Dinesh Nirmal
Vice President, Data and AI Product Development; Site Executive, IBM Silicon Valley Lab at IBM
  • Video
  • Table of contents
  • Video
AI Summit San Francisco 2019
September 26 2019, San Francisco, CA, United States
AI Summit San Francisco 2019
Video
New AI Frontiers Data, AR, and Quantum - IBM
Available
In cart
Free
Free
Free
Free
Free
Free
Add to favorites
250
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About speaker

Dinesh Nirmal
Vice President, Data and AI Product Development; Site Executive, IBM Silicon Valley Lab at IBM

Dinesh Nirmal is Vice President of Development for IBM Data + AI, with global responsibility for building the analytics, data, and AI technology in use by enterprises from governments and international finance firms to startups. Dinesh leads more than a dozen IBM Development Labs around the world. Major releases during his 2-year tenure include IBM Cloud Private for Data—an open, cloud-native information architecture for AI; IBM Integrated Analytics System; Machine Learning for z/OS; and IBM Watson Studio. Products in his portfolio have won major design awards including two Red Dot Awards and the iF Design Award. Additionally, he is the Site Executive for IBM's Silicon Valley Lab.

View the profile

About the talk

Share

Today, I want to call the next 25 minutes or so on AI in Enterprises. How are Enterprises adopting a I so he look at 00:04 it. There are three Technologies. That's really driving. The Enterprises today want is a I we all know we all hear about AI 00:14 the second one is cloud. What is quantum so I'm also going to have IBM research power Quantum with me on how AI is being implemented 00:24 with quantum. But with an AI there are three friends that developing. And those three times around data 00:33

algorithms. And the final one is around computer. So what do I mean by compute the more data the proliferation 00:43 the velocity the veracity of data happens, you need more computing power the traditional ways of computing doesn't work. So that's where Quantum 00:53 comes in. So we are going to cover that algorithms as you develop and deploy this algorithm supermodels into Enterprises. 01:03 You want to make sure there's no buys there. It's fair. It's explainable and the final one is data. 01:13

So there's a perception that you know all data in Enterprise exists in the CSV file. But it's not the case data is siloed 01:22 data is fragmented. There's no way to get access to trusted data. I mean I have worked with customers were you know, they talked about 01:32 self-service analytics but it takes 6 to 8 weeks for a data scientist to get faster data. What kind of account login out for the meditate exist within 01:41 the Enterprise's how is that sinking between the existing card lost data Galaxy that is out there to the challenges with data. And so I want to invite 01:51

I had a pleasure of the last two days working with one of our customers from in a Green Bay Associated Bank. So I also have the pleasure to invite him 02:01 on stage to talk about from the daily challenges that Steve lyric who is the SVP senior vice president and director for data management. So, thank you 02:11 Steve. So I mean, obviously you are a practitioner that you know, I'm more of a Preacher so your practitioner, but 02:19 we talked about data all the time. But even before we get today. I mean especially for the audience what does a I mean to Enterprise, I mean there's a 02:29

lot of talk about AI, you know, every Enterprise wants to adopt a are but what is a I mean to an Enterprise Turk from our perspective obviously AI is 02:37 unleashing new capabilities for our organization unleashing new business insights from customer knowledge and insight all the way through to our 02:45 systems are are functioning operating and really what we're trying to do is facilitate Innovation across the organization, but I think when I look at 02:53 it from an Enterprise perspective and point of view, it's really about how do we scale and enable and empower the right users across the organization 03:01

to be able to leverage and perform analytics across the data sets that are now being able to provide which is new day of sets new variables that never 03:07 before could have been touched. We have some scenarios pop up recently around Areas of our our industry are of our 03:17 organization are not traditionally did scientists or even statistical modelers. We have different we're finding that by enabling the users around the 03:26 bank. There's new opportunities where we had a model that was built for a research Department within our team that they've built completely 03:36

independently in a silo not on Enterprise infrastructure and wanted to we now hire trying to provide a platform for them to enable that those use 03:45 cases because we think that we're seeing a gray area between roles and responsibilities across the organization of who is actually performing 03:55 analytics data engineer who used to wear that heart is now becoming a data scientist and the same thing because 04:02 those those are kind of like marching to some extent but my data perspective like what challenges how many know I see the data silos and lot of copies 04:12

of day. A multitude of data existing, you know within and the price I mean what challenges do you see with the day that no, it's it's exactly that we 04:22 have data availability is still like you mentioned in the opening this still a big problem. I meant for Austin and we want to try to provide a 04:30 platform that's scalable and can introduce consistency and governance and control across the Enterprise but still be fast in the address. I won them 04:38 both. I'm one of the risks we have is that enterprises off and looked at it as being slow, right? And anytime you talk Enterprise. That's that's going 04:46

to take too long. So we're looking at ways and Frameworks and platforms to help us and just that they have rapidly and have that government controlled 04:53 all the way through the analytics and end in able to users through that are slow to adopt 05:00 a i there's a huge challenge in bringing a i into Enterprises. It's not that easy in traditional because you're infusing aai into that a technology 05:10 that has existed for decades third-party software. Challenges exist but finally wants you to know you don't have the model you deploy tomorrow, but 05:19

then comes in that industry like yours would banking the so much regulatory oversight. How do you do it? Like, you know, how do you explain ability 05:28 and other things that you need to do to run again that that's one of the keys of the platform in the scale that we bring from Enterprise perspective 05:35 is the the explainability of the model is just as important as model itself in a lot of cases from a regulatory and compliance perspective. So making 05:42 sure that we understand and have the process in place to know where that data came from through the entire life cycle of that data and we can explain 05:51

that on the way out. There's a perception kind of in the space, especially in the banking and then the Legacy industry is is around the fact that the 05:57 black box right? When we introduce a I were going to lose divisibility rule lose that Insight we want to make sure that we have a platform that 06:05 enables us to have that explained ability so that we can dig right in and find out all the details right to think that thing Steve mentioned one 06:11 problem. I think it's very critical for any Enterprise that wants to embark on a journey of a i it's very critical to have a platform because all the 06:19

way from ingest to relaxation if you have a set of services that's already stitched together. Becomes a lot more easier than you bringing a government 06:28 service or a data ingestion service and the visualization service in trying to stitch it together the next point that you mention Black Box. I mean, 06:37 that's also very true. So let me that takes me into that my next small snippet of a video that I want to show you but let me thank Steve for us. Thank 06:44 you Steve so much for coming on stage. I know you don't reply back, but thanks for coming back and sharing with us your experiences. 06:52

so so here I want to show you in a steep talked about the about the black box, but then we 07:04 talked about interaction with data. We have been tracking the data for a while right at me through all your computers obviously opening tracking but 07:13 now it's changing to talk to you later. I mean, if you don't think about chatbots you're talkin to the data and the date is kind of talking back to 07:21 you same thing with home assistant, you get some level of talking going back and forth. What is working do some 07:28

research on what if you walk with your data or what if you walk into your data meaning it becomes a lot more easier if you can 07:37 visualize data in a 3D perspective or if I'm a bunch of reality perspective so I can show you like a, you know, a minute video of what we have done 07:47 for explainability become so much more easier when you can visualize from a 3D perspective take a look at it. Our virtual reality environment allows 07:55 you to step into your mother like never before we can see our machine learning model waiting to be explored when we crack open the Black Box we can if 08:05

we see all 25 features selected by the model. The ones closest to us have largest contribution to the model of decision those further away. Could you 08:15 read the least? Pictures in an AI model don't operate in a vacuum the interaction between features play a huge role in the 08:25 the better. We can make sense of our mom. Let's selected features 08:33 unemployment rate and interest rate. This launch is a 3D scatter plot. We can see how multiple 08:41 features added together impact model output. Analyzing a 3D scatter plot helps us see the data from different perspectives. Let's focus on the 08:51

bottom and view it from the perspective of interest rate. You might expect people to default on their loans when the interest rate Rises, but that's 09:01 not the case that such as long as the unemployment rate remains low Rising interest rates actually make people less likely to default 09:10 insights begin jumping out 09:19 Explorer bring us one step closer to cracking open the black box and understanding how are 09:27 a I was making prediction. So you see 09:37 how you know, I talked about walking with your data walking in today that where you can really visualize from a 3D perspective. Next. Especially in 09:47

the industry is like what Steve talked about explainability becomes very critical and this really helps the second friend that we see is algorithms. 09:55 So I called you because once you deploy tomorrow people think that that's it, you know, you're deployed tomorrow, but that's where the real work 10:04 starts. I think in an Enterprise because think about it once you deploy tomorrow you want to make sure the model is not drifting. Obviously. The model 10:11 is fair. The model is not biased all those things is part of the algorithms and in the next video clip in I want to show you where we work with 10:21

Wimbledon and you will see that bias exists in the most unexpected places and you and you see the video you will never see bars in there, but I'll 10:29 point out to you so they can look at the video and then let me take you through it where buys exist. Technology is changing the way we've used for 10:39 Wimbledon this technology a groundbreaking best 10:48 highlights from multiple matches a captured simultaneous though. We both a 18 quarts in up to full max 10:57 amount of time to comply with 11:04

these two minute to the match 11:12 finishing. Why won't you play reactions listening to crowd excitement levels and analyzing the gameplay statistics IBM Watson enables 11:22 Wimbledon to deliver unmissable minors without delay. 11:32 Can you go back? So let me run through the high-level architecture of that and then if you could take a guess where the buy succeeded have a nice 11:44 time, you know, it's really hard. But what we did was that obviously there's a tremendous amount of unstructured and structured data that comes in. 11:53

I'll be created the features for example prayer player ranking venues crowd cheering crowd noise level all those things and be built a Model A Body 12:00 realizes that every time the model is picking the higher-ranked well-known players for highlights. So think about it now is always Federal 12:09 Nadal Venus Williams or maybe a nice is not talk anymore. But Serena, you know, so it always picks the top-ranked fish. So once we run through the 12:19 bikes detection software off, you know, what's an open scale be able to bring up lesser-known players because if you think about it, the features that 12:28

we were using were crouched your Venue, I mean, obviously the top-ranked player Get the center code they get larger crowd there, you tend to be more 12:38 louder. So the model continuously picked, you know, the the top ranked players, but once we ran to the Watson open scaling buys detection, we were 12:47 able to bring lesser-known play reflection Toro who had really good games that day and some really nice shot, but they were able we were able 12:56 to bring them to the top but not lie. You know, it gives an opportunity for lesser-known players to be advertised or being brought the limelites of 13:06

hit help from two perspectives, right? So that's what we did it by Static and then that's why I said sometimes buys exist in the most unexpected 13:16 places. Nobody would have thought okay or alright, there's the Federalist highlight, but it will never happen in your mind that there is bias exists. 13:25 And that's what I call it subconscious bias or mute advised that is hard to Inner detect. So what is it take? What does it 13:34 take for us to build a model that is in a press 40, right? I think there's four pillars to it one is explainable. How do you make 13:44

sure that it's explainable? So especially when you have a bank for example, you know, you want to make sure the model is explainable. If you give a 13:54 loan it's for sure the regulator will come and ask. Why did you deny This Town Fair? How do you make sure the model is fair? And there's no bias in 14:01 their tree accurate. I mean you want to make sure you want to detect model drift even before it starts happening. Right? So it's very 14:10 critical. I have some concrete examples of a obviously given the time I cannot go into it. But Martin is real, you know as the day that changes model 14:20

800 the last thing is open. I honestly truly believe that this is the biggest undertaking is the biggest undertaking of a 14:28 lifetime. If you don't get this, right, there's going to be large consequences. We got to make sure that these organisms are not held by one vendor or 14:38 one company or one Enterprises, but it's out in the open. A community is behind it to make sure that the model that be billed the training data sets 14:47 be used is not creating biased and unfair models. So it's very critical that we make it as open as possible. So if you take what's an open skill for 14:56

example, and maybe you know do what we are preaching we have contributed tremendous amount of Technology into it. So if you take for example explain 15:05 ability to take for fairness 360 robustness to get all of them are available on get a beer contributed back into the community and we are we 15:13 are more and more focused in making this open so that Bias and fairness in Odessa and exist unfairness doesn't exist in the 15:23 models going forward. The last piece is computer and for that I want to invite Abraham who's in IBM researcher in the quantum side and working with 15:33

AI and Quantum to see how do we ask the data proliferation happens? How do we make sure that we bring more computing power into it so far. What else? 15:43 Can you turn then the cubits in Quantum so far. Let me invite Abraham to come and give you a high level of you on what we're doing from IBM research. 15:52 So Abraham, thank you. Tell everyone let's begin. So 16:00 today I would like to talk to you a little bit about Quantum Computing and Ai and how those two work together. So let's begin first with Justice 16:09

opening page. How many of you have seen this picture of a golden chandelier looking thing in the past see some had so Dai Dai here is the 16:19 Quantum devices that at least IBM is working on are made out of superconductors. And so those superconductors have to be cooled down to low 16:29 temperature in order to work with those devices. So we have something here that's called a dilution refrigerator and you see several lines bringing in 16:36 microwaves to your device because microwaves are how you communicate with the device at the end of the day the chips. It's at the very bottom and we 16:44

have two bits instead of classical bits to work with. So now the question becomes how do we use these cubits? And how are they relevant to Industry 16:52 applications? And in this case, I'll show you a quickie. sample of using Quantum Computing in the field of AI So let's start very simply by 17:02 introducing the field. So the history of the field goes as far back as the early days of quantum mechanics. So here I put the earliest days of 1935 17:11 when people were trying to understand the implications of quantum mechanics as a as a field of physics very quickly. The field has come has come to a 17:21

point where we're now focused on application. So how do we extract application advantages out of quantum computer so you can see as early as 17:30 1984 IBM has been working on this field and proposing algorithms for used in cryptography. For example, the field 17:39 accelerates very quickly. And here we are in 2016. IBM was the first company to put a quantum computer on the cloud which means the field have gone 17:49 from something that's in a research lab to something that's out in the open. So anyone can access a quantum computer down. And since that time some 17:58

exciting announcements have come more recently IBM announced a new data center in New York and our largest. Quantum device yet with 53 18:07 cubits. Okay, so let's let's start with some basic. So what what are we dealing with? So let's first talk about the computers that we have 18:16 today versus quantum computer. So on the left, I put two coins here with heads and tails to indicate that we have two options. So 0 and 1 and 18:26 our computers today work with this level of information out of your transistors today. You can build and Gates or Gates not Gates and so on and build 18:36

logic based on those Gates so that the end of the day your logic is Computing on zeros and ones flipping zeros to ones who wants the zeros and vice 18:45 versa. On the other hand on the right side. I'm trying to show you a representation of the state of a cubit which is a Quantum bit here instead of 18:53 having just two states up and down. You have multiple States along the surface of that sphere in the picture. So really you can have a cubit that's 19:03 in any state along the surface of the sphere and what this tells use that the feature space that's available to you is fairly large. So this is one of 19:13

the two principles that I would like to tell you about to take advantage of quantum computers to in a real feels like artificial intelligence. So this 19:20 is one property rights to the state of a cubit can be anywhere along the surface of the sphere another property that becomes important is that first 19:30 one is called superposition. You might've heard about this the second property that becomes important. Is that 2 cubits that come together can 19:39 interact in such a way that once you have them a parts their properties are correlated. So if I flip two coins right now the outcome The two coins 19:47

doesn't necessarily need to be correlated, but with Quantum bits that's not true. The outcomes can be correlated. And so what this does is give you a 19:56 much larger features face to play with. Okay. So without going into too much detail one point that I really want to 20:05 get across the days that you can go to Quantum Computing. Ibm.com where you'll be met with a web page that looks like this and you can write code on a 20:15 quantum computer and executed yourself. So you'll be met with a page that looks like this where you can build a circuit graphically. So sing 20:24

the the old days of computing where you needed to build circuits with and Gates or Gates not Gates a similar concept so you can use a graphical 20:33 interface to build your circuits or you can write code that's very familiar to most of you here. I'm sure you can write code in Python to then program 20:42 a quantum computer and get measurements outcomes from that quantum computer. So, how did those two screens look here's a screenshot of the circuit 20:51 composer those five lines that you see here horizontal lines are individual cubits and you can see I've dragged and dropped among that series of gates 20:59

that are available few select options on those cubed soap. Generally. This is how you'd build a Quantum Circuit by dragging and dropping using the 21:08 circuit composer on that page. There's also an interface for you to write coat. So this code looks very simple right? I'm saying circuit. 21:16 XO applying xgate on Cuban 0 and you can see that very first XK8 on q80 so you can you can build your Quantum circuits very 21:26 simply right writing programs in Python like this. Okay. So now that you know that you can go ahead and build a 21:36

Quantum algorithm whether it's graphically or encode the question becomes, how do we do use Phalanx? And in this case? I want to show you how easy it 21:45 is to build a Quantum svm kernel classifier. So here that the challenge is going to be I have a data set that looks something like this and I'm going 21:55 to try to come up with a way to classify that data into the several pieces of classification. So there are two challenges that come up here, right? 22:04 The first question is how do you represent data into the quantum computer for it for it to then do the number-crunching for you? The second part that 22:13

comes up with that you come across is how do you then compute on the data once you've been quoted it? So these details have have all been written out 22:22 for you and kiss kids. So kiss kid stands for Quantum information science kit. This is an open source Quantum Computing software and in kids get there 22:32 are many layers for you to get involved at the application Level. So what I want to show you is how easy it is to build a Quantum algorithm to do this 22:42 kind of classification and kiss kit. So take a look at this piece of code. So this is all you need to get the code to classify that data. So 22:51

what I'm doing is first saying the algorithm is a q svm. So that's a Quantum svm. Kernel I'm doing a feature map, which I've described 23:00 here. So for the details of this feature map week and we can talk offline and I'm doing a multi-class extension. So because this is data with multiple 23:10 classes instead of just two classes. I can extend traditional classification models using multi-class extension and I'm saying I would like to do 23:19 1,000 repetitions of this experiment and see what a quantum computer says and it's that simple so you can build your Quantum algorithm 23:27

programmatically like this just describing the pieces that you need. I want a Quantum svm. Kernel these this is how I would like the information to be 23:35 mad and I'd like to work with this algorithm. So we have for the algorithm the input is the training data and we also test with the test date. 23:43 And that's it. You take that bit of information that you built to describe the algorithm and you just say run it and the quantum computer gives you 23:53 the results from this. So if you run it at least in this particular example, I'm able to classify this data with 80% accuracy. And as you 24:01

traditionally expect we can get the predictions out. We can get the the results and compare what our outcomes were with what we would expect to see. 24:10 So what one thing I'd like to point out here is there there was a paper published by IBM research and co-workers in a nature earlier this 24:19 year showing exactly this kind of concept to supervised learning with Quantum and has feature spaces. So taking advantage of the large features space 24:29 that's available to you in quantum mechanics to do work in terms of classifying data sought out strongly take out strongly suggest that you take a 24:38

look at this paper. end 11.5 like to leave you today is that you can still go to this website start 24:47 programming your own Quantum algorithms start learning how this field can be leveraged in your application and we have made several tools to make sure 24:57 that your you have an easy time doing this. So one of those is we've put out a learn Quantum Computing using kiss kit textbook. So that idea here is 25:06 that you can go through a textbook that's open source and available online and start learning how to program these quantum computers. We also know 25:16

that people like to learn from watching video. So we also have videos available for you online. You can take advantage of these videos and learn how 25:25 to install kids get on your computer and run code from it. And because this is open source software. You can also learn how to contribute to the open 25:33 source project. Okay, thank you all 1 one final thing. I'd like to say so there's a session today at 1:30. Where will go into a deep dive on 25:41 the kind of code that you saw earlier. I'll show you how to build Quantum circuits how things work and we'll build circuits together at that session. 25:51

So 1:30 p.m. Deep dive. I'll call Dinesh back to the stage now. Thank you. 25:58 Thank you everyone. I mean you saw what we have done from a quantum computer perspective from a data perspective from an older than perspective. So, 26:06 please, you know reach out if you have any questions we have also a boot on 516. But thank you so much for your time preciate that 26:15

Cackle comments for the website

Buy this talk

Access to the talk “New AI Frontiers Data, AR, and Quantum - IBM”
Available
In cart
Free
Free
Free
Free
Free
Free

Video

Get access to all videos “AI Summit San Francisco 2019”
Available
In cart
Free
Free
Free
Free
Free
Free
Ticket

Buy this video

Video

Access to the talk “New AI Frontiers Data, AR, and Quantum - IBM”
Available
In cart
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
512 conferences
19971 speakers
7276 hours of content