AIOps in Networking

0:00 [Music]

0:10 thank you

0:12 hey good afternoon everyone it is great

0:14 to be back in Las Vegas and if we had

0:16 been here 40 years ago when I started my

0:19 career as a wireless networking engineer

0:20 and we had bet one dollar that would end

0:23 my career as a chief AI officer we'd all

0:26 be billionaires going to Tahiti instead

0:29 of here

0:30 Zoe but besides beating the odds you

0:33 know what it really highlights is how

0:35 fast this industry is changing and how

0:38 important AI is becoming

0:41 to the industry you know you look over

0:43 our careers how fast things have changed

0:45 that I'm now a chief a officer at

0:47 Juniper

0:48 uh you know for me personally

0:50 the missed AI Adventure started around

0:53 2011 when I was at Cisco

0:56 that's when I saw Watson play Jeopardy

0:59 right and that's when we saw Watson beat

1:02 those Jeopardy championships and that's

1:04 when I realized you know if they could

1:06 build something that can play Jeopardy

1:09 we should be able to build something

1:11 that plays networking Jeopardy

1:14 right and that's when the adventure

1:16 started and you look at what we've seen

1:19 AI touch all aspects of society right AI

1:23 is going to be on par with the

1:25 Industrial Revolution the Internet it's

1:27 going to touch all aspects of things

1:29 that we work with you know if you look

1:31 what's happening in healthcare AI is

1:33 becoming a necessity for helping

1:35 diagnose cancer right you're going to

1:38 want your doctor to be using AI when

1:40 you're in there visiting him if you look

1:42 what's happening in our farming industry

1:44 right they're building AI now they can

1:47 identify each weed and just apply the

1:49 right pesticide to each weed you know

1:52 this is going to like double or triple

1:54 the production of our agriculture

1:56 industry you know and we look at what

1:58 happening with cars you know I would

2:01 also predict In Our Lifetime it will be

2:03 illegal to drive a car

2:06 you know these autonomous vehicles will

2:07 get so safe

2:09 you know that driving will be outlawed

2:12 and finally you know we also chat gbt

2:14 you know we're starting to see this

2:17 technology transfer become even more

2:19 transformational I personally like to

2:21 thank I send openai Christmas card

2:23 you know they've reduced the number of

2:25 skept cases when I started this five

2:26 years ago

2:27 I would tell you over half the room

2:29 would be AI Skeptics that has gone

2:31 significantly down

2:34 you know in the networking space

2:37 you know when sujay and I left

2:40 Cisco in 2014

2:42 . what we realized was it was going to

2:45 require a new

2:46 architecture to build Cloud AI it was

2:49 going to require a blank sheet of paper

2:52 when we left Cisco to start this

2:54 adventure

2:56 right and when we started the adventure

2:58 we started the adventure with the

2:59 wireless access point it's not because

3:01 we thought the industry needed another

3:03 access point uh be honest most of you

3:05 guys thought we were crazy Tom was one

3:08 of them he's like why are you guys

3:09 starting another wireless access point

3:10 but they didn't realize we really had a

3:13 different Vision the vision was

3:15 basically to solve this Paradigm Shift

3:17 going from managing Network elements to

3:21 one of managing the cloud we still have

3:23 to creep all those Network elements

3:24 green right we still got to make sure

3:27 I'm green but the rule paradigm shift

3:29 and we heard this from you when we were

3:31 there at Cisco was customers were tired

3:34 of controllers from crashing right

3:36 before they were going to put any

3:37 business critical app on that Network

3:39 they wanted those controllers to stop

3:41 crashing they wanted them to transform

3:44 faster right they didn't want to have to

3:46 wait a year for an update you wanted to

3:48 be able to keep up their digital

3:49 transformation they want things to

3:50 happen in the order of weeks but most

3:52 importantly they wanted to make sure

3:54 that that user was going to have a great

3:56 experience when they put that App or

3:58 that application on the network

4:01 and we started that Journey with the

4:03 wireless piece because when we were

4:05 trying to answer the question of why are

4:07 you having a poor internet experience

4:09 the access layer has about 80 percent of

4:12 the information you need to answer that

4:13 question

4:14 and we built those access points to

4:16 basically move Telemetry back to the

4:19 cloud

4:20 and so that's where we started the

4:22 adventure since we've joined Juniper now

4:24 you've seen us basically extend Marvis

4:27 and AI Ops across the wireless AP to the

4:31 switch and to the route you know the AP

4:35 brought all that connectivity

4:36 information

4:37 the router that sd-wan router brings all

4:40 the applications starts to help us to

4:42 answer questions about the applications

4:44 and what we just announced last week a

4:47 week or two ago a Mobility field day is

4:49 something called continuous learning

4:50 it's you know when you're in the data

4:53 science business

4:54 you talk to data scientists label data

4:57 as gold

4:58 it is very hard to get good label data

5:02 you know what we're doing with these

5:03 Cloud applications like zoom and teams

5:06 now is basically getting labeled data

5:08 from that user experience application

5:10 they know when you're having a bad

5:11 experience

5:12 you know and with that label data now we

5:15 can join it with network feature data

5:17 and actually build models that can

5:19 accurately predict your Zoom team's

5:22 performance

5:23 and once you can do that

5:25 you can now start to explain why you're

5:27 having bad teams so that is where AI

5:31 deep learning is starting to transfer

5:32 transform what we're doing in the

5:34 networking space

5:38 now as I said we started the journey

5:40 with the data you know if you look at

5:42 the access point

5:43 that I built 20 years ago at airspace

5:46 that access point was sending data back

5:49 to the controller every minute every two

5:51 minutes

5:52 it was sending you back synchronous you

5:55 know the reason we built the access

5:56 point was because I really did not trust

5:58 Cisco in Aruba they give me the data I

6:02 needed to answer that question

6:04 so we look at the data side that is

6:07 where we're starting to get user state

6:09 in addition to that synchronous data

6:11 we're starting to get asynchronous data

6:13 right every time that user State changes

6:16 I want that access point to send me back

6:18 to State I want to know that user State

6:20 when it changes

6:22 and the next big part of this puzzle was

6:24 really what we call these AI driven

6:26 perimeters

6:27 was what we found out

6:30 promote from you

6:31 AI aside was you wanted the cloud just

6:35 getting the data to the cloud

6:37 and that took us about a year you know

6:40 Randy got the data built you know got

6:43 the network built and then we had to

6:44 spend a year trying to figure out why

6:46 the support team was still sshing into

6:48 radios to solve problems

6:51 right

6:52 you because when you do this Cloud a

6:54 you've got to get that data back to

6:55 compute storage you gotta bring the data

6:58 to where the compute storage is where

6:59 you have unlimited compute storage

7:01 and the good thing about AWS and Google

7:03 we have unlimited compute storage

7:05 they only have to worry about is my bill

7:07 I get a bill from those guys every month

7:09 and I still that still limits you to

7:11 some extent but once you're there

7:12 there's no limit on what you can do with

7:14 the data and you guys appreciate that

7:16 because once you get the data there we

7:18 all have access to the same data

7:20 everyone in the team has access to it

7:22 the next big piece of the puzzle was

7:24 really around the data science

7:26 and it's really not around the

7:27 algorithms this is around the team

7:30 right and the big thing we did inside of

7:32 a Miss is really get our data science

7:35 team next to our customer support team

7:39 because that support team represents our

7:41 customer

7:42 and then the fourth thing the thing I'm

7:44 most proud of this is something uh

7:46 four or five years ago when I started

7:48 the whole conversational interface for

7:49 severe you know there was a big argument

7:51 whether or not we really need

7:53 conversational interfaces

7:54 you know and what I firmly believed was

7:56 in the industry we all started our

7:58 careers with CLI we're all CLI jockeys

8:00 20 years ago very proud of it and then

8:03 we slowly went to dashboards and it

8:05 makes life a little bit simpler the next

8:07 big transmission transition in the user

8:09 interface is going to be around these uh

8:12 these llms in CI you know you're going

8:15 to be talking to your network

8:17 like we see on Star Trek that is going

8:20 to be the next big trans is going to

8:21 make life easier for RIT to actually

8:23 start troubleshooting networks and

8:25 finally Studio's favorite is actions at

8:27 the end of the day

8:28 ah I don't care I don't need to know why

8:31 there's a problem I just need you to fix

8:33 it if you know why there's a problem

8:34 just go ahead and fix it and not bother

8:36 me anymore

8:40 and then finally if you look at the key

8:42 components of Marvis

8:44 you know we started with

8:47 the VNA that was our conversational

8:49 interface that's where we started the

8:51 journey four years ago you saw that with

8:54 nlu

8:56 you know what you're going to see us

8:57 start doing is start basically

8:59 integrating llm into that conversational

9:01 interface we've always had the natural

9:04 language understanding piece of the

9:05 puzzle the next piece of puzzle is to

9:08 start to generate answers that sounds

9:10 like a human and that's what llm is

9:12 trying to bring to the party is that

9:14 natural language generation part that

9:16 starts to bring your network to life

9:18 the action framework is a self-driving

9:20 piece this is the piece where you want

9:23 your network to start taking actions and

9:25 this is really going to become on an

9:27 element of trust

9:29 right this is the piece of the puzzle

9:30 where

9:31 you have to we have to earn your trust

9:34 right we have to earn your trust as an

9:36 AI assistant and I always tell people I

9:39 don't care if you're a virtual assistant

9:40 or a real person most I.T people are not

9:43 going to let you touch the knobs on

9:44 their Network

9:46 until you've earned their trust somehow

9:48 they've got you've got to make sure that

9:49 you trust that virtual assistant to

9:52 actually twist the knobs or make it

9:53 change your network and the third piece

9:55 that we're adding and this is a new

9:56 fundamental Lake in the marvelous

9:58 adventure is this continuous learning

10:00 feedback

10:02 this is where we're starting to bring

10:03 label data back from your Cloud apps

10:10 now

10:12 people ask me to AI ml for me AI is not

10:16 an algorithm AI is a concept

10:19 right and it's really the next step in

10:22 the evolution of automation

10:24 you know what we're doing with AI is

10:26 really Building Solutions that can do

10:29 tasks on par with human domain experts

10:33 right we're building solutions that

10:34 typically require cognitive skills you

10:38 know we've all built automation of

10:39 basically automate some deployment

10:43 configuration those are very basic

10:45 scripts the next step is really doing

10:48 something that typically would require a

10:50 human to do

10:52 and if you look what's going on under

10:53 the hood of Marvis

10:55 I usually break it into two sets of

10:57 algorithms you know we have kind of

10:59 these regression ml this is what we've

11:01 been doing for decades they've been

11:02 around for decades you know what really

11:04 changed in 2014 was the addition of

11:08 these deep learning networks you know

11:10 and interestingly if you look at Google

11:14 you know you look at the Google search

11:16 statistics

11:17 on AI ml it is in 2014

11:22 when AI went from Mostly a research

11:25 topic

11:27 to a reality 2014 is when we saw kind of

11:30 that perfect storm of compute storage

11:33 low-cost cloud storage again good we saw

11:36 tensorflow and all the tool sets needed

11:38 to build these things become real

11:40 you know that was the year when things

11:43 started to come together for AI

11:45 and these deep learning algorithms these

11:47 are the ones they're disrupting the

11:50 industries

11:51 chat DBT that was Transformer right

11:54 that's all the Transformer technology

11:55 that came out in 2017 from Google

11:58 lstm or the Deep learning algorithms

12:01 that are allowing us to do anomaly

12:02 detection now with very low false

12:04 positives kind of a very basic

12:06 networking capability we all talk about

12:08 that I've never really achieved because

12:10 there's always too much noise coming out

12:13 of those things waking you up at three

12:14 o'clock in the morning

12:15 and then what we're talking about now is

12:17 shappy for those who have not heard

12:19 about shapley this is basically a

12:21 technique that lets you make these AI

12:24 models explainable like you understand

12:25 what features are really driving a

12:28 prediction

12:35 and I think the other quick thing you're

12:37 going to hear Randy come up talk a

12:38 little bit about this more

12:39 was in addition to having to build a

12:43 whole new Cloud infrastructure

12:45 right that had the microservices

12:48 architecture had the pipelines when we

12:52 wanted to build real-time data

12:54 processing

12:55 and Susie will quietly reminded me Bob

12:58 when we started this we really weren't

13:00 talking about AI

13:02 back then we were really talking about

13:03 trying to build an architecture that

13:06 could do day two operations and process

13:08 data in real time

13:10 now it turns out that is the foundation

13:12 you need if you want to do AI you have

13:14 to build Cloud infrastructure they can

13:16 process data in real time

13:18 but more importantly

13:19 you know

13:21 the need for the blank sheet of paper

13:22 was around organization

13:25 right in addition to having to get all

13:27 this technical stuff done and built

13:30 we had to basically get the domain the

13:32 data science team

13:33 tied to the

13:36 support team those are your domain

13:38 experts

13:39 you know and if you look at our Cloud

13:42 support team what you notice is the Miss

13:45 support team is represents you

13:49 right we all have access to the same

13:51 data Cloud the customer the support team

13:54 the data science team you know and the

13:57 data science team's Mission and life is

14:00 to keep the support team happy

14:02 the fewer tickets that that support team

14:04 sees

14:05 is a few tickets that you guys are

14:07 sending our way

14:09 right and that is the key and I will

14:11 tell people that you know if you talk to

14:13 an infrastructure vendor if their

14:15 support team is not using their own AI

14:18 Ops tool they have not started the

14:20 journey

14:21 because this journey starts with making

14:23 sure your data science and your domain

14:25 experts are tied to the hip and that's

14:27 another thing that's very hard to do in

14:29 a large organization like Cisco and

14:31 Aruba it's hard to organize actually

14:32 change those type of things and I always

14:34 tell people

14:35 if you want to look for architecturals

14:37 changed in an industry

14:39 if you see Cisco try to get three

14:42 different bus to work together

14:44 you know something's up that is a sign

14:46 that something's changing

14:48 so keep an eye out for that

14:51 now this is what I've been talking about

14:53 I'll go a little bit deeper on this if

14:55 you look what's fundamentally Happening

14:57 Here

14:58 we're not getting label data

15:00 from your zoom and teams we're taking

15:04 that label data we're joining it to the

15:07 network features

15:09 and now we're building models that can

15:11 predict the audio latency the video

15:14 latency and really the performance

15:18 of your Zoom teams experience

15:22 now once you have a model that can

15:24 actually predict something accurately

15:26 the magical thing called the shapley

15:28 is now we can explain to you what

15:31 feature

15:32 is causing that performance so when you

15:35 have a latency if the average latency is

15:37 100 and your CEO has an average latency

15:41 of 150

15:43 we can now explain you know is the

15:45 wireless is it the client or is the WAN

15:48 who actually contributed to that latency

15:51 problem

15:52 so that's the the power of these deep

15:54 learning models that are going to be

15:56 transforming how we solve and

15:58 troubleshoot and manage networks going

15:59 forward

16:01 the other big piece of this is really

16:03 around natural language what we've done

16:04 to date you know in your current Marvis

16:06 interface

16:07 is we have Ross and nlu there so we do a

16:10 breaker job of understanding your intent

16:12 but we've never done a really good job

16:14 of basically giving you an answer that

16:16 sounds like a human wrote it you know

16:18 and that's going to be our first step on

16:20 the journey to L and we're going to

16:21 integrate lym into Marvis starting with

16:24 knowledge-based questions

16:26 now our goal is to get Marvis to the

16:28 point that it can actually pass

16:30 a networking certification test so

16:33 you'll basically have a network

16:34 certified engineer by your side helping

16:37 you troubleshoot things

16:40 the next step in this adventure

16:43 is to put this on top of your database

16:47 you know where we actually translate

16:48 text SQL

16:51 so you can now start talking to your

16:53 network like you would talk to a Star

16:55 Trek computer you know so this is my

16:57 other thing you know we all all the star

16:59 Tommy Star Trek fans that we have in the

17:01 crowd

17:01 you know okay talking computers that

17:05 technology is about to become real

17:07 we are down to the teleporter the only

17:09 thing left will be getting this tell a

17:11 problem solving talking computer is

17:13 within our lifetime within my lifetime

17:15 maybe let me shortly shortly here

17:21 okay with that I'm all invite said

17:23 you're back up

17:24 partner crime is going to take you

17:26 through reality and get us off some

17:27 PowerPoint awesome thank you Bob thank

17:29 you very much

17:32[Music]