>>Welcome, Stefanie.>>Thanks.>>It’s a pleasure to have you here. Like we were just chatting, this is not your first time here. So you spent summer here

as a research intern?>>Yeah. Many years ago in 2012.>>Yeah.>>I had a good summer

and I’m just recognizing the building and memory.>>Yeah, it’s almost like

a rite of passage, right? It’s amazing to see how many prominent researchers

in our field at one point or another have passed

through an internship in this or one of the other labs.>>Yeah. This is actually

what I got to know then, like all these other people also have interned in the same

lab back in the day.>>So of course that was still not a very early point in your machine learning

research career. You’ve been doing machine

learning research for a while. If I remember we crossed paths, I think it was 2006 in Tubingen.>>Yeah.>>When we were both undergrads. So tell us a little bit

like how did you get, I guess maybe more common these

days but it was not very typical at that time for

undergrads to get started in research and especially in

machine learning research, so what made you get into machine learning research

as an undergrad at that point?>>Yeah that’s kind of an interesting story with

lots of coincidences. So I think I started hearing about machine learning when I was actually an exchange student at UT Austin->>I see.>>- back when I was

an undergrad, and so I took some courses there and

I took a data mining course. I actually enjoyed that course. That’s where first time I heard about clustering and

such kinds of things. Then I did a research project, actually on neural networks and

computational neuroscience. Then back when I was at home, I needed a bachelor’s thesis project. That time I had roommates who were research assistants

at Max Planck Institute.>>I see.>>I was like, hey it

sounds really cool what they’re doing and it sounds very interesting and also a bit more towards the mathematical

side and algorithmic side, and I kind of enjoyed that. I was like, “Hey is

there may be someone I could ask for a project?” So I asked if they have a project

for me for bachelor’s thesis, so that’s how it started.>>So I guess the fact that you were actually doing your

undergrad in Tubingen, which was like this machine

learning powerhouse already at that time

probably played some part.>>Yes I think so. Then some random coincidences of

having met people who said like, I’m doing this and I was like, wow that’s really interesting. Then I just stuck

with it essentially. So I started working actually on independent component analysis

and matrix approximations, and I continued working on that and then expand it to

other topics. Just like that.>>Yes. So let’s talk a

little bit about that. Right you’ve had a pretty

interesting arc in terms of the variety of

research topics you’ve pursued. You started, like you mentioned, with algorithms for independent

component analysis type things. You went a fair bit on

clustering during your PhD and then it seems like

you got interested in discrete optimization

and just the body of research kind of

exploded from there. At least that’s the impression

we get from the outside. So how did you get interested in this discrete optimization

and maybe just for people who might

watch this conversation, maybe also tell us a

little bit about what discrete optimization is

and where it comes up, and how it interested you?>>Yeah, so I’ve always been

interested in algorithms. That’s almost one of the

fun things I enjoyed. Actually, the way I got into

discrete optimization was partially. I was working on

clustering and that is one specific discrete

optimization problem, meaning I have data points I need

to assign them into clusters, so I have to partition them. So there’s a discrete

solution and I have some objective function

that I want to optimize, but then of course there’s many other discrete optimization

problems meaning I have usually some binary variables and I want to optimize

over those variables. For example doing assignments, finding maybe clusters

or paths and graphs. Usually this involves graphs or sets, or partitions, something

along these lines, and where does it come

up in machine learning? In many places. So

one is in clustering. There’s many places where I say

I want to select data points, maybe I want to do exploration, I want to see which

measurements should I take to gain as much information

as possible about the world and also in

inference problems, when I want to make

discrete prediction. So one classical example that time was predicting where in

an image is an object. So I basically make a prediction for each pixel in the

image of zero or one, or is it foreground or background? So it’s an inference

problem where I’m making predictions of

all discrete variables. There’s many other problems

where maybe I want to do predictions over

graphs or networks, I want to find communities, I want to find discrete labels. Somewhere all of these are essentially discrete

optimization problems. So the interesting thing

is that typically, computationally these

can be very hard. So typically, I’ve some

objective function that I want to optimize. This can be NP-Hard, meaning it’s hard to find any algorithm that works

in polynomial time. But now, sometimes they

have specific structure. So what I was interested in that

time was a specific type of structure that enables good

algorithms, good approximations. That structure is

called submodularity. So the idea of submodularity is

that I have diminishing returns. So if I think for example

about information, I want to find a set of items that maximizes the

information that I get from them. If I have a set of items, I add one more item to the set, how much does my value

or objective increase? The more I have, the less it increases from

this one additional item. That’s the idea of submodularity, there is diminishing returns.>>Correct.>>It’s pretty simple

and hence it applies widely but it’s very

useful algorithmically. It has very nice connections

to convex optimization, greedy algorithms

combinatorial optimization.>>So speaking of more continuous

optimization and connections, so in terms of the touch points between the broader

optimization community and the machine learning

community, definitely there’s been like over the last decade, a huge growth of that literature in the

continuous optimization area. There partly it was

fueled due to this aspect that when we started thinking about the kinds of optimization problems that

arise in Machine Learning, they were maybe introducing some structures that were

not necessarily seen before or maybe some

regimes of parameters, data sizes, things

like that that are not necessarily explored enough in the optimization

literature and that led to the mushrooming of different

algorithmic styles and what people were studying. Now when you look to again

discrete optimization, it’s obviously not a young field. It’s a field that has very rich, very old literature

both in mathematics and theoretical computer science

amongst other areas. Where has Machine Learning influenced or changed a narrative in that field? Has Machine Learning added an interesting dimension

there in a similar manner?>>Yes. I think so too. So one is that Machine

Learning has also brought new problems like different

clustering problems or different inference problems

where I want to make structured predictions

or so that lead to new optimization problems. But the other one is that

maybe the focus shifts. So if you think about the classical combinatorial optimization community, theoretical computer science, the main goal was to get a polynomial time

algorithm for some things. So it could be like

something that scales on the fifth or sixth

order with the input, but it’s polynomial time. In Machine Learning,

you care also about, even linear can sometimes be very expensive and quadratic

is also very expensive. So it’s a completely different shift and there may be you are okay to have a bit of an approximation

but to be much, much faster. So then the question is, is there some trade off in-between? That is one question and the

other one is even if I have a polynomial time algorithm

and it’s a good algorithm, it’s maybe not fast enough. So one example that’s maybe very easy to understand is

the greedy algorithm. So one way to do sub-modular optimization is to

run a greedy algorithm meaning, I’m picking out of the

many items I have, I’m picking one item that’s

the most valuable and then I’m picking the next item as the best augmentation

tool dataset, and I do this over

and over until I have K items if I want to select K. So

that’s a very simple algorithm. It works very well in practice and it works very well if your

dataset is not too large. But if your dataset is large

then searching through all these many different items every time you augment your

dataset is very expensive. So now our new questions, can I accelerate this

and how much do I have to trade off the quality

of my solution for speed? Can I parallelize? This is inherently sequential, maybe I can parallelize it

or can I do sub sampling? So one reason line of work has looked at like how

parallel can I actually make these queries to not lose the approximation

guarantees that I had.>>This is the work from Harvard.>>Yes. This is the

work on adaptivity. There’s been many recent papers especially from Harvard

but also other groups. Yeah, so this is one question that is motivated by looking at

sub-modularity on larger datasets. There have been earlier

works that used other ideas such as just

sub sampling your data, how much do you have

to sub-sample and just doing the

optimization over that. Smaller set of items or

just using sub-modularity itself to do a better organization of your data and you don’t

have to query everything. So there’s been several ideas. All of these are very

interesting and then there’s also another

literature just on streaming algorithms for

greedy and so on and so forth. So all of this essentially was

motivated or fueled by saying, “I have large datasets

in Machine Learning, can I somehow run these methods on large datasets because they

work well in practice?>>In terms of the communities, does the Discrete

Optimization Research in Machine Learning and the

Discrete Optimization research in theoretical computer science

or in math and other disciplines, are they at the point where they talk to each other

and learn from each other or they’re still happening mostly

in their own ecosystems?>>I think they are increasingly

talking to each other. So they’re still theory conferences and there’s the Machine

Learning conferences but more and more you see people actually collaborate

in that space and theoretical computer science

researchers actually starting to publish at the

Machine Learning conferences. Another example is that even

mathematicians are starting to look into problems that are

interesting for Machine Learning. So another example, so

I recently also got interested in discrete

probability and distributions that

are essentially log sub-modular distributions and

a stricter version of these. That has a lot of mathematics

and algebra in this and I know that mathematicians

are starting to look at, “Hey, actually sampling is a problem that’s interesting for

Machine Learning tool.” So there’s a new dimension, a new interesting directions for those people as well and

looking at what distributions are these people interested

in and vice versa as the Machine Learning researchers learning from these people too.>>Now of course, a lot of the concerns you’re pointing

out which are leading to these new quests for theoretical research are very motivated from concrete applications. I know in your own work you’ve

always maintained a thrust for also doing research

in those applications. So tell us a little bit about

some of your applied work and often people

complain about the fact that there is theory

but it has little bearing to do on what’s going

on in practice and vice versa. So in your experience

as you’re working on better algorithms from a

theoretical standpoint and you’re also working

on just trying to get the best performing algorithm in the context of some

computer vision problem, how easy do you find progress

in one maps to the other?>>So I will say sometimes there’s a discrepancy and

sometimes they actually come together very nicely

and I like to think about things like them from a theoretical perspective,

trying to motivate. So I think a very good example is my first paper in computer vision. So that time, it was a very classical problem

of segmenting images, basically finding foreground and

background, supervised studied. Lots of research work on this and I had never written a

computer vision paper. So I had this idea of well, one problem with these methods

that time was that they were essentially penalizing the length

of the boundary of an object. I said well maybe we

can relieve this by instead saying I want to input some priors that the

boundary of the object is coherent so we can implement this via some other

mathematical tools. So the sub-modularity

tools I was mentioning. I had an approximation

algorithm for that which actually solved the

problem, approximately of course. So I just tried it and after

a little bit of trying, basically finding

the right parameter, actually worked quite well. So I said, “Well, let’s submit it to a Computer Vision conference.” I remember that time there

was another PhD student in our lab and I showed him the results and I said,

“What do you think of this?” He was like, “Wow, this

actually seems to work.” So I submitted it and it was selected for an

oral presentation at that time. So actually that was completely

motivated from an idea coming from a theoretical research which was accepted by the researchers in the application fields at that time. So that’s one example that it

can actually work quite well. I think nowadays also a lot

of theoretical research is inspired by looking at phenomena

people have observed in practice, especially a lot of the work

on theory of deep learning, trying to explain that

behavior that’s another route. That’s also fruitful. There’s sometimes still a gap between what you can

actually prove and what is actually the real model which

is decreasing but in general, I think it’s a very good

perspective to think about at least like what

theoretical tools I have. It doesn’t always work as smoothly as that first

paper I mentioned, often it needs much more engineering to actually make something

work in practice. So that was maybe a lucky case where not much of that was required.>>So speaking of like trying to explain some of the properties of deep learning

from a theoretical standpoint, I noticed that’s what you’re

going to be talking to us about in your seminar today and you’ve been working

in this area more recently. So what are the questions

you’re studying in this domain?>>So yeah, what I’ll talk about

today is a specific class of deep neural networks

that actually applies to a combinatorial objects

or graphs and networks. So in many learning problems what we want is we

want to actually make predictions on inputs that

are graphs or networks. I though we’d just put

a label on a graph say I have a molecule that

I can model as a graph. I have some information

about what are the nodes and maybe

the edges, the bonds, and the atoms essentially, and I want to predict

properties of this molecule. Or I have a social network I want to make predictions about

the nodes in that network, could also be another network of

interactions and it is used for recommender systems

when people click on items or interactions

between drugs also. So what I need is a good

representation both for nodes and of the

entire graph itself, and this has been a field

of research for long. People have long thought

about how can I best find a good vector representations

of graphs and more recently what people are doing is neural networks which are essentially

learnable representations. The way they work is they essentially find a representation

for neighborhoods of each node and then

they put them together into a representation

of the full graph. So what we have been interested in is more theoretical properties

of these networks, so one question I’ll be talking about is what is the discriminative

power of these networks? So meaning if I give it

two different graphs, will it always map them to different vector representations

or sometimes to the same one? The problem is that if you map

them to the same representation, you can never learn to assign

different labels to these graphs. So there’s an implicit restriction in what kinds of

functions you can learn. We are studying this from basically the perspective of what choices

you make when you design this neural network architecture and what you can show is

basically there’s a limit to how much they can discriminate

so they will like in the current architectures these so-called aggregation based

architectures that people use, there’s always a limit on

what you can discriminate and the limit is a popular

graph isomorphism test.>>I see.>>It’s called device

vital Lehmann test. Now the question is do all the

neural networks that people are using actually possibly achieve that limit or are they

bigger than that? It turns out that actually some of the popular architectures

are much bigger than that and you can see this by studying what are the

aggregation functions. I should maybe explained

what the aggregation is. So the way this neural networks

work is that for every node, it looks at the neighbors in the graph and collect

information from the neighbors, basically the feature

vectors of the neighbors, combines them in some way to get

a new representation of itself. Now this combine the

neighbors operation is the aggregation operation. There’s a few restriction. So for example, it has to

be permutation invariant because there’s no natural ordering of like this is

neighbor number 1 or 2. The other one is that you’re

using the same kind of aggregation operation on each

node to reduce the number of parameters and hence this

neighborhood aggregation has to be invariant to how many

neighbors you have because not every node has

the same number of neighbors. So these are the restrictions and typically the way

they work is that you apply some transformations of the neighbors and then

you do an aggregation, maybe a sum or a mean or a max pooling and then you do another non-linearity and you

do this over and over again. These aggregation operations

there are some conditions on these under which you are

more or less discriminative. At the end of the day, it boils down to an injectivity condition on

so-called multiset function. So injectivity means

different inputs lead to different outputs and multiset means that the functions

you’re actually having here get sets as inputs, this is a set of neighbors

and the neighbors can repeat, and the number of neighbors with their specific label is important. So you’ll need to actually

preserve the information about that and that helps you and you

see actually in practice also. So this is the theory

and now the question is what is the implications in practice. So you actually do see that

the networks that respect these injectivity conditions do fit the training data better

and also generalize better. So the generalization is a phenomenon we don’t

explain in this favor but it happens but you basically fit the

training data much better. So it actually does have an

effect also in practice. So this is one question

we had looked at. There’s others, for example, there’s an interplay

between the structure of the graph and the depth of the network that works

well for these graphs.>>I see.>>So there is a way to explain these interactions and

you’ll also see again in practice that there does seem

to be this phenomenon happening. More recently, we’ve

been looking at how the structure of the network aligns

with the structure of the task that you’re trying to do so this pertains mostly to questions

of reasoning where the answer is also

actually computed by an algorithm and the

network essentially has to learn to execute

that algorithm. So there’s also some interesting

theoretical questions that you actually see

reflected in experiments too>>So I guess does this way of thinking about the

architectures which are better or worse for

capturing the data, does it also give any intuitions

in terms of algorithm design for the underlying

optimization that we need to solve or is that

still being done in basically using the same

SUD style algorithms, or are there any other tricks we get to bring from the

theoretical underpinning?>>So that’s something we

are currently looking at. I think that there’s lots

of open questions there. I don’t have an answer yet but

it’s a very good question. Yeah. So basically looking

more at how the structure actually relates to the

algorithm that you’re solving. I think there are interactions

there and there’s things you can exploit

and explain probably, but I don’t have a good answer yet.>>Okay, we’ll look

forward to maybe learning about it in the next conferences. We have you here for a distinguished seminar today

but you and I both of us have, of course, quite a long way to go still in our research careers hopefully and I’m

very curious you’ve. like the nature of problems

that you’ve been studying has really become richer and more diverse as it naturally

does over the years. Really long-term, what’s something

that drives you that you would like to get to in terms

of research goals.>>I think I have a few goals

that are inter-related. So one of my goals is basically to improve machine learning

with structured objects. So all the discrete

optimization of trying to do inferences and predictions with sets and partitions

is related to that. The graph neural networks

are related to that. Basically, trying to get better at the computational

hurdles that are there, and better understanding the

models and their properties. Understanding the algorithms,

understanding how we can do much. Basically, how we can

get algorithms that we understand how they

work and why they work, and when they don’t work, and that are actually

practically applicable. So that is one thing. The other one that we

started working on is to make machine

learning more reliable. That of course relates to understanding what your

models can and cannot do, and what your algorithms

can and cannot do, and to actually

specific optimization, like so we have looked

at robust optimization. How he can, like that’s

one way how you can actually bring that

robustness into the problem that has regularizing properties and to understand these trade off. So basically using prior knowledge,

using structural knowledge, using optimization

formulations to guide the problem and I like there’s

some understanding of how we can encode what we know and what we know about the problem domain to

get better models and algorithms, and then better

optimization strategy. So that’s one big set

of inter-related goals.>>Well, that seems like it’s going

to keep you busy for a while. Now, you also had this experience of two very

different educational systems. So you grew up in the European

system and now of course you’re teaching in

the American system. I’m very curious of what you see as the relative or maybe I would say metrics

and weaknesses of the two. I will ask you a more specific

question about that in a moment, but I first just want to hear

your general impressions.>>That’s a very big question. So yes I did my undergraduate

and graduate studies in Europe. At that time, I would say my

PhD program was a bit special. It is less sold than what

typically now is done in Europe. Is basically, I did my master’s degree and then I

only did research after that, and it was only a few years. Like only three to four years. Versus in the U. S. you do your bachelors and then

you start research, and you will spent

much more time there. So these are very different systems. In between I visited the US. So I found that what I had learned

actually in the German system was very adequate to follow what

was in the American system, which I was very happy about. So I think actually the education

I received was very good. Also actually at that time at least the German system

was very much focused on giving you first abroad

foundation of mathematics etc, and then going into

the technical things. Most also possibly because

at that time we didn’t have a degree after three years it

was actually after five years. So there was a lot of time. So what I would say

though is like that, in the end I liked those

PhD programs better, where you don’t only do research, but you also have a few

courses and you can target coursework to the

research that you have, and sort of continuously

actually learn something else because it’s never easier to learn about something than if you actually have a course that you can take. Even if it doesn’t have

to be a required course. If it’s just like some fun course you can take is just more structured. So in that respect, I think that the more

structured PhD programs I would say basically you get a broader

exposure to different topics. So because if you just

work on one thing, you get more focus on

what you’re doing, but you’d see less about what else is there like

many other directions. So this is something that

when I visited the US, I notice that there’s a broader exposure you can get

just because maybe there’s a big university behind it and that is something that’s

a bit different on Europe. But at that time it was

a bit more focused. I think the PhD programs in

general are a bit different. Because in Europe it used to be

the case that you were admitted, like not into a program, but you’re basically hired by the professor and that’s

like your advisor. Versus in the U. S. you’re

admitted to the research program. So when principle you can

still change advisors. It’s a bit more flexible than that, which is sometimes a good thing

if you’re coming in and you don’t actually know what

exactly you want to do as is the case with many students.>>Yes.>>Even though it’s good

to focus soon enough.>>Now, the more specific thing

that I wanted to ask also was, so one of the things that of course, we’re becoming increasingly aware of and there’s a lot of

desire to do something, but we don’t always

have good answers is the overall lack of

diversity in STEM fields. In particular, when we

look at the statistics for machine learning conferences,

it’s pretty damning, and if you start going narrower

into machine learning you realize, sometimes in fact all of the machine learning

is still maybe being, there’s like a handful

of areas that are maybe making all of machine

learning still look good, relative to what it might really be. So I was curious in terms

of your own experiences in the German system and then

now what do you experience, what you see the students

going through as a faculty like are

there things that we can learn that we

can incorporate from the German or other

European systems that maybe are helpful in terms of improving the

diversity in STEM fields. Just what was your experience

like when you were, I’m sure you had to overcome more than a handful of challenges

to get to where you are. So what kept you going?>>So first the thing is I think the same problem

persists in many countries. Many times also in Europe I

have been the only woman.>>I see.>>So I’m not sure there’s good

things to learn from either side. I think there’s other countries

where this is different. But I think in general, so the thing is there needs to

be support at multiple levels. So one thing is getting

people interested to even start say a college education or graduate

studies in this field. So it actually needs

to start very early. I remember when I was

in high school and I was saying I’m going

into Computer Science. People are like, why do

you want to do that? There’s like go into

Management or something, but Computer Science

that’s so uncool.>>That’s not necessarily

gender-specific. I remember my dad, so my dad was an Electrical Engineer. He was like you know

these computers there they’re done like the dot

com boom has happened. This is just a fact, look like nobody is leaving

electricity behind. You should do double E.>>Okay. So there’s that one, but is?>>But I still understand

what you’re saying.>>I think just the thing of

the thought of as a woman, what are you going to NSR? All the other women were going into business or something like that, and maybe made similar. So it was just not a typical thing to consider and basically

to change that mind. I remember I had some physics teachers who also

made some remarks etc where I was at that time. So basically, even showing people that there’s interesting

things to discover. I remember, I like

mathematics at that time. So we had this thing

where you could go to some advising and they will tell you about what are maybe

good job choices or so. I said, “I like math.” They were like, “You

could study math.” I asked them, “So what do

you do when you study math?” They’re like, “Yeah, you can

work in an insurance company.” I was like, it sounds very exciting. The way they described it, sounds like I’m not

going go into math. In it maybe, I think I would

have been happy there too, but like at that point, so basically, just

knowing what exists. At that time, I also had no idea what all you may be able to

do with Machine Learning. I didn’t even know that

Machine Learning exists. Didn’t like what is the breadth of computer science. I didn’t know. I actually went for

Bioinformatics at time. I was like I like math, I like biology, I like many

of the things that you do. I wasn’t actually even specifically interested to do computer science, I have to admit, at that time. So that’s how it devolved. Then I actually went much more. I went into computer science studies. I didn’t really know how to program. We had to do a little bit

at school that was it. So at that level, just getting

people interested is one point. Then the other one is support. So I see acceptance of women in leadership positions

is very different often. There are these things of you

just got this because you’re a woman, like judgments. Then the other thing is support. So I see also other

women who say this is too much or say

being a faculty job, it’s just too stressful. I don’t want to do that. So there’s also these things

like apart from discussions we’ve had about basically

insulting and offending, but at different levels. So one other level of support is

for example, say you’re a woman, you have a child, how flexible is

the work experience with that? That I have to say actually, I’ve been very lucky. I had a child two years ago, and I had lots of support

from the department. I wasn’t sure at that time. It’s a male-dominated

field. I don’t know. In many cases in my career, I have been the only woman

starting from PhD student. For some time, the

only other woman was the person who was the administrative

assistant in the office. So I wasn’t sure, but actually, I was very lucky. I will say it was a good thing that actually there was

support and flexibility, and there’s childcare

options available. Otherwise, it would be hard for me to do the job and just follow up. So I think if you want to support

women, there’s many dimensions. So there’s this thing

about getting respect, but there’s also this thing

about starting early to get people interested just to have enough women

interested in the field. If there’s only two women applying, what can I do in 100-men? The other one is to continue

giving support and saying, “Okay, it’s actually feasible.” Say took on my own family and

job and all of these questions, I think are important.>>So one dimension of support, I also heard at least or perceived

in your response was that sometimes it’s important if you have role models

that you can look up to. Did you have such people when you were just getting

started in the field that helped you overcome

those initial hesitations?>>Actually, I didn’t

have a female role model. Frankly speaking, I think all the people I knew

in the field were male. I mean, at least in

the more senior roles. But what I did have was very

good support from my advisors, people who have been working with, who were continuously

basically believed in me, and were encouraging me. That is I think very important. That there’s some people who

believe that you can do it even say when paper gets rejected, Tetra, these things happen, and there’s people who make your

believe that you can do it, and that things work out. That is I think very important. So someone who gives

you basically support, and that can also be

a network of people. These can be role models as well. I think it’s good to

have role models, but I think very important

role is played by advisors and people that

you’ve worked with. If they believe in

you and they give you the feeling that you’re valued

and you’re strong, you can do it. If they give you the

feeling that know everything you do is just worthless, then I don’t think you’ll go very far unless you’re really

strong as a person also. That is maybe less women

specific and more general. I’m sure this isn’t for everyone.>>Absolutely, I think

we’ve all or the course of our PhDs had difficult moments where we’ve doubted our

ability to get through things. Having a good support structure and having encouraging mentors

absolutely plays a big role. I think in particular, often we also see that make people hesitate when taking up more mathematical or

more theoretical topics in terms of their research areas. In those it’s even, I would say all the more important to have the right support structure. So I think we’re getting

up on time here. Are there any final words of

encouragement or wisdom for our viewers who might

just be thinking about how to get started

in their research or how to navigate this difficult terrain that

you would like to offer?>>I think for all young researchers, I would say honest, be open-minded. Learn about many different

things, don’t discard anything, that’s a very important

thing and be courageous, explore on your things,

and just have fun.>>Yeah. I think the last

bit often gets ignored. Well it was a pleasure. Again, we’ll look forward to learning more technical

aspects of your work in the seminar today. Thanks a lot.>>Thank you.

This different approach by Microsoft is a nice innovation. (Maybe I have missed others). And it's classy in having a well prepared host chatting to someone interesting who is able to followed up on other vids.

good dicussion