Bob Friday Talks: From protein folding to teaching Marvis about Zoom

0:01 hello I'm Bob Friday and thanks for

0:03 joining us on another episode of Bob

0:05 Friday talks I am joined today by former

0:08 professor of structure biology at Lon

0:10 University and now director of data

0:12 science here at Juniper Miss naaj panu

0:16 today we're going to be discussing his

0:17 journey from professorship to network

0:20 data scientists welcome naaj you know

0:23 maybe we can start our conversation

0:24 today a little bit about how did you you

0:26 know what got you in be a professor I

0:30 think what got me into being a professor

0:32 but also the transition to from

0:35 structural biology to networking is my

0:37 love of mathematics that's a common link

0:40 I love Applied Mathematics I love

0:42 solving problems with mathematics

0:44 testing out the problems by writing

0:47 software and then validating it okay now

0:50 now structured biology you know exactly

0:52 what does that mean so you were working

0:54 on so what we were trying to do was

0:59 developing algorithms to solve the

1:01 three-dimensional Atomic coordinates of

1:04 protein or DNA molecules so why is that

1:07 important essentially they are the

1:10 building blocks of life with that

1:12 knowledge of the three-dimensional

1:14 structure we can understand how the

1:16 protein functions and the basis for

1:18 diseases so many diseases are caused by

1:21 mutations in that protein with that

1:23 structure we can correlate to the

1:25 function and then develop drugs to

1:28 combat disease Okay so and I've read a

1:30 couple of your papers it sounds like you

1:32 know you know before joining us you're

1:33 working on solving the problem of

1:35 immortality aging you know I almost

1:38 regret hiring you you know well know so

1:42 tell me exactly how the proteins aging

1:44 all relate to each other so at the

1:47 fundamental level aging or cancer is

1:51 caused by damage in DNA or can be caused

1:55 through damage in DNA we our body and

1:58 has a whole system

2:00 of DNA repair proteins that repair

2:03 protein repair damage DNA all the time

2:07 so that these mutations aren't

2:09 propagated to have malfunctioning

2:11 proteins and as we get older or if we're

2:15 exposed

2:16 to radiation for example Sun the DNA

2:20 damages more and hence we have these

2:23 problems so I one of the uh aspects I

2:27 was working on in the Netherlands was in

2:30 collaboration with experts trying to

2:33 determine these DNA repair proteins and

2:36 understanding the mechanism of action

2:38 okay so I've got to ask you now so

2:41 you're working on networking trying to

2:42 solve networking problems and Building

2:44 Solutions on par with domain experts how

2:47 close were you to solving this aging

2:49 problem oh I think I was more into uh

2:53 solving the particular DNA repair

2:55 Pathways associated with certain

2:57 diseases the Aging problem is still

3:00 ongoing trying to understand how we can

3:03 reverse aging it's an active area of

3:05 research at the moment and but with more

3:08 knowledge that we have on all the

3:10 different Pathways and how you know our

3:14 DNA can become damag the more we'll get

3:17 to solving this problem okay I don't

3:19 want to feel too bad about hiring you

3:21 like you know solving aging that's a

3:22 pretty critical problem to be solve now

3:24 I understand also that you know when

3:26 you're working on this problem really AI

3:29 wasn't around at the time maybe describe

3:32 a little bit to the audience you know

3:33 what techniques I know you were visiting

3:35 linear accelerators and stuff whatever

3:37 you know when the hell did you need to

3:39 go to linear accelerator for so we

3:42 employed statistical methods in order to

3:45 solve uh the structures of proteins so

3:48 what we had to do first was collect the

3:50 experimental data so we first grew

3:54 crystals of the protein of interest and

3:56 then we would take it to for example

3:59 synchrotron

4:00 at that synchrotrons we had access to

4:04 highly intense x-ray beams with that we

4:07 could collect a defraction pattern that

4:09 defraction pattern gave us the

4:11 information we needed to solve the

4:14 three-dimensional structure of

4:16 proteins and through statistical methods

4:19 we are then able to determine the atomic

4:22 coordinates and um analyze the molecules

4:26 okay so you were building a big

4:28 simulator you're were trying to build

4:29 simulation to simulate how all these

4:31 proteins and amino acids were folding

4:33 was that the basic technique so indeed

4:36 we developed a model in order to

4:39 determine the atomic coordinates but it

4:41 was based on the experimental data but

4:43 then as I moved to the Bay Area a

4:47 transition occurred and that was

4:49 actually using AI methods So based on

4:53 the hundreds of thousands of protein

4:55 structures solved Deep Mind and Google

4:59 got involved to try to solve destruction

5:02 try try to solve structures without the

5:05 need to go to the synchrotron to collect

5:08 the experimental data so they trained a

5:10 model to try to predict the fold of a

5:13 protein molecule and they were very

5:15 successful at doing that okay so this is

5:17 the transition when you decided to

5:19 transition from academic to back into

5:21 industry and data science you know maybe

5:25 a little bit more about that transition

5:27 and you know is this a case where AI

5:29 actually put you out of business you

5:30 know you how many years were you working

5:32 on this protein problem I had started as

5:34 an undergraduate in

5:37 1996 developing algorithms and applying

5:40 them to protein structures um I was very

5:45 interested in AI yet I realized in order

5:48 to do it and understand AI methods

5:52 coming to the Bay Area would be the best

5:55 solution for me so indeed it was a huge

5:57 transition going from academic to an

6:00 industrial environment but at the end it

6:02 was all about the mathematics and

6:05 another Advantage coming to the Bay Area

6:08 you have access to a tremendous amount

6:11 of data with that data for example what

6:14 we have at Juniper Mist we can tackle

6:17 big problems and model the situation

6:20 okay so so the the team at Google deep

6:22 mine actually were able to solve this

6:24 problem with some sort of deep learning

6:26 AI technique and eliminating the need

6:28 for the simulator the models yes you

6:30 know can you give the audience a little

6:32 bit about you know do you how much do

6:34 you know about deep mine and what type

6:36 of models were they using to you know

6:38 solve this protein problem similar to

6:40 what we see with chat gbt or these

6:42 different type of models the models are

6:44 not Transformer based but they are deep

6:47 learning they um so in the case of GPT

6:51 we have billions of data parameters for

6:54 the case of Deep Mind and their deep

6:57 learning model they have access to maybe

7:00 100,000 but they probably truncated

7:02 their training data to a bit less but

7:05 there's a great deal of information

7:07 within each of those structures and they

7:09 were able to develop a Target function

7:12 that um understood the protein folding

7:15 problem and then validate the model

7:18 since we they had access to all this

7:20 data so this was a revolution because

7:23 previously as you mentioned most people

7:25 were just trying to they considered it a

7:27 big optimization problem they weren't

7:29 looking at the vast amount of data

7:31 available and they were just thinking oh

7:33 if I optimize this problem and I search

7:35 All Through parameter space maybe I

7:38 could get to the answer they ignored all

7:40 the data that my colleagues my fellow

7:43 structural biologists had and that was a

7:46 difference that deep mind okay so so now

7:49 you're at Juniper miss you know you're

7:52 working on trying to build a solution

7:54 that can manage operate networks on par

7:57 with human domain experts you know you

8:00 know one minute you're trying to solve

8:02 aging proteins now we're overall trying

8:04 to solve build something on par of human

8:07 it domain experts similarities proteins

8:10 networking I think the key similarity is

8:14 just the scientific approach you know

8:17 you have to define a problem you have to

8:20 know the data that you have and the

8:22 limitations so in the case of juniper

8:24 Mist we had access to both zoom and

8:28 teams data

8:29 this gave us an indication of the

8:31 quality of the data but thanks to your

8:35 efforts in part we had all this data in

8:38 the cloud that we can combine with that

8:40 data so we can determine which network

8:43 parameters or client parameters are

8:45 affecting Zoom calls so we have

8:47 developed a model to do this an AI based

8:51 model and because we have access to

8:54 millions of data we're able to then

8:57 train a very accurate model and then

9:00 more importantly validate that model you

9:03 know I make a barrel of wine and you

9:04 know what I've learned is that organic

9:07 chemistry is a hell of a lot harder than

9:09 Wireless you know trying to make a good

9:11 bottle of wine wireless wireless is much

9:13 easier you know what do you think

9:15 solving the protein problem solving the

9:17 networking problem same level easier or

9:20 harder that's a great question by the

9:22 way the best approach is just to

9:25 iteratively improve those methods so at

9:28 the start of the protein folding problem

9:30 people as you mentioned were just trying

9:32 to simulate the model right they weren't

9:36 using the vast amount of data but then

9:38 they learned to use all the data um

9:42 there's still some networking problems

9:44 that exist that we will tackle in the

9:47 future we started off with getting this

9:50 label data from zoom and teams we had a

9:53 well-defined problem we had a huge data

9:56 set and then we could try to explain the

10:00 issues now the powerful thing we have

10:03 all of this expertise at Juniper mist

10:06 and they could validate that problem as

10:08 we validated whether we came up with the

10:10 correct solution or not and that's I

10:13 think the key we have the endtoend

10:16 process all the way up to validation so

10:20 I realize I might not have answered your

10:22 problem completely like what is more

10:25 difficult but you know this I think is a

10:28 big first step to having AI methods in

10:32 networking including like the post

10:35 connection problem which we can then

10:37 create even more problems in general

10:40 problems and maybe on the journey give

10:43 the audience a little bit you know you

10:45 know in terms of building a model that

10:46 can accurately predict zooming teams you

10:49 know where are you on the journey you

10:51 know can you actually accurately predict

10:52 a zoom teams user experience now or are

10:55 we still working on that we are able to

10:57 accurately predict a

11:00 zoom um client minute so for example

11:04 what do I mean by that if a client were

11:06 to run have a zoom call we can predict

11:10 for example the latency or the packet

11:13 loss that client is um experiencing or

11:17 will experience how do I know that based

11:19 on the metrics based on our validation

11:22 training data set and but what's but so

11:26 what so why what's the big deal about

11:28 predic latency and packet loss the nice

11:32 thing that comes along with AI is this

11:35 whole idea of explainability once we

11:37 have an accurate model which we know

11:39 from our statistics we can try to

11:41 explain that and that is what our

11:43 customers really like they like looking

11:46 at the feature ranking so what are the

11:48 important Network parameters so for

11:50 example is it a capacity issue is it a

11:53 coverage issue or is it a client issue

11:56 that's what our customers want and

11:58 that's uh what we're providing them so

12:01 and we validated it not only internally

12:04 but externally as well now now I've

12:06 heard you talk about shapley and you

12:08 know maybe for the data scientists in

12:10 our audience here maybe give a little

12:12 bit explanation you know Mutual

12:13 information shapley how does this

12:15 shapley help you explain the network

12:18 features responsible for the prediction

12:20 so at the core of shapley is it

12:24 takes I have this concept that I've

12:26 mentioned a few times and let me Define

12:28 it I call it a feat that's what's called

12:30 in what's used in AI what this is is

12:32 essentially a network parameter we are

12:35 using to try to predict the Model

12:37 Behavior so some example of features

12:39 that we're using are the client RSSI or

12:43 the number of clients on the AP or you

12:46 know the return trip time from a

12:49 particular AP so shapley considers every

12:53 single combination of all of those

12:55 features in order to explain which one

12:59 is dominant and that's what we want to

13:01 know which feature is dominant that's

13:04 leading to this poor Network performance

13:07 and the mathematically proven fact is

13:11 that shapley values are objective and

13:14 unbiased so we know that it can provide

13:18 this information so in contrast Mutual

13:22 information it only looks at pairwise

13:25 comparisons so you can only compare the

13:28 for example the latency from Zoom with

13:31 that particular feature you can compare

13:33 at a multivariant level okay so so it

13:36 sound like we have a we have a model we

13:38 have shapley you know maybe give the

13:40 audience a little bit how much dat how

13:42 much data did it take to train this

13:43 model we talking like terabyte is this

13:46 like chaty PT where you took the whole

13:48 internet and trained this model on so at

13:50 the moment what goes into training a

13:52 model is usually a million data points

13:55 so what I mean by a data point is that

13:58 we have the labeled latency that we're

14:02 getting from either Zoom or teams and

14:05 then we are combining that data with all

14:07 of our Juniper misted um network data so

14:11 combining that all we train the model

14:14 and we use that model to then predict

14:17 what our clients are uh experiencing you

14:21 know maybe for the audience you know an

14:23 example of how this is really helping

14:25 customers or you know the needle and the

14:27 Hast stack problem that it teams face

14:29 so we've uh uncovered a few issues but

14:32 one of the thing that we've seen that

14:34 has been quite prevalent is the idea of

14:37 a customer connected to a VPN server so

14:41 before both zoom and Microsoft teams

14:44 have identified that a VPN server can be

14:48 deleterious to a call but something that

14:51 we have discovered is that it's actually

14:53 not just the VPN server that is an issue

14:57 but actually the distance from the

14:59 client to the VPN server and if there is

15:03 a large distance say someone connecting

15:06 from India go connecting to Australia

15:10 that leads to a very poor uh video or

15:13 audio performance okay now R well I have

15:15 to say I'm starting to feel a little bad

15:17 you know solving the Aging problem

15:19 solving the network problem still not

15:20 sure I made the right decision you know

15:22 but maybe for other young professors and

15:25 data science coming out any words of

15:27 wisdom and advice on there Venture into

15:29 the data science AI world for me at

15:33 least from my experience as I started

15:35 off with I loved mathematics I found the

15:38 area that I really enjoyed doing and I

15:42 found the place that um I've always been

15:45 lucky that I found a place where I can

15:48 do this and apply it to problems which I

15:50 think are very critical to me so that's

15:54 the number one thing if you have that

15:57 passion you will be able to find

15:59 something that you enjoy doing you know

16:02 whether it's structural biology where

16:04 there's still a lot of room for

16:06 improvements whether it's networking or

16:09 you know even natural language

16:11 processing developing something like GPT

16:14 there's so many areas of AI so for I

16:19 love for example thinking about ethical

16:21 AI thinking about um the hallucination

16:25 problems of GPT how can we get around

16:27 that I mean these are all open areas and

16:30 you know there will always be room and

16:33 always be uh just keep in mind I think

16:37 the guiding principles are the

16:39 scientific uh method of

16:42 experimentation testing and validation

16:45 and then reiterating well now Raj I want

16:48 to thank you it's been a pleasure

16:49 working with you I can say and I want to

16:51 thank everyone for joining us today on

16:53 Bob Friday talks and look forward to

16:55 seeing you on the next episode