Let me start with my country and
the all need clean water to drink

this picture here shows the support by the
distribution network really interesting

what it taught us the content of gets
into this network to an accident

of this activity this continues you could
spread it we're trying to affect a large

part of this network and cause them to
population to event that you may be

interested in applying sensors in this
network such a successful in the right

here which allows us to tap these
continuations reactivity prevent heart.

I think their senses are very expensive
and with limited budget the have to

decide if we should place those
accessible to achieve as much protection.

Instead of deploying sensors.

We're also interested in the using more
sensors to explore the environment for

example we've been working with
a ninety percent scientists

on the autonomous
an environmental monitoring for

example monitoring water quality takes in
order to be able to do that you need to

be able to monitor large spatial anomalous
just temperature distribution and

fluorescence and one of the two
that is used by the sciences.

That is in these robotic part of the city
with the sensor to try to figure out for

example figure out what is the temperature
in the vertical slice taken

through the slick you don't totally
every measurement take some time and

also take some time for the size of
the child between these locations and

a limited amount of time you can only
get a small amount of observations.

The question is from the from those who
would have been to make predictions

of what the temperature is at
the kitchens we don't know missions and

a question as well as she says in order
to get the most accurate picture.

So these are the kitchens
where you wish to

censor the knowledge you
learned something about.

Well also going to talk about implications
in the social networks and with

the advent of last year like that word
such as Facebook and Twitter and so on.

If people gain new insights about its
host site work feels like functions.

And if you look if you want
to do that need to be able

to understand structural properties
of these large social networks.

So for example this shows a network
deferred from the information public.

If you are going to talk about later this
tutorial and if he has a better word.

You may ask.

Well it's not one of maybe the most
informal influential knows it's not

the highest however there is the impact
in spreading information across.

So the system of education and
how these and

any others require solving
this we don't miss it and

turns out that many of those can be
formulated exactly the way that young

mentioned in the site authority in
the previous talk and it can be used to

the problems of maximizing SAP functions
start with a finite set this could be

location to place our sensors we can make
measurements of the mobile sensors but

also in the start of this
information diffusion that work and

yet you should list so
that small subset of those dots.

Maximize something to the function and
still seems to trivial many of those

utility functions actually turned out to
be so much because of the problems that

are going to talk about
things such as the result.

So the first part of the story.

I'm going to talk about at the kitchen
smart maximization inside the right

words to a targeted level of objective
functions arising in those domains

that talk about optimization problems
by particular what kind of constraints

rise to those occasions and result that
they got the algorithms that can be used

in order to solve those problems even
though I want to focus on the algorithms

because you already have heard quite
that although in the previous part and

you hear a lot more about how it was
in the other parts of this writer

true then I'd like to show off
some of those I wouldn't really

write very strong columns on
just going to the first half.

The story will and
in the second part of the Tory are going

to talk about an interesting
connection with the smart musician and

she learned in looking at problems
that the on the power problem insisted

even through data and to be able to
sort out the opposition problems.

Based on top of that they've
learned about decision told Time.

So let's start with a very simple
projections talk about sense of placement

and just cut to a picture.

Maybe you want to place a bunch of
sensors in to believe in order to monitor

temperature for
example there's a fire at one recently or

formalizing the problem would
be to see that maybe the sensors

tell you something about the temperature
of the video perfectly predicts

what the temperature is in
a certain summaries of the Hi dear.

Hold on to us or
anything outside that range.

If I needed something
then if I wish to a place

I could fall that's a problem in that if
we met by this replacing the environment.

I have said the apartment key for
inspection protection place the sensors

and sensor placement simply
a subset of the patients and

the value or bodies of those locations
to start the total area covered and

that's exactly this coverage function that
I've talked about my previous OK that's

one point about the function I really get
something of white from his mere right.

But of course that's a very
very simplistic model but

it seems that the second part
to decide everything or alter.

Nothing else for now it turns out
that a lot more realistic objective

functions smartness Well if you somewhat
realistic model of anything about science.

It may be how some statistical model
that describes the phenomena that

you're interested in regard to the more
the temperature the different locations in

this building and spread variables to
the other reading variable Y. yes it

look each an S. discussed temperature
there and he's ready variables a car or.

Example that model of our assumptions
about statistical dependencies between

these variables not a variable to tell
which itself is what we like to infer but

it's unobserved.

But it one of those who gain
information we can place answers.

OK For example a complete sets imitation.

Yes and that would give us a noisy
observation excess of the temperature.

OK That is one vehicle we can
think about modeling sensor

and if you have such a model
the most you can do is you can

describe our assumptions about how
temperatures of them looking at

this current in the different locations
a lot of questions about what kind of

exhibitions we're going to get if you
have sensors that different locations.

You can describe describe probability
distribution of less probabilistic article

all this has to go models are being used
and I once we have such a model we can

use it in order to gauge
the informativeness of pain and

one very actually to do that is
sufficiently Asian which is a popular

takes a set of the kitchens and
quantify how much the observations at

those locations help to reduce the ice
certainty about the temperature holes.

So it's a difference of two terms the
first is that certainty twenty five times

the challenge of the of the system usual
Americans would like to affirm infer minus

their residual actually yes certainty that
they have left about the temperature given

us additions at the locations that you
have access to this is a function and this

allows us to quantify how useful it is to
how its missions are different locations.

Right.
And for

example if the sped up sensors
shown here to me providers.

I was very I was a lot of information
about the temperature anywhere but

if you place a very close together.

Well you have done the scene right in
the missing cards and butane LOL well

that's exactly the reason why similarity
comes in is going to the kitchens.

And it turns out our guys.

I've started.

This is actually part of the subject of
Oculus modular so we can define family

a set of variables white one through Byron
he said are obsessed with lights and

fur and various X. one threats are
variables that can potentially looks or

do you want to select sort
of those variables and

detect the functions just as
much information but turns out

that the function is always more Islamic
but in some cases it's not a model.

It's logs.

But it turns out that if
the variables starts at agency made

satisfy an actual conditional
independence property.

Maybe the sense that indicated measurable
then it turns out that this objective

function is in fact the smarter function
and it's not that hard to show actually so

let me just start you through
about three bets at the Prince.

So what we can do is to form some
of the list you can look at what's

the marginal gain of any of your solution
S. given that you're going to have made

observations that said evocations and
using some basic information related

problems these you can write up as
the difference of two terms the first term

is the entropy of the measurement location
just give it up submissions and locations.

Right.

And it turns out that

is it on Mondays it actually needs a very
basic information theoretic inequality.

Also on call information that hurts all
five of this one was it evens out yet

for years and there are years.

Sure I should be about this mess
medication and get them out of the jam.

Basically you're able noisily send it out.

OK And that's just a concept independent
of OK so last night I heard the monologue.

I just want someone to try and
subtract a concert from it and I've

heard all the about makes model and since
you already know that approaches are you.

As my brothers.

Want to talk it's OK So this is a very
natural function resting a lot of those

patients and if you have to start at
the function you can follow that sensibly

small problem was just kind of constrain
organizational thoughts with much of

the talking as well as the best Q.B.
Asians are actually able to be right

there and it turns out that this is that
you have a problem for us was for a lot of

the other patients guys that arise object
to functions the rest of the supplications

by this period seem they
are good outcomes for

maximizing Thomas Mucha factors
particularly over the green young

actually works well and
that is the sort of cellular sort by

the most efficient will see
that it's on the safe side to

take money from some of the function
of the discounts in fact approximation.

But it's constant practice
about six years and for

a lot of their objective function is the
rise in these kind of sensory problems.

This approximation factor is
the fact the best possible for us.

I cover that sort of idea at all.

And fortunately she can also
provide better than this

but there's a lot of practice
of course practical problems.

Often times one gets a lot
better if you guarantees.

So this is some of
the results of your data.

So you take the temperature
sensors deployed in the building.

I political asylum and
then what you do is B.S.

actually fit a model to that later and see
how much you feel to me to insist it be.

I'm one of those answers and the profit
shows whole commissions mission Greece.

Increases it used to be out of them and
it turns out it's for

the problem we can actually cover the
small enough to get some fresh involved

in order to fire up the video a bit scary.

So let me tell you all I
want to take that vacation.

If you actually try this.

It's the politicians who people said
a lot he said in cars classrooms chairs

office environments but the big bill sort
of tells about what they're up to for

example start paying attention to the top
of the top to the very hard right and

recognizing closer to these could be very
useful for the different occasions but

if you're the driver assistance system
that you want to be the area of their

drivers actually drive slower but
how do you get this information but

you could put sensors on the chair
right here is a very full size so

that has a lot of the resolutions about
once as a first person to me that

you heard from the you can see it put
the person on the chair or the pressure

distributes crisis person leaning left
leaning forward and slouching right.

It can I guess kind of guess what
the flight is different posters meters by

looking at polls at those pressure mounts.

OK there's a sort of a very simple values
to get a pretty good gestation accuracy

to say two percent on ten paces of
fosters across subjects trains on

subject tested other subjects so it's it's
pretty obvious that it's got a problem.

Unfortunately this sensor
is quite expensive so

cost several times not you think
about putting that in the car right.

Don't want to have the price tag was by
several thousand dollars just to have this

purchase and so or
maybe get some of the iris.

Monitors and say that's not a problem
that you know that you could

fit a promise to do this to the state or

expense to try to think about where to put
those sensors in order to get good actors.

If you can figure out some of
the sensor readings that is difficult

to discern the variables and
then you try to predict the poster

given the sensor readings
at the locations that the.

Served.

This would be our guest and

suggest a sensor sensor and
instead use a study right.

And before you have percent accuracy
this very very expensive sensor.

About the same records for us that one.

Let's see here let me
go back to this water.

I want to work a problem.

So the studies of your
sort of quite a bit.

It turns out that this problem of
deciding what to place at center what

the distribution networks receives
a lot of attention the community

in fact the organize the competition
about this charge and this competition.

They give out and I'm on my small reading
The problem here you are treading water

distribution network insisting
about the twelve thousand nodes and

they also get all the simulator because
I develop it in model Protection Agency

which allows us to make predictions about
where to go to initially go get started

and most of the competition has started
very large set of outlandishness

several million of these
new goals to place a start.

Most of the objectives for example when
it's time to detection by them maximize

their predicted population using very
disparate time realistic the city's models

which are for coming here very quickly and

you go through the distance
to a place of the phone.

Well.
On average.

This is ready to put in if you have is
somewhere over the bucket and now it turns

out that the leaves of a digital cell the
jackpot is actually smaller the user so

that here just explain what happens if you
have one network and maybe just one time.

It introduces one of the goals of
the cycle of discrimination threats

by the facts a part of this network and

now what you can do is you can think
of what's the benefit to write if you

put on the sensor somewhere does
not work right here in this case.

It allows us to detect this event at
this particular point in time and

prevent all the shots time.

Now if you have another sense
of this in this network.

I believe your text is going to mediation
there really are the end of the Chief

a lot of potential and it turns
out that here you get the function

which essentially is to find
one of the person army bases.

You can tell it is to do some are the
scenario of other public intimidation

spreads and if they are there
ended you right in that scenario.

All right all said Eat
a census of the place and

start with cites waterfront is
just a mix of budget numbers and

the overall objective is to not average
just a bunch of smaller functions

areas together which I'm a bit smarter had
a lot of money smart a lot of pressure and

if you could use the greedy out of them to
optimize the fact that's just what you did

he tell you what some of the results
of this in this challenge.

There are thirteen participants
both record in your industry and

the performance measure of respect to
thirty different settings on the site

their methods used to come out of
this elections actually quite varied.

So there's a lot of local search receipts
genetic algorithms that can develop.

I'm looking at such
risks to actually start

just based on the meeting all the people
looking at that were saying I know a lot

about what the networks here's where I'm
going to put the sensors but didn't bother

to about this in this particular location
and steps also issues that I'm trying

to appease instead of programming to
actually solve this problem not me but in

terms of the problems it's what works and
I was so large that they wanted to solve.

So they have to pick a small subset of
the problem and try to optimize for

that and here.

Maybe you're missing the mark is
the missing Mark soon as we know

these are not our numbers.

Yes these are numbers that the onus and

source of the groups working on this
proposition are before all of the sudden

he made for
quite a while very puzzled by this.

It's really simple.

We are actually we try to
understand what's what happened.

And so yes it is often the issue
of why the thing that happened

while I told you this these large
number of particles in every sphere.

We're going to get several billion and
I'll tell you what the simulator is for

you to see if the effects
of each of those events and

times of that is actually quite
by chance more than it's actually

all the other entries in the competition
piece of it to be on the optimization

of a small subset of those scenarios but
the others.

We actually went ahead just suited
all of those provisions which

generated a lot of data
using some compression

you just put it this is also
very very accurately in value.

It's every kind of place and
it's a good value.

It's very very good thing.

Very precisely but
it also means we have to look at sixteen

gigabytes of data just about every for
success.

OK So let's look at the running
time it's about first of

all exhaustive search is hopeless and you
kind of you're going to pare successors.

But even if you just run the query up
on just adding one sensor at a time.

This is a highlight in my eyes the
committee takes one and a lot of time so

it takes thirty hours to place for
the sensors and for

all your time additional with multiple six
weeks and exactly the power that long.

For for instruction in terms of
this is where a supporter Larry

really came to the rescue and
then let me just tell you this.

It suggests that inside needs many

factory years back it's very kind
of it's like St Paul practice

inside of him actually accelerated video
with him and said he works like this.

So one has to be out and do.

Well the first iteration that
calculates the marginal benefit.

How much just every single bit healthier
you do it for all the evidence you have

right all the twelve thousand sets of look
at this because then you get the best.

That's what we are now the evil
committee if you go ahead and

do the same thing might be calculated
smart looking for all its bitterness.

That's right and
you do that until you place it all.

And that's what makes it expensive if you
have to do what's right for you now or

the insight that you have this
that you can use more narrowly but

it will get wiser the money benefits can
never increase it's what you can do is you

can essentially just keep an ordered list
of the marginal benefits from the previous

duration just sort of according to them
and now if you want to find the next

item if you just reevaluate
the item on top of the list right.

And usually diminishing
marginal bettors that value

is smaller than an hour about from abuse
situation but it can never go right.

So if it's the decreases.

Well you just reinserted
put it back into the scale

that we look at the next element right.

And if he found that
either to be evaluated

the marginal gain of the settlements.

If it stays on top of all the other and
if you could pick.

Well you know that we
found next to that right.

It's a very simple idea but
it is actually useful and

side different looking to
fall back to very late start.

Two days instead of six
years of get the benefit

from the candidates objective very
very accurately but it is very slow.

OK So that was just a few examples
of massive problems problems.

There's a lot of other
interesting directions for

example can't cope with the serial margins
can cope with more complex constraints.

What about if you mean signals OK So

let me talk a bit of those talking
about this area is that right.

So far I'm talking about
placing sensors to do

well on average trying to contain anything
happening at random knows what to detect

them as quickly as possible but if someone
knows what we put our sensors right.

It's just a test because it's wider
knowledge in order to achieve much.

And in fact if you compare
those two placements here which

are very different times
of the average case

the value of the worst case I
could actually be quite right.

So the question is about worst case and
why we feel that it's actually and

consider that objective factors of
information are very definitely

here happen very quickly and so

our scenario of your
job as a very mine but

it is if you have is already
here you have the right.

And it's one of the worst
small scale successes

and it's all our problems in fact.

We don't just happens to objectives but
a large number of them and

what not to do is to well use fault.

So this can be.

Very actually cast
Max Middletons it probably and

in the scenario each of those functions
have a right this is much lower to

the middle or much smarter
functions its size not smaller

I think the question is can you solve
this guy's opposition problem here and

yours is the very start.

So it's a quite simple algorithm supply
will actually fits in this little black

box consists of two steps first guess
the value of the optimal solution.

Let's call it C. and then find the
cheapest solution which achieves value C.

with respect to all these little five.

OK so now it's just conceptually
about how the government works but

it can actually be implemented very fishy
so getting the value can be implemented

using high single circuit and solving the
second problem of finding a cheap solution

which is she sell your beast see
with respect to all these monotonic

smart object functions that fight can
be solved using simple real projects.

These are not going to
go into more detail.

I was expecting is out of this
very simple light yet but

it actually works pretty well so
what I can show.

Well it's first of all just give
you a solution which has value at

least as high as the darkness in the group
get if you have kids answers but

you have to put more than he sets out to
keep such as it's by perseverance and

efficiency and the Southeast quality
depends not really on the parts.

But it turns out you can actually not
get better parts Michigan some reason.

Let's hear some examples of your
system picking in most cases there's

this Norcross mission here if you want to
ever you can't even explain a few small

fraction of the are more value and
if you love those.

All the officers.

I tells of cost and
you can't prove I started out

OK so this is the other thing is not
talking about too much to be telling me

they're going to have kids so this is just
a normal to want to treat the people that

I mentioned to you want to be
in the office of the star and

so you get what you would like to do is we
would like to know what is the temperature

rise of the city this
hour of this phenomenal

right to talk to you want to figure
out what's the ph value of the heart

of this big into the flow charts
are taken for this day but

you can't or you know it's more
information is using your sensor and

from those you would have
to predict what is there.

Here's Dallas Little Pieces of
You don't have an accent right.

I don't know how to do this for
those predictions but

use problematic models like the one of the
start of the beginning of this off the top

right and the last to estimate
it's pretty R. and B. can all ask.

Well it actually says in order to me.

Most are not some British or
worst case picture.

So I'd like to make sure that it's
confidence parents are squished together

experts right and
the parents of the setting

as I have heard all of
these kind of this expect

exactly first one to predict the very
basis that was difficult and that turns

out to be a lot of opportunities that
sort of condition for the clearance and

that's the problem of minimizing
the worst case that the variance.

Exactly.

Here is for something that you
talked about earlier to minimize

the next Americans and so
you can use arguments right.

Well suggested a state
of the art students and

our highly functioning simulated
annealing are with us a lot of it

just because it's community
seven hundred have to be.

But for
the time probably the sightings and

here is how it's out of the phones on this
as a problem to be accessed a number of

observations to make the white excess is
the worst case prediction there are those

that like it shows the values that
we are just really trying to prove

as much as possible about with respect
to this what's good relation or

not it's silly to milling out of this
quite a bit better than if we are and

this kind of Charles out of the
destructive use actually I was actually

pretty close to there
to the still there and

sometimes the slogan search is a bit
better but it's very interesting.

Why this is here.

Another problem instance I was the goal of
it in my city already pretty reciprocation

one of the descent right from the Pacific
northwest of the United States and

again some of the falls pretty
close to the State of the art.

One thing I have to say is that this is
the audience lot faster and there's more

money just to tweak but it's just another
one of the off particles one hundred

sixty one two places in order to minimize
the worst case of detection turning

over the conditions in areas that people
separate and there's something interesting

happens if you think we are going to be
does not prove the worst detection time at

all reasonable that is that honesty
in place at least transfers.

There's at least some event that none
of the sensors that detects this little

house and there's just nothing
no signal it becomes less and

leaves leading out with our randomized
eventually finds a good solution but

actually there's quite a bit worse.

It's hard to

get more sense of that movement

than this because it is so important
that we just carry solutions out.

Certain values came the scene.

So that's actually I'm sorry I
thought a serious problem for

about this is a period of
Harriet's that was seeking.

OK OK so I just want to say here.

There's a lot of interesting aspects.

So far we've been talking about
cutting edge to constrain

opposition What's the best
the televisions if you play successors.

But it happens there's a lot of
other interesting constraints for

example it may be more expensive to
put sensors in different locations.

Right.
So it's very actually how NASA constraints

that have a much more money function and
survive the question right.

And it's using for example this
question of the moon really to me by

Yes it's really approximation our need for

constraints arise very naturally
actually if you think about making

decisions over time and see a bunch of
sensors and you want to decide which

of those answers should vary at
what point in time you can see

the problems maximizing multiple functions
subject to a partition a very constrained.

I don't have to explore
the literature by the it's about.

But there's a lot of other
interesting constraints.

It's a for example if you have a small
sensor the wrongs are all the needed bards

and the actually have to come straight
to whatever solutions he picks

are not too far apart right you have
to the model car takes to move from

one of the lengths to the next and
what can vary I should model that

by modeling the locations that enable
solutions of those in the ground and

then requiring that there are solutions
that are connected together.

It's that selected are connected to
this crime scene caster problem that's

this is one of your interior problems and
such a detector we have been doing

some very interesting work and different
folks several years back are proof that.

Well that's the only
time I'll be going but

still how it's going to feel that
is approximation guarantees for so

much more in general and
it turns out that function is not

just one of hundreds of Marlowe's it also
has an actual candidate for example if

you take a position that are very far
apart the function is approximately.

More like but if you take a function that
takes two she said very little spying.

Then you get the machine terms which
is very arduous consisting of you

can guess it's like but
there's probably also if you think about

our place in the network of wireless
sensors that need to talk to each other

right to transmit messages to the station
for the purposes of analysis.

Then it has a problem.

Essentially the guess is that the set
of nodes in the ground with the smart

function find the words it is right and
required but it was selected disconnected

connected in the graph and
doesn't study the pointed right side of

the problem by kind of this would
all get rid of the prosecution's

together the objective function satisfies
the additional bookended conditions.

You can get is that these are just
some ideas about the constraints of

the right at the start of the race of
these applications which will be different

from the constraints of this talk to me

so that I'm going to talk about sex and
social networks.

Let's move on to change gears a bit and

talk about the different application
of me talk about security and here.

Let's talk about one very
interesting smart musician problem.

Remember it's just a beautiful paper
by David Templeton clever going

to the toilet.

Now almost ten years back.

That's sort of a lot of
interest in this area and

it's what informs
maximization it's a trend.

Suppose you want to run
the marketing campaign.

To create a demand for some new products
want to get people on Facebook pivotal.

To get them to buy some of
the product like a cell phone.

What.
You could do is put Think about exploiting

social networks for
might try to give the product for

free to a bunch of people and
hopefully convince

difference that this is by the time
the whole model of this influence spreads.

Well one way to think about that is to use
the problems to model we have this social

network here.

Each edge just indicated by some
probability just in terms of influencing

the product for type to enlist.

They just search terms to different
spot maybe with thirty percent chance.

I should influence power trails and

she's fifty percent chance to influence
char Egypt and they're trying to get fifty

percent chance to Flynn's bar which may
fail and forty percent chance to start.

It's all right.

As you can think up some
of the spread of confusion.

Remember how well it is given
this model you could ask.

Well if I have to pick one hundred people
who want to give the product for free to

maximize the by product that

we can cast just last year on
the charges if a problem or

a ground starts to set of people or
objective function takes a neutral site of

hundred people that target
investors expected or funds.

I mean those are in the air.

And that turns out to be smart and
actually very interesting

to see sort of Yes is the proof and
can give or has it.

So instead of analyzing this complex
process as a unfolds over time.

Yeah your story is typical for

essentially flip coins associated
with all the interest in advance.

Ok flip all these kinds of events and

some of these ideas become active and
some of these.

It's become an actor you
know you can say well

if I give the product to Ellis I can
read off who she's going to be right.

He's going to be Charlie Barker Dorothy.

But nobody else writes and so I can look
at what is the value right from a set and

the outcome of kind of let's let's call
it COD The inevitable the flip side of

that and dice turns out just to be second
probably right to overcome structure.

OK And so no again I don't make up for

this kind of overall objective
here is just the expectation.

With respect to the idea is
the fact that it's made to

function with

yes

yes I think that it was a lot of right.

It's also your books Dres

So this

is just one small part of this for the
next whole bunch of other models that turn

out to lead to market musician problems
there's been quite a number of interesting

results from the scientists to mainstream
to study which of those models right.

So I was more functions just in
order just to work on competitive

flaws maximization various people
thought about I think about Coke and

Pepsi both want to sell maximize the
number of customers that they reach for

the start of the writing I mean it's all
understand how they should write it.

Well.

It's game study press and
that game is a very very

interesting very nice indication along
those lines the want to highlight It's

a very interesting setting so if you think
about the network that's why marketing and

selling You have to talk with people.

That's what have the ownership
interest right now and

Molly that suppose everyone in this
network has a part it costs so

much more that what I asked for is to
tell their friends about this product.

OK And so we can ask.

Well this there may be some be
a bit of incentivizing it's

not worth it when it's just red flags and
so

I did look can I ask is there
an incentive compatible negatives and

maybe a simple listing costs
brought on from this network so

that they have not set a flight model
put it in terms of the estimates.

So you aren't forced to doesn't
tend to prove that there is a very

nice mechanism for monotonic smartly
maximization just incentive compatible.

So nobody has incentive to lie.

It's what if you force we don't spend
more money to help which is good and

provides confidence
constructive proximity.

You're not going to talk about how far to
the depends on other projects the function

but I want to talk about the politics
of this influence maximisation people

don't have that it's
just a two thousand and

twelve and actually got the best
people report at the conference and

it's a very nice set of data actually
from what did what did he do well.

He ran an experiment and
this mechanicals for

most of your audience but
it's the platform of the last year.

Sort of give people tax and
pay a little bit of money but

small crowd source certain certain times.

And he's actually decided the task
that he ran on is in the character.

He asked the participants.

How many friends you have on Facebook but

how much do you want to be paid in
order to put and advertisement or

selfishness travel agency on your
Facebook page and if you do.

That's very automatically shows up
on the pages of that uses friends

like you really have to suspect that
the some are so actually going to get and

that he has actually told them
that it's more on both participate

in this task you go to
pick some subset and

pay you the cost for the really have to
settle for small reporting the cost but

these are all you are either
actually put I don't think so.

They're actually right so that suits me
just as good looking static as the years

of the past from this paper the exits
the number of friends divides to costs for

some high part of the very
pleasure of many friends and

he will charge more than
a few a few friends.

That's not really to be the case
it's what the system he says but

that's very funny kind of
distributions I can see while

Naslund whatever dollar amounts
that are the disabled by five.

For example and thoughts are simply for
the system usually and

then you run experiments that he
crawled out of the physical network and

got stuck around twelve one
illegals from the previous one and

just use those distributions here.

All right.

What are the sort decides what is the
private cost of the different looks inside

its beauty sleep experiment that is
informed by this by this data and

then if you read various books on our
inner influence propagation schemes right

and similar I try to understand
hold of the who come here with

their algorithms that actually get to
know the true cost but how much really.

It's at

the bottom of this because like

trials.

Yes it was all right but

let's get this stuff right.

Get right.

Some of us are very rare.

OK Just want to

tell you that all of

this and lots

lots of people right.

Sorry sorry sorry sorry I thought there

was a hole there right to try
to get an overview of what's.

At that time time I said it's
like many parts of the world

but I did ask that question to
turn one to automation problem.

Well if you're willing to spend ten
minutes each day which laws are needed in

order to figure out how many more
that is a whole lot of work.

I think it's a process and
you just the machinery.

How in the theory the law writes about it.

I think over time.

One of the postings getting
the link to each other for

in this case of information that's read
the letter of lots of lawyers here

and that's supposed to be one
of those largely by word of next

in this case a lot about this
particular case keep very real.

I don't know about that before all these
other people learn about the story right

and thought given to the F.B.I.
as well which law should be in

order to learn about its biggest kids
reading through their cell blogs.

It's pretty smart.

Actually it's quite similar to this

was never a problem by the both of those
settings you have network and you have

something that spreads you just like
it's going to mean if it's pretty good.

Perhaps it was never read or information
be disseminated through links and

the last sentence was a pretty small
set of nodes that attack this because

I think it's very similar structure in
both settings the fact it's more or

less it is always having to go up and
so he has for example what happens if

you step on your wrist it's first like
a lot of red marks are pretty bad yet

it's not very far to the point of
a good chance to protect him because

kids very early but you could have some of
the above if you read it with a number of

postings links like Peter and other
respects but if you actually optimize.

I don't actually want to talk
about that aspect he said it's

like yes for
several of your postings here.

You start this book just part of why

all the talk about this is
clearly you have to see

the fire the seeds are right there.

Look at some of her words and

you can also look at many
times it is a problem.

Exhaustive search aside hopeless
takes quite a lot of time and

so I told you but that's just back

when it just gets about
seven times because I

did see it's

a very very practical aspect that I don't

want to talk about it's actually kind

of in the future novels from one small
That's my present pleasant start so

far talk about what to take these

kids but oftentimes these are actually
very difficult to answer.

For example about blogs actually very

very severe write about
particular stories but

it on the story to tell you
where you get to see if you can.

To see their bodies different
glorified body start but

not to influence to try
to get these six traces.

He says kids even that were running
time to find the right stuff and

know the number but
you don't see where it's OK.

And actually bar is another thing
about their God You see that people go

to the doctor but you don't see
a factor supernatural said but

you know as well you're sort of traces of
diffusion can be slurred to the body alive

in the first thing and
it's usually of the mind's it's much more

like what you get it is not what's

inside the no network and you get to
see these traces of the fusion as

a number of sites then the second
if you're right that's right.

Maybe they're right about
a particular story and

what I would like to do is give those
observations elected Fergus's I like

that who tends to
influence women to die or

to backward you want to
go further surgery and

you should be able to compete with
smartest artists what explain that.

Well the challenge is what are the
brightest constraints within one of their

cherished Well there's many
little beads explaining it.

So if this is my experience
then there's different

trees different ways of influence
how they put explains things for

example here marketing influence to all
but one of those at least three and so on.

Right.
It's one of the floor right.

But this is another way of
knowing if it's good but

if it's the only
difference that's a treat.

I see only gets about one
second to think about and

what will all sort of once you but
that's that's not OK so.

Well I'm going to ask if you can
come up with a networked small

explains the truth about sparkles that
you that we need to some of you to

describe how such a treat real money
that you can promise to keep a bite

on a promise to model that describes how
likely this is to see a particular traits

given some underlying diffusion treaty.

OK And then what you can do is you
can try to find a sparse crowd.

The next mice the likelihood
of the most likely tree

for all states and dark in Britain.

That's a lot of time.

It's a model objective function and

you can use the radio with musician and
see hear some results and

synthetically That's all right to
recover their legs from truth.

It's a few years but ran aground generates
a whole bunch of different outcomes.

This is using here chronically

grandmother left with
this post a list back and

here is what the potholes its
position will be called Curves R.C.

curves for different I wouldn't say
the scale is that it's a single

baseline dissects you just look at it.

Informativeness in some sense of
every single day ahead of you Dr

pixel talking sense and
sense it's because of

how he knows with his marketed
independently from each other and

it's there that doesn't work too well and
if you actually use this.

I'm going to start to your subjective
function just applied reality then

you get very focused on this one
already graph models another and

graph model the both of you gets
very hard to see really how

you can find usually of more than
one side and this is good all range

of the networks with the from structures
so right part of the terms of the board.

All right let me try to take a step and

especially those stats how targets
that it looked political topics.

What was it was just a short process.

Literally petition marks and
these are a bunch of means

during the last presidential
election right and right.

Time he's the leader of the right but
also just on

the phone with us for a couple of the time
but you get this means you get to see

Ben these the current notes you know
that's the traits that you consider

OK then you can run the other thing is
they are using about twelve million.

I'll use articles from without.

One of the new sources.

This is good for network was
interesting Nancy is that a lot of

that I get very I'm talk to much but you
can for example try to understand a step.

So the community decided for

example see what a cluster of notes
that correspond civil rights or

political but if you see blogs and

blogs more related to its
attainment technically or

you know what's in the OF Can
it to you to be blogs and

so it's very few trying to get so
excited about the structure of

them that a lot of this time there's
a lot of really interesting.

I find discrete optimization problems and
sites and so for networks.

Fortunately many of them are smarter and

there's a lot of interesting and varied
optimization problems that arise from

looking at it was
a constraint on if problems.

It's never talked about.

Yeah it was just an extremely useful
since I think it's really a success story

in just a maybe the last couple years and

that's also interesting constrained
optimization the more complex constraints

the Rices connectedness constraints for
example but I think more sense

for him and he's got a problem
that's also very interesting to

consider for
the first part of the story and

then the second part is to try to get
to talk about whether it's thank you.

Like

our writers but
that said it was just funny.

It was actually very excited to have it.

One of the VERY much
more a lot of good and

something for all your approaches.

Just look at it rather

than just my own little
fusing very very very simple.

It's sort of three words.

I think very very very general structure

very well that's rare.

Right

right right.

And for trying.