host of things and so the data scientist goes and sets all of these individual, knobs they train a model and they get some results and if the result is not performing, the way that they expect it to then, they, need to go and iterate and they go back and they pick another, model let me choose another set of features let me choose another algorithm when we choose some other parameters, and train again and they. Repeat this process over, and over and over in a very tedious fashion, too before they finally come up with a model that they think is the highest performing model that they can have and so, the question is what. If there were a better way what, if this could be automated and so that's what we've done with automated, machine learning, instead. Of needing to go through that tedious process where. You're using all the expertise, the data science has around, how do I tune these hyper parameter tunings and how do I use the, output of the previous run to inform the next one now all I need to do is start with here's the data set that I want to use here's. The thing that I want to optimize I want to predict the price of this car here's a list of you know the last thousand, car sales and and, please go in and choose the right one, and what automated machine learning does is, it uses itself, a machine learned algorithm, to go and recommend, what's, going to be the best, you. Know set of parameters, is best algorithm, that it can go that it can use to train this and, it will go and it's do its iteration as well learning. From. The runs that it does on your data and the more that you use this algorithm and more that you use automated, machine learning the more it learns what, works well for your data and so, it's very efficient, it works really well at, making, sure that it you, know doesn't spend a lot of time and, compute, on bad, cases, on failed runs you, know a naive approach that, you'll see some other people doing is just to try everything, it's. Hugely, expensive from a computational perspective and it takes a tremendous amount of time this, works much faster and, much cheaper and so, is a really much, more powerful, way to, make. Your data for scientists more productive, you can sort of accelerate, is this data set learn about what, are the key features that are going to train on it what are the key sort of algorithms, work well on it and just give them a tremendous boost and their productivity, and so, we've been building in real end and working on this system and so it's available today in public preview but, even more exciting than sort of talking about it is you, know we work with customers and developing, things and so we've, been working with British Petroleum to, come in and test, this out and really make sure that this works the way that we expect it to so, I'd like to invite Manisha onstage to come and talk to what British Petroleum's, been doing with, automated machine learning. Thank. You Eric hello. Everyone my name is Manish Naik and I work as a principal, in digital innovation organization, in BP our. Team is responsible for scanning, and applying external. Emerging. Digital technology, innovation, to group's requirements. Now. I'm sure most, of you have heard of BP BP.
Is One of the largest energy companies, in the world operating, across the oil and gas value chain if. You look at the infographic, starting with upstream, where, we are in the business of exploration, and production of oil and gas moving. Into the midstream part where we have a complex, supply chain a world-class. Commodities, trading business further. Into downstream, where we have manufacturing, facilities all over the world a world-class, refining, and marketing business. As well as a growing, renewable energy portfolio now. As you can imagine across, this value chain we generate a lot of data we have lots of examples of data science and machine learning projects, today, specifically, I'm going to talk about a very simple, but powerful example. From upstream, where, we have a strong partnership with Microsoft, and I'll, also touch upon our experience with automated machine learning let. Me flick to the notebook, environment. So. This is essentially the data scientists, play area and what. You're looking at here is a representation. Of a hydrocarbon reservoir. What's. Reservoir, simply. Put a reservoir is an accumulation of hydrocarbons. In the Earth's crust. What. Is of most importance, to us here is to be able to predict how, much of that hydrocarbon. In percentage. Is actually, recoverable this. Is called recovery factor, now. This. Is a very complex, prediction, lots, of data involved, lots of different geological, properties and we. Have world-class, reservoir. Engineers who used skill and intuition, to make this prediction today however. This, tends to be a little bit manual, resource. Intensive, and is, subject, to some qualitative input, we. Decided to explore machine learning as an approach to figure out two things one, can. We improve the quality of the prediction and two can, we remove some of the human bias, let's. See how we did this. This. Is a fairly standard data, science workflow, and the, first step what I am doing here is. Loading. The data now. The. Data, folks in the room would know that the, task of. Collecting. Organizing. Cleaning transforming. The data getting it in shape to do machine learning is is, a massive activity in itself not, to mention some of the feature selection activity, which has happened before, the. Next step after late loading, the training data is to find the best algorithm, for this data now this is where the skill and expertise, of the data scientist comes in this. Tends to be a fairly trial. And error driven, process but in this case we know that this is a prediction problem we. Know from experience of.
Working With this, data what kind of algorithms, are going to perform well so, we start with those algorithms, as you can see here the. Data scientist has six. To seven different algorithms, to work with and. These. Are the results from those algorithms, and. This. Plot suggests. That in this evaluation. Random. Forest regressor seems, to be the best performing model, now. This, is just the starting point and depending, on the amount of data and the type of problem, this, can. Be done tens if not hundreds, of times now. That we have a set of algorithm, to, work with the next step in the data science workflow, is to tune the algorithm, to fit the use case better, in. Machine learning parlance, this is called hyper parameter tuning and again. This is a very iterative process and relies heavily on, the data scientists, skill. So. These are the different parameters in, this case that the data scientist is trying and the, way this works is you test different parameters, back. Against. These algorithms, set. Of algorithms here, and you, try and plot the results on the basis of. Which algorithms. And parameter, combination, are showing, the best score and least, error. Now. That what, I've described to, you right now is a fairly. Standard. Machine, learning process and, as, as. Hopefully, become clearer it, tends to be iterative in nature and relies heavily on, the data scientists, expertise, can be time-consuming as, well, now. We. Have some. Of the strongest data science teams in upstream in, BP and it is in our best interest, to ensure that the data scientists, are productive and they, have the right tools in their arm oh that's. Where automated, machine learning comes in let's. Look at how the same process that I described above is done, could, be done using automated, machine learning what. I start with is I load the configuration and, here I'm just giving automated. Machine learning broad direction, I'm telling it this is a prediction problem and not a classification problem and, I'm, feeding it train and test data along, with some other parameters. After. That with. Just this one line of code, automated. Machine learning on, its own runs, through the different algorithms within the prediction family, and the. Different parameter combinations, that were manually tested. By the by the data scientist so. As you can see here this, is the result of some of the evaluations, and you can see that it's going, through not just different algorithms, but different parameter combinations, as well now. This is where the, power of the cloud comes in it, would be tricky to run this on your local machine but, from the local machine itself, you, could auto. Scale on the cloud using some of the flexibility, of of, the cloud environment. If. We look at some of the results from the automated machine learning then. This, this plot shows us, that, not. Only in terms of the actual algorithm, but also the, metrics, and the score the, results are very much comparable to what the data scientist produced, there. Are various, different ways to plot and visualize these results as well within the notebook environment. And. Before. We get to the actual output I just want to say by no means, am I saying that automated machine learning always will. Give you the best outcome. Of the model how, hopefully, what is clear is it's. A very good starting point and it will help reduce some of those iterations that we looked at before now. Talking about the model itself I can tell you that this model is in production in BP as. Some of these plots suggest. The. There's, very little difference between our training and test samples, from a variant standpoint which. Means that the model is performing well. There. Are about 300, or so experts using this model on a regular, basis and we have till, date made, more than 1500, predictions. In. Summary. We. Think that automated machine learning has the potential to become a core part of our data science process and I, think we could derive value in three ways first being fairly obvious, it, will increase, the. Productivity of, our data scientists, dramatically.
Increasing. The time, to market for machine learning projects. Second. Is, as. Our, data scientists, continue to work in their existing, fashion they. Will also run their models against, automated, machine learning and the. More results, they get that are comparable like this the. More trust they will develop in automated machine learning in which case we would we would develop a benchmarking, process, which. Which would suggest that every machine learning project. Model. Could. Be run, against, automated, machine learning which kind of exhausts, the what-if scenario that data scientists, always have which is what, if I tried this different algorithm, with this different combination, an automated, machine learning could take care of that thus improving, quality, lastly. We, have a lot of domain experts across, the group in BP these. Guys are data heavy analytical. In nature, not. Done, machine, learning before but we think and the still early days in our thinking with this but we, think that this could empower some, of our domain expertise, experts, to leverage machine learning and we, think that we could allow them to experiment on their own in a controlled environment and, perhaps pass. Their results to a central data science team who could productionize, it so. I hope I was able to give you a glimpse of BP our thinking. Around structuring, machine learning projects, and how. Automated, machine learning might fit into this thank, you back, to you Eric. Thank. You. Thank. You manish again. Very exciting, to work with a customer and and instead of just thinking, about how great this can be to see it really adding value to a customer in their, business so, we. Think this is gonna be a really powerful technique, and. We're looking to see this accelerate, and add a lot of productivity, to data scientists across the board as well as broaden, the availability, of it so we're very excited about it, going. Back to our machine learning platform, and the five different layers if we're now on layer four and, we talk about the powerful infrastructure, that we have so, this is the beauty of the cloud as it brings all this compu
2018-10-03 17:04