 Hello, folks. How was everyone doing? Good conference so far? Very good So I was listening to the bioforma for Sudeep the previous speaker and it looks like he went through like 10 different companies in 10 years And I'm thinking, you know, I spent the last 18 years at IBM. Is that a good thing or a bad thing? But then he said one thing which really was interesting. He said No offense, but you guys have suddenly because started becoming cool No, okay, I'll take that Because you know, you guys are doing open source. You guys are a catering to developers. I'll take that. Thank you Sudeep so in my role now I actually travel a lot I meet with customers and Work with them on real machine learning use cases so going beyond the the theoretical stuff to actual implementations and I wanted to share some perspectives with you and Also what I see in terms of a huge role that open source is playing in advancing the technologies Now I'm sure, you know pretty much everyone knows the fundamentals But I'll take maybe a minute What is machine learning there is the the definition on Wikipedia that goes all the way back to 1959 You know, you know computers that learn without being explicitly programmed There's Arthur Samuel from the time when he built the the checker program right at IBM You know computers that learn automatically sounds kind of like Skynet So I Tend to describe this in in in my terms. I say, you know, it's basically about Understanding patterns and data using algorithms now it goes a little bit beyond that you may have structured data You may have unstructured data and you're essentially trying to find patterns and insights in that data And then maybe in certain cases use those Use that understanding to predict what is going to happen. So you know about supervised and unsupervised supervised when you have Labeled data with the known outcomes. So as fraud already happened your does your historical data already have that information Was across the law for successful? So if you have that historical known outcome data, you use that to create a model Understand patterns and use that to predict and then unsupervised where you know, you don't even know what the Outcomes where you're asking the algorithms to try and understand the relationship between Different parts of the data understand cluster. Okay, so that is the basics now as as I interact with customers what I see is very Innovative and interesting use cases. So some of these are obvious ones like in the finance industry fraud detection Is is like the number one thing any bank you walk into? This is the number one thing that they want to focus on now going from batch Scoring of of credit card transactions to in in transaction real-time scoring predicting fraud There are other very interesting use cases healthcare, right? So using the ABCs and predicting You know the chance or likelihood of you of a patient becoming diabetic preventive interventions Trying to avoid emergency room visits by predicting certain lifestyle Outcomes these are all happening as we speak across across Industries, I thought I'll share a couple of these And also then start talking about The role of a spark that I see you're happening in here. So this is a this is an interesting one This is not a traditional Industry, this is a company that works with inner city buildings inner city Businesses in improving the energy efficiency of the building so they will work with non-profit organizations churches and the like and Help those those businesses understand the energy profile of their building and help optimize it Their current model is that they will send an engineer or a couple of engineers They go out and they sit down and they measure and they talk and interview the superintendent of the building and collect a lot Of data and then they've got a pretty sophisticated model which which tries to calculate what is called the energy unit the the efficiency of the building Eui and and then it comes back into okay, how can I improve this and so on there's a problem with this Sending engineers to each and every non-profit is not a very scalable model The non-profits and churches usually cannot afford that initial assessment in in a few thousand dollars range And they want to scale this they want to scale this operation Now the way they did this is and it and this is a there's an actual thing that as they have rolled out It's very interesting instead of sending engineers. They make a mobile app available to the building superintendent so the building superintendent has This app on their phone and instead of asking and answering a bunch of questions, you know you start the app and The app tells him to go outside and take a picture of the building from outside Then go down to the basement and take a picture of the boiler take a picture of the the heating system take a picture of the cooling system and all of this information is uploaded into into the environment and Basically pieces of information are extracted from those images Fed into a machine learning model and it comes back with a score of what our efficient it's building is and Then if you want further intervention if you want the company to go out and improve the energy efficiency That's when they engage So at the fringe at the at the at the mobile app level now that is using deep learning to understand the visual images You know things like oh, you know that building is made of stone has got large windows and it's got a high ceiling to base ratio And this does not have a lot of shade on the sides. These are all being extracted out through custom visual recognition Models and then it builds a set of profiles or attributes in machine language machine learning Lingo it is actually building out the feature set that is going to be used to score against the model that has been built and deployed Coming back with a score very interesting use case Next one is is another interesting one what we call a celebrity experience. So in here the idea is that You have court France up by the way, that's a mainframe. It's an IBM keynote, right? I got to put a mainframe up there somewhere But all kidding apart, you know, this is where most of the enterprises have your core transactional data The question is how will you use that data in new and interesting ways? So what what this scenario is about is, you know, you've got a transactional history You've got credit card transaction data there volumes of it So you kind of have an understanding of how this how customers are using credit cards But couple that with other data social media data weather data, etc And build in real-time offers for the customer as they are transacting with the bank Now what what we call a celebrity experience for a story uses that phrase celebrity experience where the end user feels that He's a celebrity that the business knows him personally and is not treating him as a segment But it's treating him as an individual with an understanding of what his likes and dislikes are So lots of lots of variations of this in the retail industry in banking and so on where you target offers or to provide a Customized experience to your end user through using machine learning now machine learning itself is not new So what exactly is different now, right? So I think there are three major forces that are coming together to make this happen If you go about this in reverse order from what is on the screen Availability of different types of data You have transactional data all along But now when you when you supplement that with other types of data, for example all sender records Who would have thought calls and the records could be a valuable? It's a pressure throw of information that can be used to couple with your transactional data social media in a Twitter or what a Facebook etc etc blogs Whatever that your customer has put out there in the social sphere These things become Additional data that can be brought into enhancing your model now external data may not always enrich your model It is an exp it is typically a data scientist job to see if adding external data and Riches your data enriches your model and improves the accuracy or performance of the model Advances in compute performance and the wave of open language. Let's talk about a little bit about that if you look at About a decade ago So I said machine learning itself is not me But what most enterprises spend their dollars where in was in the infrastructure needed to run those algorithms The the amount of money spent on compute and storage was huge And there was not enough money left for investing in where the actual heart of the matter was the algorithms But now with advances in compute and storage technology that is not the challenge Per se you have much more room much more Headroom to invest in in terms of your actual algorithms natural development, especially with advances in GPU technology for deep learning Especially coupling that with your standard processors. You're all probably familiar with the With the GPUs from NVIDIA and how that can be part that can be paired with standard processor technology for example with the power processors couple with NVIDIA using the NV link an order of magnitude Improvement in performance, right? So you've got much more Ability to innovate in the actual algorithms and the and the neural nets and so on And the other part that is happening is that data science has become a team sport It used to be especially if you're doing SPSS or SAS or one of the existing technologies There was a very contained space you needed a PhD to understand that space And I'm just kidding. You don't need a PhD But you know you you need you needed to have extreme expertise in that specific area And it was very difficult for the business analysts the app developer the data engineer to collaborate on machine learning projects Now where the industry is headed is this collaborative idea team sport where all of these people can collaborate So the concept of democratization is catching on And at the at the sort of the star in this gallery of players is a batches park So you've got many different open source offerings and frameworks and languages out there you got TensorFlow for deep learning you got scikit-learn spark is becoming sort of a star in this in this gallery of players for machine learning and and IBM is Very consciously Embrace this we are we have made a commitment to embrace spark And we are doing that commitment in one of in in two ways The first one is a commitment to contribute to this ecosystem The second is a car is a commitment to adopt So if you look at the first one the commitment to contribute, this is where we talk about This park technology center is an investment where we've got a bunch of bunch of developers who are constantly their only job Is to produce code that is contributed to this park ecosystem If you look at some of the things that have happened, you know some of the metrics, you know Lines of code contributed to spark You know the number of commits in spark 2.0, etc. And the increase in the level of contribution to the spark ecosystem It is incredible. I believe we are the number one contributor to this ecosystem right now Especially in the machine learning space if you look at the top three Contribution areas you'll see that there is spark sequel itself then there is spice park So python is the most popular language for what one of the most popular languages for machine learning and Bringing the python and spark environments together with pike spark is a is a large area of contribution and the ML libraries themselves these are the top three of our contributions amongst the the the overwhelming high contributions To the overall spark ecosystem. So that is the commitment to contribute and that will continue But we also have a commitment to adopt. It is not just that we write code and and and donate it We actually are embracing this internally off-premises on-premises on the cloud on the mainframe I was not joking about the mainframe by the way it was it is actually the first on-premises platform that we put spark on and put broad machine learning on so on on on my on mainframes on other hardware platforms in Connectors to various data sources for connectors Cloud services you have spark as a service in two clicks you can get a spark service provision for you. You can you have Beta science experience, which is a platform for data scientists, which makes machine learning easy And and brings a collaborative team support effort to to the data science If you look at data science experience and what we are doing there I would if you have not tried that out I would highly recommend that go on data science or IBM comm and try that out you you get a Variety of options to play around with data science and machine learning constructs both hardcore Programmatic approaches in Jupyter notebooks as well as visual guided interfaces that let you play with these environments If you if you're if you're building ML models It'll actually what guide you through the process of creating ML models You select and choose between different options in a drop-down you try out your data You look at the accuracy that is coming out of the out of the models and pick and deploy one and the deployment is actually pretty cool Now when in one click the model gets deployed into a spark environment, but beyond that we get to this concept of operationalizing What what I mean by that is if you if you have deployed a model. It's well and good. It performs great Area under the rock curve is great or you know It's wonderful, but a week from now the model starts degrading and you don't know what's going on So the concept of constantly monitoring the model Looking at the performance of the model and providing that as a feedback loop into in a close feedback So that the model can constantly be retrained if you need to is Pretty interesting aspect pretty interesting concept So that the model that you built today and performs well today Continues to perform well six months from now as new data comes in and as new patterns evolve in your data This concept of deploy monitor Constantly improve your model is something that I would suggest that you take a look at and Spark basically allows this connectivity to different data sources provides a pipeline for your ML But it also enables very interesting use cases Blockchain is another hot topic these days So here is an example where we are using spark connectors into the data that is held in a blockchain environment This happens to be a supply chain blockchain So we are looking at Supply chain data and trying to understand or predict delays in shipments So the shipment delays could happen because of a lot of different things related to supply chain itself It could be related to the things that are being shipped. It could be the carrier. It could be a destination and Source it could be time of shipment. It could be External things like weather so using machine learning in the context of this data Made possible through spark. So spark allows you to connect to the underlying data sources bring it into DSX data science Experience understand factors contributing to the delays and shipments and then build an ML model That can be used to predict whether there will be a delay as well as what the delay will be right interesting use cases evolving and finally one word about Watson so Watson is set of a cognitive functions in terms of speech recognition language understanding natural language Image recognition and so on. So what do they to have in common? We see a we see many customer use cases where the friend and interaction uses one or more of Watson's cloud-based services. So for example Call center records so speech to text happening at the call center Which are then used to extract pieces of information build a model Build a build a set of features which are then fed into the model to do a scoring real time and then guide the call center discussion Right, so speech to text in the front end using Watson coupled with a spark ML model in the back end coming together so the concept of Part for machine learning processing on spark coupled with pre-built deep learning services on Watson Make some very interesting cognitive applications. I want to end there. My time has run out Happy to discuss offline I would highly encourage you to take a look at some of these things machine learning and data science URLs are here We feel free to to play around. Thank you for your time