>> Kirsten Wiley: Good afternoon and welcome. My... I'm here to introduce and welcome Dr. Sandy Pentland, who...

advertisement
>> Kirsten Wiley: Good afternoon and welcome. My name is Kirsten Wiley, and
I'm here to introduce and welcome Dr. Sandy Pentland, who is visiting us as part
of the Microsoft Research visiting speaker series.
Dr. Pentland is here today to discuss his new book, Honest Signals: How They
Shape Our World. In his book Dr. Pentland writes how a surprising amount of
human behavior can be reliably predicted from unconscious communication
mechanisms, such as a separate communication channel.
Using a specially designed digital sensor worn like an ID badge to monitor and
analyze the back-and-forth patterns of signaling among groups of people, he and
his researchers found that this second channel of communication revolved not
around worlds but around social relations.
Using this network intelligence theory of social signaling we can harness the
intelligence of our social network to become better managers, workers, and
communicators.
Dr. Pentland is a pioneer in the fields of organizational engineering, mobile
information systems and computational social science. He co-directs the Digital
Life Consortium at MIT, a group of more than 20 multinational corporation
exploring new ways to innovate and oversees the Next Billion Network. In 1997
Newsweek named Pentland one of the hundred Americans likely to shape this
century. Please join me in welcoming Dr. Pentland to Microsoft.
[applause].
>> Alex (Sandy) Pentland: Glad to be here. Good to see some sold friends.
And glad people are interested in the stuff we've been up to for the last almost
decade now. So I thought I had just sort of jump right in and see where it goes.
So honest signals is a term of the art in evolutionary biologist and we all know
that animals, social animals communicate by a signaling to each other in order to
coordinate activities. Honest Signals are an evolutionary stable type of signaling
that evolves whenever there's competition for resources and there's a necessity
for bringing individuals together to coordinate their activities.
So an example is the battle between the sexes. You need to bring two specific
individuals together to be able to do anything but there's general competition. All
animals have Honest Signals, bees and things like that have been most studied,
but you also see it in apes. And one of the things that you might want to ask is
what are they in people.
The difficulty in studying them in people has traditionally been getting
measurements. And the other little fact that we have language to communicate
as well as signal mechanisms makes it more complicated. So let's dive in and
ask what are the sort of signals that people might have. So after reading a lot of
the literature, I decided to focus on four which I thought were good candidates for
honest signals. Do we have a -- I guess we don't. Oh, here's something. Oh,
okay. See. Honest signals of confusion.
So four that I focused on because their easy to measurement -- measure and
they're present most places are these. So the first one is actually a signal that's
related to your autonomic nervous system. So this is your fight or flight system.
It's the thing that when you're scared or when you're really interested, you get
jazzed up. So think about little kids when they get interested they bounce
around, right. So they're much more active. Or a dog when a dog is interested.
What does it do? Well, it barks more and jumps around, his tail wags, right?
Higher activity.
And so higher activity of certain sorts, more verbalization, faster accelerations in
your movement, things like that, is something that's highly correlated with your
autonomic nervous system being fired up. But think about this just for an
instance. So I said that you can judge something about this sort of fight or flight
being jazzed up interested stuff not only in humans without words but you can
also estimate these things in species like dogs. You can tell when a dog is
excited, right? How is it that you can do that? You don't have any expressions,
you don't know anything about their internal mental state. They don't have
words. But yet you can tell quite reliably when they're excited about something.
So this is an example of a very old signaling mechanism that occurs in many
social species. Another one is attention, another very old one has to do with your
thalamic and subthalamic areas. How can you tell when somebody is paying
attention to you. Or for that matter, how can you tell when a dog is paying
attention to you? Because they have thalamic nuclei, also.
Well, here's an answer. If I put you in a psychology experiment where I showed
you a light and you had to push a button, it would take you about 400
milliseconds to push that button. Maybe if you were really fast you would get it
down to 300 milliseconds. But when I'm talking to you and we're taking turns, the
gaps between our speaking and gaps between my jumping in when you turn off
as down in the tens of milliseconds or less.
Similarly, if we are interacting with something, your eye motion, your head motion
is very tightly coordinated with mine. Again, down to the sort of millisecond
range. Much faster than you can do without paying attention and without
modelling the other person. So you get this very tight influence measurement
between people when people are paying attention. And actually this is used as a
clinical diagnostic in preverbal infants for language development.
So for almost 30 years now when you wanted to screen kids for language
development problems like four month olds, what you would do is you'd look at
the coupling between infant motion and mother's motion. And if they were tight
like that, then that meant the infant was paying attention to the mother and it was
likely that they would develop normally in terms of language. And if you didn't
see that, there was likely to be a problem.
>>: So when you say coupling, how is it different than [inaudible].
>> Alex (Sandy) Pentland: This is a timing measurement. You model the two
things, the two actors as Markov processes.
>>: It's action reaction.
>> Alex (Sandy) Pentland: It's action reaction, right. Mimicry is one that's more
or less special to people. We have mirror neurons. You may have seen a little
video of a three-hour-old infant in the mother's arms. The mother sticks out its
tongue and the infant sticks out its tongue in response. Think about that. This
baby's had his eyes open for three hours. It can't control its body. It didn't know
anything about the environment and yet it can mimic an internal muscle of its
mother. Maybe the first time it's really seen it have. And yet it does it. So this is
a hard wired system in us. Some of the apes have some mirror neurons. And
what it results is in a mimicry. So when you and I sit across from each other I'll
nod my head, you'll nod your head, I'll go like this, you go like this. People copy
each other. And when they do that and you ask them yards you'll find elevated
judgments about empathy and trust.
So an example of this that is really quite striking is Jeremy Bailins [phonetic] at
Stanford did an experiment where he had a little computerized agent that was
trying to sell you something, and in another part of the experiment had a little
camera that was watching you, and so the also agent would now move its head
the way you moved your head but with a four second delay. So first of all people
didn't detect that, even though it's sort of drop dead stupid mimicry. And second
of all, they were 20 percent more -- 20 percent more bought into the thing that
was being sold which is a behavior that had to do with carrying your identity card.
So the presence of mimicry in this situation even in a computer agent resulted in
much more adoption of a behavior. And you see that across many, many
different studies. Tips for waitresses, negotiations, sales. The presence of
mimicry is highly correlated with empathy an trust and makes these sorts of
interactions go much smoother.
A final one I'll draw your attention to that we looked at is what you might call
consistency of motion. So think of Tiger Woods and the golf club or Barishnokoff
dancing. Experts of a fluidity of movement that's really remarkable. People can
pick it out. They use it as a marker of expertise. So in talking if you're very fluid
in your reduction, people think you know more about the topic than otherwise.
So it's good to practice your talk. Okay. And if you aren't very fluid in it, then
they think you're either conflicted about it or you don't know what the right answer
is so you're open to things.
So for instance, I'll show you some things here in just a moment. One of the
things here that's true about this is this is another sort of standard psychological
measure. So it's been used as a measure of cognitive load. When people are
doing one activity and then you ask them to do another activity, they'll become
irregular in the first activity. So if you're driving and asked to count backwards
from 100, your steering becomes erratic, the entropy of your steering control
goes up.
So this again is another sort of marker of a brain system function, in this case
your motor system and cerebellar system. So all of these actually are things that
provoke reactions in others. So here we have mood contagions. If we're in a
group and I become very enthusiastic, you'll tend to back enthusiastic. If you
become enthusiastic, I tend to become enthusiastic. So there's this infection that
happens as a function of this activity variable. You can understand this as
adaptive behavior. If you think about people go out to attack the mammoth, right,
it's good to get everybody to go and everybody to be enthusiastic so if you all get
sort of geared up before hand you have much more likelihood of success than
otherwise. And incidentally, you see this in all sorts of pack hunting animals,
same type of behavior, they get all jazzed up before the attack.
Attention tells you when people are paying attention to you, obviously, it's good to
know. Consistency. Who are you going to copy. Mimicry seems to be adaptive
because it's important in tacit learning. Anyway, so those are the things that I've
been studying. And the sort of take away from all of that is that you think about
behavioral neuron economics and reading brain state with these big FMRI
machines and stuff like this. Well, that's all well and good, but you can actually
use very simple devices to read qualitative measures of brain state. So this is a
sociometer. It knows where I am, it knows whom I'm facing, it knows about my
body motion, it knows about my tone of voice. It knows -- I think those are the
main things it knows. And it reads those honest signals as I move around.
You can do most of the same things off of a smart phone. The main limitation of
Smartphone is you don't wear it in a fixed location, sometimes you leave it in your
purse, so the data collection is a little irregular. And using these things are
deploying lots of these on whole organizations or all population in the city, you
could begin to x-ray the behavioral function of the city; in other words, how are
people interested in things, what are they paying attention to and so on and so
forth.
So first thing, so this is a little spinoff from the group. I mentioned that activity is
a measure of autonomic nervous system. Well, depression, which is the number
one cause of lost days of work in this country, is often described as a blunting of
the autonomic nervous system. So you ought to be able to measure it, and in
fact, you can, you can do almost in fact as well as the gold standard which the
Hamilton Depression Index. That's a pretty interesting thing.
Here's another one that's more interesting to other people. We worked with
Vertex, which is a large caller center operator representing Tesco, which is a
high touch retailer in the U.K. And we looked at operators in the vertex call
center, and we listened to their tone of voice in the first 90 seconds, and using
that we could predict whether or not the call would be successful by Tesco
standards with just under 90 percent accuracy.
So what was the answer? Well, if the call center operator listened a lot and had
very variable [inaudible], then it was very likely that this would be a successful
call. Pretty good. Actually I have some economic consequence and we're doing
a much large certify study with Bank of America call centers beginning soon.
So what it seems to be is that these honest signals shape your life in lots of
different ways. So examples that I think are interesting, we looked at mock
salary negations. These are mid career people that are negotiating the salary,
ones of senior VP the other person is transferring in. They have to gosh salary
and perks and things. It wasn't a real negotiation but people actually got paid
based on how they came out and they got their grade depending on how it came
out. And the signaling in the first five minutes when they were just sort of setting
up predicted the outcome with about 30, 35 percent of the variance accounted
for. To put that in context, that's bigger than the other sort of factors that people
have studied like motivation, strategy, et cetera, et cetera, et cetera, it's in fact
the largest thing that anybody has ever found.
We looked at sales. I just show you sales. Here's a fun one. We looked at
hiring -- dating rather. So I went to bars where people were doing speed dating,
and so people talk to members of the opposite sex for three minutes and then
they write down whether or not they want to exchange contact information. And
we could predict without knowing anything about the person, without knowing
anything about what was said whether or not they would exchange contact
information with about 75 percent accuracy where the chance rate was about 15
percent. Interestingly the woman signaling is the one that would determine the
outcome and the man signaling had no effect on predicting the outcome.
[laughter].
Another interesting thing actually is that even though the trading of contact
information was supposed to be secret, the men only agreed to trade contact
information when the woman agreed to. So somehow they were able to read the
woman even though the decision to trade contact information was supposed to
be secret, it may have occurred after the interaction. It's interesting.
And this is one that I particularly like. We got a group of mid career people,
stone fellows at MIT who were pitching business plans, the real business plans
for business plan competition and being judged by other people who were cell
phone fellows, so mid career people who had done this a lot of times, and just by
listening to the tone of voice you could predict the rating of the business plan
almost as well as a human could predict how another human would rate the
business plan. So human level performance without knowing anything about the
plan or anything about the person.
So what was going on there? Well, if you don't know too much about an area
and a purported expert comes up to you, and they sound very enthusiastic and
they sound very fluid so they sound like they know what they're talking about,
probably the best strategy is to trust them, right? And that seemed to what
people doing. So pretty amazing. So the next time you get a business plan pitch
or a pitch for a new thing, get it written down, go in the corner, think about it first.
On the other hand, one aspect about this is one of VC's fun things, they like to
visit the corporation and see the buzz. So what that has to do with this is how do
people react to each other, how do they signal among each other? Do they all
get excited, do they all listen to each other, do they sound like they're committed?
So we may be able to rebuzz, which is an important factor in funding decisions,
from this sort of technology.
So let me just go back one slide here, excusing my poor PC burp. So we've
done a lot of different experiments, 2300 hours of experiments with 800 people.
In most of these experiments we find that we can count for about 40 percent of
the variation in decision making. It's not subjective generally, this is hard
decision's making. Did they buy it, did the they exchange information? How
many dollars did they get in it also includes subjective things, also, put I'm going
to focus on the hard addition making. So these are objective outcomes that can
predict 40 percent of the variance on average. It's pretty content across more
than a dozen experiments and some very large scale things which I'll tell you
about in a minute.
Now, 40 percent of the variance may not sound like very much, but that means
on average we can get a about 80 percent accuracy in predicting things and 40
percent of the variance is for instance about the contribution of your genetic
makeup to you. So it's a big factor. And this is all done without knowing anything
about the baseline of the person, the personality, the facts of the matter is just
looking at their behavior, they're non-linguistic behavior over short period of time.
Nor is this affect or emotion. Because what we're doing is we're integrating these
signals over periods of 30 seconds to five minutes. It's like the texture of gesture
rather than the specific gesture. It's much longer than linguistic phenomena.
None of these things have very specific effective or emotional content. So it
really seems to be something that is a signaling mechanism that we have that we
inherited before we had language. So as I said at the very beginning, let me ask
the question again. What did -- what did we do before we had language? So
language and humans is only 50,000, 100,000 years old, according to genetic
estimates.
So not so long ago, when we already had tribes when we could hunt, when we
could defend ourselves, when we could hunt all the large mammals to extinction,
we were able to coordinate ourselves very well. How did we do that without
today's language? Must have been some sort of signaling mechanism.
Evolutionary biology tells us that these honest signals develop regularly and
stably to coordinate groups, to do activities like that. So we probably had them.
These have all the hallmarks of these signaling mechanisms. And then what
would have happened is later we would have evolved language and it would
have come on top of those. And the way evolution works is it co-ops earlier
mechanisms, it doesn't replace them. So you would expect the signaling
mechanisms to be related to the conscious decision making, related to language,
but not completely subsumed by. And I'll show you some more about that.
>>: So I like the way you [inaudible] and things happening so you've got to -- it
looks like there's correlation here, right.
>> Alex (Sandy) Pentland: It is a correlation.
>>: I'm talking about R. And then for the different types of signals did you use
some sort of [inaudible] more bang for the buck in [inaudible].
>> Alex (Sandy) Pentland: Yeah. Right. So same four things across different
situations and you see different proportions. And I'll show you how the
proportions sort of stack up in just a minute, okay.
>>: So in this, how did you [inaudible] linguistic information before you
[inaudible].
>> Alex (Sandy) Pentland: You don't measure it.
>>: [inaudible] I mean [inaudible].
>> Alex (Sandy) Pentland: Yeah.
>>: So [inaudible].
>> Alex (Sandy) Pentland: No. No. So for instance take the negotiation, okay.
So the negotiation happened in any of four different languages in the Greek,
English, Japanese. During the first five minutes is when we were listening.
People were mostly putting their case out on the table. They weren't actually into
the strategy and push and pull. And so what we'd looked at is averaged over five
minutes how much did you talk? That has very distant relationship to linguistic
content, okay. Or averaged over that five minutes, how much mimicry did you
do? Again essentially zero. Very, very distant relationship to linguistic content.
Okay. Pitching division. All right. Everybody had a business plan. Everybody
said all these things. What they said had essentially no predictive value as far as
we could tell. People confuse style and presentation with content almost
completely. The correlation between radions of style and radions of content were
in the high 90s.
>>: Are these the same person you do this kind of [inaudible].
>> Alex (Sandy) Pentland: Same people. Well, I mean, in the same person
you'll find those correlations if that's what you're saying. But we've done this
across people without taking in account individual variations.
>>: Did you make attempts to decouple the content from presentation like in the
example so -- I mean, is there correlation between making good content and
presenting good content?
>> Alex (Sandy) Pentland: In the case of the negotiation the content is fixed
pretty much. I mean people are taking variations on it, but it's pretty much the
same.
>>: As far as the business plans these are the people who had written the
business plans?
>> Alex (Sandy) Pentland: Yes.
>>: [inaudible].
>> Alex (Sandy) Pentland: So, so, so the bottom line here is that, yes, there's
linguistic content, and, yes, that can change your signaling behavior. I can say
something that makes you excited, right?
>>: It's also like learned efficacy, you know, people who are good academically
are more confident that they're going to keep well and [inaudible].
>> Alex (Sandy) Pentland: Absolutely.
>>: So it's a how do you determine which causes the other, does someone
present well because they have a good idea or do they ->> Alex (Sandy) Pentland: So based on what we see but more importantly
based on evolutionary evidence, the answer is something like for instance where
Conomon talked about in his know bell prize lecture. We have two decision
making mechanisms, we have an evolutionary old one which he called the low
road which we think of as intuitive decision making, it's largely unconscious. It's
very heavily influenced by the social factors. And we have the conscious one
which almost everything is designed for these days. And the two interact in very
complicated ways. And particularly oar time they evolve together.
So let me show you some other things here.
>>: I think what people are getting started on is can you explain the concept of
variance. Because they're talking about well maybe I had a really great plan or
evidence a tone of voice or something that's really great that's in that
unexplained variance [inaudible].
>> Alex (Sandy) Pentland: Who? I'm sure that that is in the unexplained
variance, yes.
>>: But if you're watching in effect what the [inaudible].
>> Alex (Sandy) Pentland: I'm watching with the volume off, yes, I am.
>>: And you're saying ->> Alex (Sandy) Pentland: So in all of these things, when people study the
content, all of the content variables account for roughly 40 percent of the
variance. So that means all of the content variables, all of things you taught
about in school are about the same efficacy as these non linguistic signals.
Okay? So that's certainly true in the negotiation, it's true in hiring behavior. It
seems to be true in business plan pitches. Even more so. So the idea of two
systems that are coupled in fairly complicated ways, let me show you sense you
people are asking all these questions -- well, no, let me just continue on the way I
was going and we'll come back to it.
So this is an example where we're getting more people in a room together, so
those are all dyadic interactions. So this is more people. So this is 1300 patients
over a period of a month. 70 nurses and doctors the wore these little badges.
And just looking at the patterns of who talks to who and how much they move
around, how active are they at various points, you can -- and then at the end of
each day the doctors and the nurses would say how productive they thought they
were, how stressed they were, and we also log whenever there is a problem with
the patients and whether there was a delay with the process so things got off
track. And so just by looking at the patterns you could estimate their productivity
answers with high correlation, their stress with very high correlation. So this
gives you accuracy in estimating answers that are in the sort of 90 percent range.
You could detect when there were problems with very, very high accuracy, and
when the delay and the process was occurring the similar sort of accuracy. So
there's very sort of surface level behaviors.
Here's another one might be interested in. So this is a German bank. It has five
departments, managers, development, sales, support, customer service. So they
developed web products, so they put up a big advertising campaign on the web,
in this case for mortgage products.
The blue here is their e-mail patterns of communication between the different
departments. So each frame I put up here is a whole day. It's the average over
a day. E-mail pattern, face to face pattern. Okay. So you see first of all, the
e-mail pattern is relatively static over time. The face-to-face pattern, though,
varies rather dramatically. And this is what we found in many organizations.
E-mail is not a proxy for communication. In fact, in the studies that people have
seen, face-to-face stuff is often where the delicate complex information is, and it
toss not follow the pattern of e-mail. There's a more complicated relationship I
can tell you about. But let's start this again. Has anybody noticed anything funny
about this? They're developing a new product here.
>>: [inaudible] customer service.
>> Alex (Sandy) Pentland: Sales and customer service don't talk. In fact,
nobody talks to customer service.
>>: Right.
>> Alex (Sandy) Pentland: So they're developing this big product, and they have
meetings and everything and nobody bothers to involve customer service. So
what do you think's going to happen when they launch their product? No, it's
coming up in another couple days, 16th, 17th, 18th, 19th, 20th. Okay. Launch.
Didn't work. Big old day meetings with customer service. So that's a really
simple example. But it also turns out that you can tell which people in this
organization feel overloaded. Of will you can't do it from their e-mail. You can't
do it from the pattern of face-to-face communication. But if you look at the
combination, you can do a very good job of predicting people's ratings of their
personal overload. You can also do a very good job of predicting which groups
feel that they're poorly managed and which groups feel that they're well
managed. The poor groups. Again, not from e-mail, not from face-to-face, but
from the combination, the groups that feel that they're poorly managed have a
star like pattern where all the information goes through one person. That person
often feels overloaded. Okay? Sir.
>>: This is comparing e-mail to face-to-face, is there any consideration to some
of the emerging patterns of social networking and how they [inaudible] e-mail or
[inaudible] of the two?
>> Alex (Sandy) Pentland: Two studies done. So you asked with the
relationship between conscious things and the signaling mechanisms. So we did
a study of small groups, four people, resolving a classic social science problem
which is called the survival experiment. You crashed in the northern woods, you
have to take things with you to get back to civilization. Okay? It's sort of used by
for almost 50 years now as a way of looking at the dynamics of groups. And
what social scientists do is they take groups like this and they categorize people
into different social roles and different task roles so a protagonist, a supporter,
neutral attacker, giver of information, orienter of information and discussion,
followers, et cetera. Right? And so this was done by a series of trained people
at a reasonable iterative reliability on it, it's instated on I think a 15 second by 15
second basis. Very laborious thing to do.
And then we looked at the signaling behavior. And it turns out that using just the
signaling behavior you can predict the conscious roles that people are doing with
accuracy nearly as good as the psychologist. So what does that say? Well, that
says that you're signaling behavior is tightly coupled with your social role and
your task role. So when you're in the role of protagonist, you have a particular
signaling behavior.
Eats not just your behavior, though, you have to actually look at the behavior of
all the people in the group to be able to do a really almost human level job of
classification. So that tells you a little bit more about the coupling between these
things. So again, not -- in fact, these are all in Italian if you don't know Italian you
could classify these social roles and these information roles by looking at the
signaling behavior. Should be a little surprising. But from an evolutionary point
of view, it's not surprising at all. If you had signaling mechanisms to coordinate
people before there was language, you would have dominance displays,
supported delays, attackers, all those things would be part of coming to
coordinate a group. And as we developed language what would we do? We
would hijack that mechanism and add conscious argument on top of it. You don't
throw away things that work, you add to them. And that's the sort of thing you'd
see from them.
So there are consequences of these signaling mechanisms an of these different
roles. One of the things that we see is that people who have high influence,
that's this conditional probability relationship I'm talking, so does my talking
influence the pattern of your talking? And there's some other things that go along
with leaders that you'll see in just a second. They tend to be in the center of star
shaped networks. So in small groups you find the patterns of communication go
through these people. In fact, you can identify these people with fairly high
accuracy just from their signaling behavior and confirm that by asking them and
the other people in the group who are the dominant people. So you can classify
them by this behavior, this sort of influence they have on conversation, or by the
patterns of communication.
More surprisingly if you look at a large laboratory, so this is the media lab, so
only 25 people but it's for a period of two weeks, the people who are most
central, have the highest between us centrality are the people the highest
influence values.
So, in other words, the connectors behave differently than other people. In fact,
they're the ones that set the pace in conversations. Interestingly, they have very
variables prosody when they do things and they tend to be more active than
other people. And the correlation between the network measure which people
weren't aware of, they weren't aware where they were and the network, and their
behavior, which again they were not aware of, was .9. So 1700 hours of data, it's
not a little sample, it's not a large number of people, but it's a reasonable number
of days, a lot of hours of data, and you get a number like that. Pretty amazing.
Who are they in well, there's this influence. They tend to be high active. They
have variable prosody. So they sound like their open input, they don't sound like
their experts. From the badges we find that there's a dominance body language,
so the people who are connectors like that, the people who are dominant in a
group stand differently relative to other people than people who are not in that
category. So if I am a dominant person, I'll face you directly. If I'm not a
dominant person, I'll face away when we talk about it. There's an old joke about
geeks, you know, who's the lead alpha geek. The guy who looks at the other
person's shoes. I like that. I'm a geek.
So there's some evidence of that. It's a reasonable strong correlation. I didn't
think we'd see that, but we did. It's also correlated with personality things. So
the people who are the dominant protagonists have high openness scores and
low neuroticism scores. Again, a little bit surprising but a sort of folk knowledge.
Sort of thought you might see that.
These star networks that are caused by having a dominant person are not
always good. So for instance we did an experiment recently where we looked at
people doing 20 questions tests, so they brain stormed for a bit, come up with
possibilities and then they stopped and made decisions about which things to
guess. And in some groups we had a dominant person just by chance, other
groups we didn't. Dominance here could be determined any of three ways which
were, came out to give the same answer. You could ask people who was the
dominant person. You could ask people if they were dominant. Those two
agreed. You could also look for these behavioral characteristics and that agreed
to with good accuracy.
If you had a dominant person in this group, you had the star shape of
communication and your decision making was faster by 20 percent, that didn't
mean it's better, that just means it's faster, but you also found that if you had
them in the brainstorming session, the number of ideas generated dropped by 25
percent. Not good. But.
>>: [inaudible] ideas are good ideas are thrown away.
>> Alex (Sandy) Pentland: Well, the thing is, is there's not enough sort of parallel
conversations might be. Right? There's a -- in some sense, in some way there's
a regulatory effect by having all the lines of communication go through this
dominant person. So this is a fundamental problem for organizations. So high
centrality leader is best for decisions. On the other hand, a low centrality team
configuration is the best for this sort of brainstorming, learning integration. And
organizations have to do both of course. So how do you do it?
So we're turned to a bunch of data that we have. We have something like a
thousand full work days of phone data people doing these things I wouldn't say
phone. That's badge data. And 37 years of continuous phone data for 91
people, tracking patterns of association, and looking at things like creative output,
okay. So it appears, and this has been in the literature before, that what
happens in natural situations and you know take that with a bit of salt, is that
people alternate these two styles of network over time. So in this German bank
data if you looked at the groups that were creative groups, you found that for part
of the day they would have star shaped network, where there was strong
leadership, and part of the day they had no leadership, they were very integrated.
And they would alternate between these. In MediaLab, we have these sort of
daily or yearly surveys where we ask about creative productive output of your
group. How productive do you think your group is. And we found that this
pattern of exploration and integration had an enormously high correlation with
people's judgments about creative output. So it's an interesting thing to sort of
see. Again, this sort of pulsing pattern of exploration and integration. If not a
new pattern, as I make a point in my book. You see exactly this pattern in bee
networks. This is how bees decide how to move the hive. They go out and
explore, they come back and integrate, so they signal each other. If somebody
has found some really good stuff they signal really, really powerfully with high
activity, then more bees go out in that direction next time and that continues until
you hit a tipping point and then the whole hive gets up and moves. Okay? That
same pattern is what we're seeing here. Okay? Now, that's the evidence. When
we studied another group, this is the data systems group, so what they did is
salesmen would call in requirements for a customer and then the people would
configure servers and computers and networks for customer. We studied pattern
and communication there. We found two things. One is is that rich media such
as face-to-face stuff seemed to be more suitable for how to tacit information, sort
of the lore of the tribe, social support, things like that. Abstract media seemed to
be better suited for declarative stuff so that if I text, right.
When we looked at the reach of productivity -- so one of the advantages of this
particular organization is we have hard productivity numbers. We know how long
it takes individuals to configure solutions for the customer and we know how
many times they through up and the customer says no, no, no, no, that's not
what I wanted, I want something different. Okay? So we can measure
productivity and number of mistakes with very good precision and we can
compare that to things like information and transactions. But what we find is
information reach. So this is the number of different people you talk to does
indeed help productivity and you see increments of seven percent or so from
that. So things like Wikis, things like, you know, other sorts of digital media that,
you know, extends your reach do seem like they're likely to help. But the biggest
thing came from the face-to-face network. So now the face to face network is
interesting because in this organization it wasn't even part of the organization.
Not on the work charts, it wasn't managed, people didn't even know it exists. It
was just what you do around the water cooler and management didn't pay any
attention to it. But the cohesiveness of your face-to-face network generated
productivity. Shouldn't say that. People with the highest productivity had the
most cohesive face-to-face networks.
So if you look at the cohesiveness of persons face-to-face network, that
correlated very highly with their productivity and the gain there was bigger than
for the access to additional information. Or in fact, any other factor. So what
does this cohesiveness mean? That means if I go to the water cooler and I talk
to five people that they also talk to each other, okay. So it's closure of the social
group. Are we all people who talk to each other, do we all trade information? So
it's a measure of the closure more closed, which means the more we all talk
among each other as opposed to individual isolated conversations, the higher the
productivity. So what's happening there? Well with, there's sort of two ways to
think about it. One is it's social support. You have a group of people that talk to
each other. So when you have a hard day or something goes bad, there's
people that talk to each other. That's one hypothesis. Certainly part of the
factor. Another thing, though, is there's sort of lore of the tribe. So don't worry
about that, he always says that, it doesn't mean anything, don't pay attention
when they do that. But these attitudes about things that get transmitted
face-to-face, it's sort of tacit learnings about how to do things and the biggest
effect on productivity.
They're also the thing that get managed out the door whenever you would go for
productivity. Right? You look at the data and say we don't have any data about
this face-to-face stuff, but we do know the number of e-mails that people send.
So let's manage it to get the best e-mail, you know, put in a new Wiki, isn't that
great. Okay. We're going to -- put they're missing in this case the most
important factor. So we're taking this finding and extending it to a large call
center beginning in January. So face-to-face behavior matters. It sometimes
matters a lot. That sort of pattern of interaction between people is important.
How can you change it? Let me give you a couple of examples. So this is a
group we did -- exercise we did in Japan. These are groups that are building
Goldberg devices so everybody gets a part of the device. They have to build this
thing. It has to work with all the other groups. So there are half Japanese and
half American groups. Just to make a good working relationship. And of course
the first day you get an American pontificating everybody and the Japanese are
not doing that much accept among themselves. So we gave people at the end of
each day feedback about who they talked to, how interactive they were, and how
much they spoke. So the size of the thing is how much you spoke. The
greenness is how interactive you were. So you like lots of back and forth. And
the thickness of the line is how much you spoke to that person. So we showed
them that -- we didn't give them a theory about what's good, but in fact people
have a background theory about what's good and the background theory is
everybody believes that everybody should be interactive and everybody should
talk. However we saw with decision making that's not a good theory, that's
probably a bad theory, at least for speed of decision making. It might result in
better decisions, we don't know that, but it is good for brainstorming. So a day
one they got this back, day 2. So we just gave it to them and let them talk about
it. And by the end of the week you see a much more interactive pattern with
everybody talking. So this type of feedback is stuff that people are willing to do.
You're basically taking the intuitions we have about face-to-face and social stuff,
quantifying it and putting it on the table. Letting people decide what to do with it.
However they want. And what we need to do of course is have some science
that says this is a good pattern in this situation and this is a bad pattern in this
situation so maybe you should think about that when you go talk to people again
the next day.
Another application of this is distance teams. A big problem, everybody has to
interact with people all over the globe. You don't know the people. People have
teleconferencing stems, telepresence systems, isn't that wonderful. The general
consensus is they don't work really well. One diagnosis is that they enforce this
sort of star configuration of connectivity. There's one person on the floor at any
one time. And what you can't see is you can't see the back channels between
everybody else during that time. I mean, maybe you can hear the keyboard
typing or something like that.
And so our hypothesis was instead of having this, what we would try and do is
put back in the back channels, make all of this much more visible to people. And
so this is a recent paper at CSCW, Tammy Kim was the leader here. So what
we did was we had people wear the badges. The badges then communicated to
the phones and among each other. And it made a little simple display and the
ambient display that was on the table in front of them. So if everyone was being
interactive, you got a nice green bright bulb. If people were not being interactive,
you got a sort of dead little lifeless ball. And the position of the ball represented
how equal the contributions were.
So in this case, this person is pontificating. So everybody claimed they weren't
paying much attention to it. As far as we could tell they in fact weren't paying
much attention to it. But their pattern of conversation changed rather
dramatically. And for distance it made the distance teams look like the
face-to-face teams, particularly in the case when they were dominant individuals.
Roughly what it did was it got have been engaged, have been acting dominant so
there was no more dominance effect. And that was very good for brainstorming
and very bad for decision making. Okay? So an interesting first step we had a
copy of this system here in the building I guess is actually the next building over
and we're continuing on to do some stuff like this.
Moving on, last couple of minutes here I want to talk about stuff that we're doing
now. So the things that I've talked about have been at most a couple hundred
people. So couple hundred people with smart phones or couple hundred people
with these badge systems. And what we're looking to do now is scale this up so
many hundreds of users and a couple thousands of users and then many
thousands of users. And this is in partnership with a city of Boston who was
going to support it from their side and Bank of America, who's supplying the
necessary dollars. And so the things that we're doing is this is a slightly older
experiment, most of the functionality of the badge can be done with a
Smartphone. The main thing is people don't wear smart phones in fixed
locations, they stick them in purses and things. But in a previous experiment
where we gave out roughly a hundred phones to a hundred people and
eventually some of the phones died so we got data, nine months of data for 91
people, we found that we could measure location and proximity and call logs and
things like that quite reasonably. People got phones and some cash in return.
They did surveys. Human subjects approval, all that good stuff, don't worry
about the privacy at this point. And we found by analyzing these proximity
graphs you could do a really bang up job at identifying social structure. So we
could identify self reported close friends with high 90 percent accuracy. We
could identify work group self identified work group members with very high
accuracy. And we could identify who was the boss and who was the subordinate
with very high accuracy.
And this is all using this influence measure. And there was papers about it and
you can look at it. A more interesting one is a current experiment where we've
given 100 window mobile devices to all the kids in a dorm. So it's about 100
people. Some of them were going to drop out. They're programmed the same
way, so we know about the social structure, who hangs with who, who calls who,
where do they do it, so on and so forth. Again, they get a nice free phone for it
and they get paid and they fill out lots of questionnaires. So what we can do is
we can compare things like voting or politics to social association or disease
propagation to how you hang with people, when. And one of the things we did is
we gave them Indy Music service so they can listen to music for free and share it
with their friends. Okay? And we can watch at things like who adopts what sort
of music and who shares music with who. And the answer just as with the
German bang is that if you look at their communication patterns of who calls who,
but also their proximity patterns, who hangs with who and what sort of situations,
you can predict the sharing of media with pretty good accuracy. That means you
can identify the people in the social group that are influencers are the people that
are -- you'd like to get to first if you want to propagate things. And interestingly,
you can't do it from their self reported relationships or from the digital
communications alone. You need to know something about the face-to-face
network to do this. It's a preliminary results. We only have a few weeks of data
there. But seems like it's very consistent with what we've seen before.
This is a little spinoff that came out of my group, sense networks. So this is taxi
cab data. Cell phones of taxi cab drivers. So you know where people are
moving around in the city of San Francisco. The black dots are popular places
like bars and restaurants and so forth. And you can analyze the flow patterns
between these. And what you find is is that the people that go to some places
don't go to other places. So there's really these sort of almost isolate tribes in the
city of San Francisco. So if you see a person go to this bar and this bar, they're
never going to go to that other one, but they're also certainly going to go to this
third one. So their behavioral characteristics across all sorts of things during the
entire day, during the entire week, break into these smaller subgroups. And the
interesting thing is that you can predict a lot of things about their other behavior
from their mobility and their visiting behavior. So you can use this for instance to
stratify customers in ways that you couldn't have otherwise done. So you can
think of this as there may be people have exactly the same demographics that
both lived in the marina, they both make $100,000 a year, but they could have
very different behaviors. And you should treat those people differently.
So many people are interested in this. Bank of America is interested in this as a
credit scoring mechanism. Robert Wood Johnson is interested in this as a health
profiling mechanism. Best Buy is interested in this as a customer relationship
mechanism. Because the evidence is is that which tribe you're a member of has
a greater influence on your decision making than your demographic data. And
almost everything that you do and every other company does is based on
demographics which seems to be the short end of the stick.
Now, I'll come to the big gorilla in just a minute here, but let me just show you a
couple things. So cute thing. Incidentally, all this data is anonymous that we're
showing you in this example. You didn't need to know who these people were.
You just need to know the patterns of flow. And if you download on to your
phone different patterns, you can classify yourself again without giving up your
identity as to what sort of person you are. And you can answer the question
where are people like me. If you do this in several cities you can do things like
where are the people like me in this city that I have never been to before. Sort of
interesting. It shows the power of these sorts of relationships.
You can also do things with this type of data using -- which essentially the same
mathematical mechanisms of traffic prediction. And I know is it [inaudible] and
folks sort of spun out something here recently? I don't know if we're better than
his system or not, but it's the same sort of theory. So this is a couple of years of
data for San Jose or San Jose Costa Rica, not San Jose, California. But we
found we could predict traffic speed within five miles an hour a couple days in
advance. And we could predict traffic jams in the morning with a couple hours of
warning. So what you saw when somebody dug, you know, big trench or
something in the road, early on in the morning five a.m., four a.m., you'd begin to
get these sort of ripples of behavior that were predictive of traffic jams when the
rush really started which is pretty good. Let's see. On the next one. You can do
many other things with these tribes. You can tell what sort of people are walking
by your door. So if you're Starbucks, that's pretty good. Turns out they don't
have anything like this at the moment.
You can predict store and store sales better than conventional techniques and
you can do it before the sales are released. Basically what sort of people are
going in and out of this store? Right? If you know that, you know a lot about
what the sales are. People know about the Pentagon pizza story, so when you
see pizza's going to the Pentagon late at night you should be worried because
something's up. Same is true of behavior in the financial district. When you see
all the young bankers go into work early and stay late, it's time to get into bonds.
So on a broader scale, you can imagine that this is -- has the possibility of really
transforming a lot of the systems that we have. So this is all the Sprint users in
the US one day it's cycling again and again. But now you can see where the
people are. So maps are no longer static, maps have people. Technology
review things this is poised to change the world. Isn't that wonderful? Moving.
Location based data.
I run a group called the Next Billion Network which is looking at these sort of
applications in developing countries. I thought I'd just mention that. The big
thing in all of this that is a show stopper potentially is privacy. I mean what I've
showed you the last little bits I mean some of them are fun, they are experiments,
you sign IRB forms, you get paid, you opt in, no problem, right? It's just like an
experiment. But what happens when you start doing things like the San
Francisco thing I showed you? Well, that's actually not a problem either because
they opted in and it's anonymous. We don't know who the people are, we don't
track them. But you can certainly see where this is going, right? And you know
so I talked to my friends at, you know, Bank of America, and they say well gee
we'd like to know if you go out and party all the time or whether you're a good
do-be and you go to work at eight a.m. and take the kids to the -- you know to the
church on Sunday or whatever. That raises all sorts of big brother things, right?
So one of the very first things is trying to figure out what to do about this sort of
data. Because I think it's actually much more invasive than the sort of document
data that we have. And the thing that I'm proposing and seems to be getting
some traction is that we treat data the way we treat money. So information is
wealth. You wouldn't just let people take your money out of your pocket and do
things with it, right? Don't let them do that with your information either. So the
principles is you own your own data. If I give you my data, it's just like a bank
relationship. I give it to you, you can do things. If I don't like what you're doing,
you have to give the it back to me, and you really have to get rid of it.
And so there needs to be regulatory things around that. And that may sound a
little onerous, but there's a lot of sort of feeling that that's going to have to be the
way that it goes. So the idea that companies can have access to data but rather
than the privacy, I give you the data now you have it for ever, it's really more like
money. I give you things, you can use it only for the stated purposes. If I don't
like it, you have to give it back. So it's moving to an accountability framework
when the data seem to have value and the customer maintenance control. One
of the things that's interesting about this is that just like money, you have to have
taxes. So there's a need for commons. So it turns out by using this data you can
do things like disease surveillance, you can see how it breaks the flu, you can
see traffic patterns. Here's a really interesting thing. This is from all of the
mobile data in the U.K. for a period of one month, so it's four million man persons
or person months worth of -- person years worth of data. And we compared that
with the council data. The councils are the small government organizations that
run neighbors in the U.K. And they calculate something called a deprivation
index, which is predictive of infant mortality and crime rates and things like that.
And it turns out that you can predict the depravation index from the pattern of
mobility and phone calls. Roughly just as there are ghettos which are places that
are in the city but they are not connected to the city, and so they're places where
there are health problems and crime problems but also information ghettos,
which are groups of people who simply don't interact with the rest of society. And
those are the people who have the worst outcomes. So identifying those is a
really interesting thing because you can imagine having greater transparency of
government. I mean, can you imagine having a map that you put up that shows
on a block by block basis, you know, how are the kids doing? How are you doing
financially. How is the government services reaching your community? And
being able to see that on a block-by-block, week-by-week basis. Being able to
go to the mayor of a town and say Mr. Mayor you said that new power plant
wasn't going to impact health, but look at this map. You can see downwind of the
power plant kids are getting sick. Let's talk about that. Or your new jobs
program, that's really wonderful. But you know, I don't see any people working
more hours, I don't see more people working. Let's talk about that.
So this notion of bringing accountability and transparency to government is
possible with this sort of data. So I don't think we want to lose that. So that's it.
These are my guys. They are doing all this. Tammy Kim. Danielle Olgene
[phonetic] designed the physical thing. Ben Weaver [phonetic] who did a lot of
the stuff with organizations. Wen Dong [phonetic] does a lot of mathematical
analysis. And some of the other folks.
>> Kirsten Wiley: Okay.
[applause].
>> Alex (Sandy) Pentland: Good.
>>: So your device track what kind of signals, the GPS signal, also [inaudible]
analysis on [inaudible].
>> Alex (Sandy) Pentland: This one has -- uses signal strength to track location
not GPS. It has accelerometry, three degrees of freedom. It has infrared to see
other devices that you're facing. And it has some low power DSP to be able to
do tone of voice analysis.
>>: [inaudible]. And then all the people participating [inaudible].
>> Alex (Sandy) Pentland: And in Smartphone experiments you program the
smart phones to do as much of that as you can.
>>: [inaudible] cell phone signal or [inaudible].
>> Alex (Sandy) Pentland: These are sort of was it a chip con 2.4 gigahertz low
bandwidth radio. So you just stick a couple of them in a building and you can
see where people are in the building.
>>: So the [inaudible].
>> Alex (Sandy) Pentland: You need to have a few of these things. And they
plug into USB ports. So you ->>: [inaudible].
>> Alex (Sandy) Pentland: It depends if you want the geometry. I mean usually
we don't actually carry about the geometry, we care that people are together.
>>: Right. Okay. [inaudible].
>> Alex (Sandy) Pentland: Right. So that's one way to do it, and we also have
Bluetooth scans so you can ask what other Bluetooth devices are in range. You
can talk to your cell phone with Bluetooth if you want to. Stuff like that.
>>: [inaudible].
>> Alex (Sandy) Pentland: So there's two platforms. There's the sociometer that
we use because it gives us cleaner, more complete data. And then we program
smart phones.
>>: Oh, I see.
>> Alex (Sandy) Pentland: Okay. And the smart phones give you many of the
same signals but not all the same signals.
>>: If you want to put [inaudible].
>> Alex (Sandy) Pentland: Well, we've done cameras and audio together, and
the answer seems to be that as a sort of generic thing they're about the same.
Okay. Now, that didn't include facial expression because we haven't done that. I
guess about the relationship is remember what we're doing is we're averaging
things over time period. So we're saying over the last minute how active were
you. Okay? Facial stuff is usually much more quickly. So you can do things like
how many facial expressions were there. That's fine. But that turns out to be
very redundant with hand gestures and very redundant with vocal activity. On
the other hand there's this fine grade stuff that we call expression on or fact,
which you do see more in the face. So that's another dimension that we haven't
been looking at.
>>: When you have the [inaudible] you don't [inaudible].
>> Alex (Sandy) Pentland: You have body motion. You know if the people are
walking, sitting, gesticulating, nodding their head.
>>: [inaudible].
>> Alex (Sandy) Pentland: Yes, three degree of freedom accelerometer. But
what you don't know is you couldn't do sign language, for instance, or finger
spelling or something. Uh-huh?
>>: The title on the signals implies some level of inability to manipulate. Have
you experimented at all with for example with the attempting to persuade on
business plans, for example, taking an inherently bad business plan and having
someone try to pitch the business plan with the right kinds of tones and to see if
that actually would take advantage of the system or is it -- is it so subconscious
that it actually didn't ->> Alex (Sandy) Pentland: So there's a couple different things you can do. We
did a poker experiment, where it's very much the people's advantages not to
show the signaling. And some of these are traditional what are called tells in
poker because they tell you about things. It turns out you can still tell when
people are bluffing and when they have high risk hands with high good accuracy.
>>: With the [inaudible].
>> Alex (Sandy) Pentland: So the idea there is that even when there's a strong
monetary advantage and you're very, very experienced it's hard to do a good job
of expressing these things. Another data point is we tried to teach people to
behave differently. Just I'll say be more active, right, turns out people have great
difficulty in doing that and tying their shoes at the same time. The moment you
start thinking about something, your ability to focus on this stuff goes out the
window. It's just too overloaded.
The way you can do it is something like method acting. So if I say you're going to
pretend to be the hard ass boss, right, think about what a hard ass boss is, how
they act, things like that. Your signaling will change, also. So if you adopt a
social role in that sort of method acting way, your signaling will change with it.
Now, people say well gee, well now I can manipulate people, right? Well, not so
fast. So for instance there's training programs for doctors that are like method
acting because it's been shown that doctors that are empathetic get sued less
often than doctors who aren't empathetic, so what they do is they put them
through a method acting thing to appear pare more empathetic, right? Typical
insurance scam, right? However, at the end of it what you find is their judgments
about things changed, and in fact they are more empathetic than they were. So
that's true of most of these signals. For instance if I act more excited, I actually
become more excited. My risk judgments change, my decisions change. Yours
will also if we're interacting. Same thing with the mimicry. If I real start mimicking
you, you'll start to feel more trustworthy and, you know, trusting and you'll start
mimicking me back. So these things actually do influence each other.
A model I like, you're probably familiar with predictably irrational and some of this
behavior stuff. So we've these different roles we play, different states, different
social states that we're in. When we're in these different states, we have
different parameters for decision making, for risk taking, and things like that. And
along with those things, those roles, your signaling changes. So it's really a
sense of who you're -- what person you're being, what stereo type you're being,
what social role you're being. Changes both your conscious and your
unconscious behavior.
So to think about the business plan pitch, and this isn't the perfect experiment,
let's take a business plan that's sort of silly and take two people that are very,
very similar. And one person pitches it but they think it's silly and they are not
very practiced at it. And then the other person who is really enthusiastic about it
and is very, very practiced at it pitches it, which one do you think is going to be
rated better? Same words, same plan, let's make them visually almost identical
people. But one person sounds enthusiastic, who is very fluent in their
production and they're going to be rated much higher.
>>: You're saying it's reenforcing effect that ->> Alex (Sandy) Pentland: Reenforcing effect. And of course it's complicated
because, you know, as you pointed out, there's interaction between these
systems, typically over time.
>>: [inaudible].
>> Alex (Sandy) Pentland: It's hard to isolate, that's right. So you say causal.
Well, you know, causal's complicated. If I get excited, it changes my decision
making, which means my conscience decision making changes and it's not a
trivial effect, right? Back and forth, back and forth, in fact there's some people,
like they talked to Gazzaniga who did the split brain stuff. He would say that
most of your decision making is not conscious, it's done unconsciously, and then
you rationalize it in your conscience thinking. So that's an extreme view. I don't
quite buy that extreme view. You obviously don't either. But I'll tell you, it's a
mainstream view and there's a lot of evidence for it. Right? Let me give you
another little piece of evidence. So the bit of series of experiments recently that
show that if you think about complex problems your far worse at performance
than if you trust your intuition. Even more interesting there's about a 20 percent
advantage for sleeping on things. Quite literally for recognizing patterns, making
good decisions, things like that, if you literally sleep on it and then make that sort
of gut decision, the quantities are like 20 percent better at all of those tasks. So
I'm not quite sure what all that means but it argues that just as Conomon was
suggesting that there are two decision making mechanisms, two types of
communication mechanisms also in our minds, one's that conscience that we're
aware of, it's very good at simple problems and fairly clean problems and begins
to break down when things get messy and complicated, so that's the whole
Conomon Traverski [phonetic] rationality stuff. And there's another one that is an
intuitive system that is much more socially embedded, much less consciously
accessible. It's good at tradeoffs. But of course it can be misled in all sorts of
interesting ways, too.
So the take home may be for you guys, right, it's where you work is that almost
all the systems we have, almost all the IT systems we have are designed for one
system and not the other system. In fact, they're designed without thinking about
the other system even though there's a lot of evidence, not just my evidence, but
a lot of evidence that this other system is really important. Your decision making
is as strongly influenced by your social context and what your friends do is any
arguments that you want. And if you want evidence of that, just look at a
cigarette pack. Right? It says this will kill you. Does it stop people? No. On the
other hand -- you know, come on, how black and white can you be? And then on
the other hand if your friends all stop smoking, what are you going to do? Very
high likelihood you're going to stop smoking.
>>: [inaudible] does start to mimic these sorts of interactions, is there any place
where ->> Alex (Sandy) Pentland: Short answer is no. Longer answer is there are some
CFCW things that try to get a broader view of what all is happening. But they
tend to be things that are a little bit like the little green ball I showed, they have
little icons of people and the little icons sort of vibrate or get bigger depending on
not the text that the person is doing but some other activity, you know, like how
much they're moving around or something like that, which can be camera based.
And there's some other things where you know, you have these sort of -- had an
interesting description recently, Cisco has this telepresence system and I was
talking to one of the guys that was in charge of that whole system, and he was
describing the situation where they had John Chambers in the center and then
they had set up these telepresence things all around him because he was going
to tell the whole company that the sky had fallen or something, I don't know what
it was. But the result was is that the people in Bangalore could see through the
screens to see the people in London. And they could also see the people in Palo
Alto. So all of the back channels were suddenly there again. So when John said
something, you could see the people in Bangalore went, you know, right? And
he said that was a really different experience than the normal sort of
telepresence thing where you can't see all those little back channels. And that's
just, you know, one more person's story. But I think there were other things
maybe like that.
>>: An good example that was during the debates, the presidential debates
when they had the blog back channels and the multithreading that was actually
happening and also the twittering analysis that was being done while the debates
were going and how the activity was changing.
>> Alex (Sandy) Pentland: Yes. One of the things I'd like to do actually is take
that type of thing and add to it things like how nervous is the person, how excited
is the person acting, because you know, there's always this phenomena of
somebody wants to jump in, right? So why not make things sort of bigger and
brighter if they're doing that or, you know, if people -- here's one of the things
we've seen. If in a group of people somebody says something and everybody
goes they do it synchronously, they really do. And what that means is that's an
important thing. It doesn't mean they agree with it. It means it's something to
think about.
So you get these group activities that are often not linguistic activities, they're
nodding heads or something like that or mmmm that type of thing. And the fact
that the group did that means that there's something about this thing that
happened, this linguistic thing that happened that's important to the group. And I
think annotating things by those sorts of group decisions would be a pretty
interesting thing to do. Yeah?
>>: Another related thing is the concept or the phenomenon, the [inaudible].
Have you done anything, have you self related to any of --
>> Alex (Sandy) Pentland: Yes. So the evidence is both [inaudible] and laughter
are similar sort of mechanisms. They're prelinguistic. They interrupt the
linguistic channel in a way that argues that they're genetically older. They're very
strongly related to dominance. And other things I'm not a really expert on the it. I
didn't pick yawns and I didn't pick laughter because it proved hard to build a
laughter detector. And I didn't think there was enough laughter to justify the effort
and get another thing to measure. Yawns are another thing. Of course they're
important, but they don't happen all that frequently. So I wanted things that were
sort of, you know, I would have a good chance of getting signal all the time. And
there are other ones, too, like we were just talking at lunch about body language.
So if you lean forward, do you lean back. You saw in the stuff that we were
doing that there's postures relative to you. Do I face you or do I face away from
you? So those things seem to be signaling mechanisms also. And you know,
we haven't taken use of those but you know they would be easy to incorporate.
>> Kirsten Wiley: [inaudible] questions.
>> Alex (Sandy) Pentland: One more question.
>>: Is the data field any differences between now men and women communicate
both in the friend networks or face-to-face or in a business setting?
>> Alex (Sandy) Pentland: Yeah. There are some things. Let's see. Well, one
of the fundamental things is one of the early ones we did is we found that in our
laboratory men and women talk the same amount, but women talk to much
shorter bits and men talk to much longer bits. In some of our experiments this
were some interesting gender differences. But I'd have to go back to tell you
what the truth is without focusing on them. Mainly dating is one where there's a
control locust issue. I've tried not to folks on linguistics or gender the or
background variables because I want to make an argument basically that says
look, here's this signal that independent of anything else has this productive
value or predictive value.
>> Kirsten Wiley: Thank you.
>> Alex (Sandy) Pentland: Sure.
[applause]
Download