0018-9162/05/$20.00 © 2005 IEEE March 2005 33

C O V E R F E A T U R E

P u b l i s h e d b y t h e I E E E C o m p u t e r S o c i e t y

Socially AwareComputation andCommunication

W ouldn’t it be wonderful if peoplecould work together more smoothlyand productively? Imagine a worldin which it is normal to openly speakyour concerns and to have a fair and

honest group discussion, in which people areenthusiastic about carrying through group deci-sions in a transparent and comprehensive way.Given the variety and frequency of jokes about badmeetings, and indeed about failed communicationin general, such meetings and enthusiasm seem des-tined to remain wishful thinking.

Although developers of communication-supporttools have certainly tried to create products thatsupport group thinking, they usually do so withoutadequately accounting for social context, so that alltoo often these systems are jarring and even down-right rude. In fact, most people would agree thattoday’s communication technology seems to be atwar with human society. Pagers buzz, cell phonesinterrupt, and e-mail begs for attention until wehave to pause and wonder if we are being assimi-lated into some sort of unhappy Borg Collective.

Technologists have responded with interfaces thatwink at us and call us by name, filters that attemptto shield us from the digital onslaught, and smartdevices that organize our lives by gossiping behindour backs. The result usually feels as if the intent isto keep us isolated, wandering like a clueless extrain a computer-controlled game.

These solutions, while well-meaning, ultimatelyfail because they ignore the core problem: Com-puters are socially ignorant. Researchers seem to

have forgotten that people are social animals andthat their roles in human organizations define thequality of their lives. Technology must account forthis by recognizing that communication is alwayssocially situated and that discussions are not justwords but part of a larger social dialogue.

This web of social interaction forms a sort of col-lective intelligence; it is the unspoken shared under-standing that enforces the dominance hierarchy andpasses judgment about whether your proposal fitswith “the way things are done around here.”

Successful human communicators acknowledgethis collective intelligence and work with it; digitalcommunications must begin to do the same bybuilding tools that can accurately quantify socialcontext and teach computers about successful socialbehavior.

At MIT, our research group is taking first stepstoward quantifying social context in human com-munication. We have developed three sociallyaware platforms that objectively measure severalaspects of social context, including nonlinguisticsocial signals measured by analyzing the person’stone of voice, facial movement, or gesture.1

We have found nonlinguistic social signals to beparticularly powerful for analyzing and predictinghuman behavior, sometimes exceeding even experthuman capabilities. These tools measure social con-text, which lets the communications system supportsocial and organizational roles instead of viewingthe individual as an isolated entity. Sample applica-tions include automatically patching people intosocially important conversations, instigating con-

By building machines that understand social signaling and social context,technologists can dramatically improve collective decision making andhelp keep remote users in the loop.

Alex (Sandy)PentlandMassachusetts Institute of Technology

34 Computer

versations among people to build a more solidsocial network, and reinforcing family ties.

SOCIAL SIGNALSPsychologists have firmly established that

social signals are a powerful determinant ofhuman behavior and speculate that theymight have evolved as a way to establish hier-archy and group cohesion.2,3

Most culture-specific social communicationsare conscious, but other social signals functionas a subconscious collective discussion aboutrelationships, resources, risks, and rewards. In

essence, they become a subconscious “social mind”that interacts with the conscious individual mind. Inmany situations the nonlinguistic signals that serveas the basis for this collective social discussion arejust as important as conscious content in determin-ing human behavior.2-5

A mental partnershipImagine a tribe on the African veldt. Each day

the adults gather and hunt, and in the evening theyreturn to sit around a central clearing where theyrecount the day’s events and observations and dis-cuss what to do tomorrow.

During the discussion, social signals, such asbody posture and tone of voice, reflect the powerhierarchy as well as individual desires. Each bit ofnew information comes with some collective socialsignaling that clearly communicates to each indi-vidual what the group thinks about that news oridea. By the discussion’s end, the group has mademany collective decisions, and the iron hand ofsocial pressure will enforce the required individualbehaviors.

Dominance displays have since given way tooffice politics, but the mechanism and result haven’tchanged much. The collective mind still uses socialsignals to guide individual behavior.

What are they?Body language, facial expression, and tone of

voice are some of the nonlinguistic signals thatunderpin this mental partnership. You might seesomeone taking charge of a conversation, for exam-ple, or hear a person setting the conversationaltone—skills often associated with higher social sta-tus or leadership.

Others seem more adept at establishing a friendlyinteraction, which indicates skill at social connec-tion, a trait many successful salespeople exhibit.5

Prosodic style—also called tone of voice, roughlythe way people vary pitch and volume in speak-

ing—is perhaps the most powerful channel for thesenonlinguistic social signals because it is the leastsubject to conscious control.3

Social psychologists have found social signals tobe extremely powerful in predicting human behav-ior across a wide range of school, business, gov-ernment, and family situations. With only a fewminutes of observation, an expert psychologist canregularly predict behavioral outcome with about70 percent accuracy.3

Amazingly, observing such thin slices of behav-ior can accurately predict important life events—divorce, student performance, and criminalconviction—even though these events might notoccur until months, or sometimes years, later.

PREDICTING SOCIAL OUTCOMESFollowing the social psychologists’ example, a

test for our ability to automatically measure socialsignals should also be a test of our ability to pre-dict outcomes from observing a “thin slice” ofhuman interactions. Could we predict humanbehavior without listening to words or knowingabout the people involved?

Our research group has built a computer systemthat objectively measures a set of nonlinguisticsocial signals, such as engagement, mirroring,activity level, and stress, by looking at tone of voiceover one-minute periods.1 Unlike most otherresearchers, our goal was to measure signals ofspeaker attitude rather than trying to puzzle outthe speaker’s instantaneous internal state. Con-sequently, we treated prosody and gesture as alonger-term motion texture rather than focusingon individual motions or accents. Although peo-ple are largely unconscious of this type of behav-ior, other researchers2,3,6,7 have shown that similarmeasurements are predictive of infant languagedevelopment, empathy judgments, attitude, andeven personality development in children.

Using our social perception machine, we couldlisten in to the social signals within conversations,while ignoring the words themselves. We found thatafter a few minutes of listening, we could predict

• who would exchange business cards at a meet-ing;

• which couples would exchange phone num-bers at a bar;

• who would come out ahead in a negotiation;• who was a connector within a workgroup;

and• a range of subjective judgments, including

whether or not a person felt a negotiation was

Prosodic style isthe most revealing

channel for social signals

because it is theleast subject to

conscious control.

honest and fair or a conversation was inter-esting.

After excluding cases in which we didn’t haveenough signal to make a decision, our predictionaccuracy averaged almost 90 percent. The“Measuring Prediction Accuracy” sidebar tells howwe calculated accuracy.

Achieving this level of accuracy is pretty amaz-ing, especially given that experiments using humanjudges have typically shown considerably less accu-racy. Moreover, the decisions we examined areamong the most important in life: finding a mate,getting a job, negotiating a salary, and finding aplace in a social network. These are activities forwhich humans prepare intellectually and strategi-cally for decades.

What is surprising is that the largely subcon-scious social signaling that occurs at the start of theinteraction appears to be more predictive thaneither the contextual facts (attractiveness and expe-rience) or the linguistic structure (strategy chosen,arguments employed, and so on).

QUANTIFYING SOCIAL SIGNALSThe machine understanding community has

studied human communication on many scales—phonemes, words, phrases, and dialogs, for exam-ple—and researchers have analyzed both semanticand prosodic structures. However, the sort oflonger-term, multiutterance structure associatedwith signaling social attitude (interested, attracted,confrontational, friendly, and so on) has receivedlittle attention.

To quantify these social signals, we began byreading the voice analysis and social science litera-ture and eventually developed texture-like measuresfor four types of social signaling: activity level,engagement, stress, and mirroring.1

By using these measurements to tap into thesocial signaling in face-to-face discussions, we couldidentify learned statistical regularities to anticipateoutcomes. In addition to vocal measures of socialsignaling, facial and hand gesture equivalents to theaudio features are being developed, and experi-ments using these visual features are under way.

Activity levelActivity level—the simplest measure—is how

much you participate in the conversation. For theactivity-level measure, we use a two-level hiddenMarkov model (HMM) to segment the speechstream of each person into voiced and nonvoicedsegments and then group the voiced segments into

speaking and nonspeaking. We then measure con-versational activity level by the percentage of speak-ing time.

EngagementIn broad terms, engagement is how involved a

person is in the current interaction. Is he drivingthe conversation? Is she setting the tone?

We measure engagement by the influence eachperson’s pattern of speaking versus not speakinghas on the other person’s pattern. Essentially, it isthe measure of who drives the pattern of conver-sational turn taking. When two people are inter-acting, their individual turn-taking dynamicsinfluence each other, which we can model as aMarkov process.6

March 2005 35

Measuring Prediction Accuracy

We calculated a linear predictor of outcome by a cross-validated lin-ear regression between the four audio social signaling features (describedelsewhere) and behavioral outcome. We then compared this predictorto the actual behavioral outcome. The histogram in Figure A shows atypical case, in which the data is “Would you like to work with this per-son or not?”

In a typical case with a three-class linear decision (yes, not enoughinformation, no) the yes/no accuracy is almost 90 percent. Accuracy istypically around 80 percent with a two-class linear decision rule, wherewe make a decision for every case. More generally, linear predictorsbased on the measured social signals typically have a correlation of 0.65,ranging from around 0.40 to as much as 0.90.

Most experiments involved around 90 participants, typically 25 to35 years old, with one-third being female. Recent papers and technicalnotes about these experiments are available at http://hd.media.mit.edu.

Figure A. Histogram for the data on “Would you like to work with this personor not?” The blue bars are “no” answers, the red bars are “yes.” Greaterpredictor values mean that a “yes” is more likely. Placing the yes/no boundary at 1.4 yields a 72 percent decision accuracy.

0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6Predictor

Freq

uenc

y of

dat

a

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

5.0

0

36 Computer

By quantifying the influence each participant hason the other, we obtain a measure of their engage-ment. To measure these influences, we use anHMM to model their individual turn taking andmeasure the coupling of these two dynamic systemsto estimate the influence each has on the other’sturn-taking dynamics.8

Our method is similar to the classic work ofJoseph Jaffe and colleagues,6 who found thatengagement between infant-mother dyads is pre-dictive of language development. Our formulationgeneralizes those parameters so that we can calcu-late the direction of influence and analyze conver-sations involving many participants.

StressStress is the variation in prosodic emphasis. For

each voiced segment we extract the mean pitch (fre-quency of the fundamental format) and the spectralentropy. Averaging over longer periods providesestimates of the mean-scaled standard deviation ofthe format frequency and spectral entropy (roughly,variation in the base frequency and frequencyspread).

The sum of these standard deviations becomes ameasure of speaker stress; such stress can be eitherpurposeful (prosodic emphasis) or unintentional(caused by discomfort). Other research has usedsimilar measures of vocal stress to detect deceptionand to predict the development of personality traitssuch as extroversion in very young children.

MirroringMirroring occurs when one participant subcon-

sciously copies another participant’s prosody andgesture. Considered a signal of empathy, mirroringcan positively influence the outcome of a negotia-tion and other interpersonal interactions.7

In our experiments, the distribution of utter-ance length is often bimodal. Sentences and sen-tence fragments typically occur at several-secondand longer time scales. At time scales less than onesecond, the utterances include both short inter-jections (“Uh-huh.”) and back-and-forth ex-changes typically consisting of single words(“OK?” “OK!” “Done?” “Yup.”). The frequencyof these short exchanges is our measure of mir-roring behavior.

INSIDE A SOCIALLY AWARE SYSTEMWe have incorporated these social signaling mea-

surements into the development of three sociallyaware communications systems. Figures 1 through3 show these systems in use. The Laibowitz and

Figure 1. Badge-like platform. Built on the Laibowitz andParadiso Uberbadge, this system allows social contextsensing by infrared, audio, and motion so that wearerscan automatically bookmark interesting people anddemonstrations, and it displays messages designed tobuild social networks.

Figure 3. The Serendipity system. Built on the Nokia 6600 phone, the system senses the proximity of otherpeople and compares their interests to make sociallyappropriate introductions.

Figure 2. The GroupMedia system. The system, builtaround a Sharp Zaurus PDA, measures attraction signaling in dating and other social events. The system can also provide feedback to users and patchremote users in to interesting or socially important conversations. The interface in the image, which is stillin the experimental stage, is a split screen of messagesand biosignals from another user.

Paradiso Uberbadge is a badge-like platform,9

GroupMedia10 is based on the Sharp Zaurus PDA,and Serendipity11 is based on the Nokia 6600mobile telephone.

In each system, the basic element of social con-text is the identity of people in the user’s immedi-ate presence. The systems use several methods todetermine identity, including Bluetooth-based prox-imity detection, infrared (IR) or radio-frequency(RF) tags, and vocal analysis.

To this basic context, it is possible to add audiofeature analysis, sensors for head and body move-ment, and even biosignals, such as galvanic skinresponse (GSR). These sensing capabilities providea quantitative measure of social context for theuser’s immediate, face-to-face situation. The resultis a lightweight, unobtrusive, wearable system thatcan identify face-to-face interactions, capture col-lective social information, extract meaningfulgroup descriptors, and transmit the group contextto remote group members.

When the system detects a face-to-face interac-tion, defined as the combination of proximity andconversational turn taking, it specifies a group con-text that consists of the participants’ identities, thefour social signals, and the compressed audio (andpossibly video) information stream.

The system then creates a social gateway thatcontains the group context information and letspreapproved members of the social or work groupaccess the ongoing conversation and group contextinformation. The social gateway uses real-timemachine learning methods to identify relevantgroup context changes. A distance-separated usercan then access these changes.

A NEW LEVEL OF COMMUNICATIONEnabling machines to know social context will

enhance many forms of socially aware communi-cation. A simple use of social context is to providepeople with feedback on their own interactions.Did you sound forceful during a negotiation? Didyou sound interested when you were talking toyour spouse? Did you sound like a good team mem-ber during the teleconference? Such feedback canpotentially head off many unnecessary problems,as the “The Face of Socially Aware Communi-cation” sidebar describes.

The same sort of analysis can also be useful forrobots and voice interfaces. Although word selec-tion and dialog strategy are important to a suc-cessful human-machine interaction, our experi-ments and those of others show that social signal-ing could be even more important.

What was that name?An obvious use of social context is to help build

social networks. At some time, nearly everyone hasmet an interesting person and then has lost that per-son’s business card or forgotten that person’s name.On the basis of an audio analysis and observationsof body motion, our Uberbadge-based system9 cankeep track of all interactions during which youseem interested in the other person and e-mail youthe names and particulars of those individuals atthe end of the day.

Building social capitalSocial capital is the ability to leverage your social

network by knowing who knows what and know-ing to whom you should speak to get things done.It is perhaps the central social skill for any entre-preneurial effort, yet many people find it difficult.We are therefore building systems that can help aperson build social capital.

One example is the Serendipity11 system, whichis implemented on Bluetooth-enabled mobilephones and built on BlueAware, an application thatscans for other Bluetooth devices in the user’s prox-imity. When Serendipity discovers a new devicenearby, it automatically sends a message to a socialgateway server with the discovered device’s ID. If it

March 2005 37

The Face of Socially Aware Communication

Four applications are being considered for commercial application:• Mood Ring (aka “jerk-o-meter”). Women often complain that men

don’t pay attention to them when they talk on the phone. TheMood Ring is a cell phone application that monitors conversationsbetween a husband and wife and alerts the husband with a specialring tone if he is sounding inattentive or uninterested.

• Comfort Connection. What most of us miss when dealing with afinancial institution is a friendly, trustworthy human to talk to.Comfort Connection classifies your preferred style of interactionduring an initial interview, and then hooks you up with a servicerepresentative with whom you will feel comfortable working.

• Personal Trainer. One of the problems with the subconscious natureof social signals is that we are often unaware of how we sound toothers. Consequently we often fail to put our best foot forward,most commonly when we are confused or stressed—exactly whenit matters most. The Personal Trainer runs on a mobile telephoneapplication and provides feedback at the end of each telephone callabout how you sounded: aggressive, friendly, interested, firm, orcooperative. This feedback is valuable in helping you learn to pre-sent yourself in the manner you intend.

• Winning Combination. Businesses depend on buying low and sell-ing high, and in most businesses this means having purchase agentsand sales agents that come out ahead of the competition. The dif-ficulty is that the style of a particular agent is not optimal for allclients—for certain pairings the company agent will tend to comeout ahead, for other pairings that person will tend to lose. WinningCombination classifies the speaking style of each purchase or salesagent, and makes sure that the agent is paired with the right client.

38 Computer

finds a match, it sends a customized picturemessage to each user, introducing them to oneanother.

The real power of this system is that it canbe used to create, verify, and better charac-terize relationships in online social networksystems, such as Friendster or Orkut.

If two people hang out after work, they areprobably social friends. If they meet only atwork or not at all, they are likely to have avery different relationship. The system canrefine the relationship characterization byanalyzing the social signaling that occurs dur-ing phone calls between the two people. The

phone extracts the social signaling features as abackground process so that it can provide feedbackto the user about how that person sounded and tobuild a profile of the interactions the user had withthe other person.

Staying in the loopA major problem with distributed workgroups

is keeping yourself in the loop. Socially mediatedcommunications, such as GroupMedia,10 can helpwith this problem by patching people into impor-tant conversations. When it detects a potentiallyinteresting conversation, the system notifies a dis-tant group member. Whether or not a certain mem-ber receives notification depends on measuredinterest levels, direction of information flow, andgroup membership.

A distant group member who receives a notifica-tion has several options. These options include sub-scribing to the information and receiving the rawaudio signal plus annotations of the social context,receiving a notification from the system only in caseof especially interesting comments, or storing theaudio signal with social annotations for later review.

Suppose, for example, most of your workgrouphas gathered, the information flow is from the boss,and the interest level is high. You might be wise topatch into the audio and track the measured levelof group interest for each participant’s comments.The group context information and the linking-innotification that the system gateway provides canincrease both the group cohesion and your under-standing of the raw audio.

The same framework could also enhance thesocial life of close friends. Suppose two or three ofyour closest friends have discovered an amazingband at a bar and are having a great time. The sys-tem could detect the situation and, given appro-priate prior authorization, automatically send youan invitation to join your friends. Although such a

system wouldn’t be to everyone’s taste, this ideagenerally gets a thumbs-up from college under-graduates.

Group dynamicsSocial scientists have carefully studied how

groups of people make decisions and the role ofsocial context in that process. Unfortunately, theyhave found that socially mediated decision makinghas some serious problems, including group polar-ization, group think, and several other types of irra-tional behaviors that consistently undermine groupdecision making.2,4 Improving group functionrequires the ability to monitor the social commu-nication and provide real-time intervention.Human experts—facilitators or moderators—cando that effectively, but to date machines have beenblind to the social signals that are such an impor-tant part of a human group’s function.

The challenge, then, is how to make a computerrecognize social signaling patterns. In salary nego-tiations, for example, we found that lower-statusindividuals do better when showing more mirror-ing, which communicates that they are team play-ers. In a potential dating situation, the key variablewas the female’s activity level, which indicatedinterest. By knowing that certain signaling patternsreliably lead to these desired states, the computercan begin to gently guide the conversation to ahappy ending by providing timely feedback.

Similarly, the ability to measure social variableslike interest and trust ought to enable more pro-ductive discussions, while the ability to measuresocial competition offers the possibility of reduc-ing problems like groupthink and polarization. Ifthe computer can measure the early signs of prob-lems, it can intervene before the situation becomesunsalvageable.

To explore these ideas, every student in myDigital Anthropology seminar used a GroupMediasystem so that our team could analyze the groupinteraction.12 Real-time displays of participantinteraction could be generated and publicly dis-played to reflect the roles and dyadic relationshipswithin a class. In Figure 4, the advisees (s2, s7, s8)have a high probability of conceding the floor totheir professor (s9).

This type of analysis can help develop a deeperdiscussion. Comments that give rise to wide vari-ations in individual reaction can cause the discus-sion to focus on the reason for the disparity, andthose interested can retrieve these controversialtopics for further analysis and debate later. Theanalysis also permits the clustering of opinions

Improving groupfunction requires

the ability to monitor the social

communication and provide real-time

intervention.

and comments using collaborative filtering. In thisway, people can readily see opinion groupings,which sets the stage for inter- and intragroupdebates.

Personal relationshipsSocial awareness may also be able to help rein-

force family ties, an important capability in this ageof constant mobility. Sensing when family mem-bers have had an unusually good, or unusually bad,experience can promote supportive communica-tion between them.

In one version, the system would randomly leavephone messages reminding family members to calleach other. However, when it senses that there hasbeen an unusual experience—a serious argument,an especially fun conversation, or an unusuallyintense meeting—the system would leave remindersfor others to call. The system would not tell peopleexactly why they should call, because doing so couldviolate people’s privacy. Instead, the reminderswould strengthen the family network by encourag-ing conversations precisely when family membersare most likely to appreciate them.

S ocial signaling seems to provide an indepen-dent channel of communication, one that isquantifiable and can provide an important

new dimension of communication support. The implications of a system that can measure

social context are staggering for a mobile, geo-graphically dispersed society. Propagating socialcontext could transform distance learning, forexample, letting users become better integrated intoongoing projects and discussions, and thus improv-ing social interaction, teamwork, and social net-working. Teleconferencing might become morereflective of actual human contact, since partici-pants could quantify the communication’s value.Automatic help desks might be able to abandontheir robotic, information-only delivery or theirinappropriately cheerful replies.

Our current systems are just a first step towardgenerally useful communications tools. We mustincrease the reliability of our social context mea-surements and learn how to better use them tomodulate communication. Much of our ongoingresearch is focusing on building meaningful math-ematical models for estimating social variables andexperimentally validating their use in a distance col-laboration framework.

Considering the personal and societal effects ofsocially aware communications systems brings to

mind Marshall McLuhan’s “the medium is the mes-sage.” By designing systems that are aware ofhuman social signaling, and that adapt themselvesto human social context, we may be able to removethe medium’s message and replace it with the tra-ditional messaging of face-to-face communication.

Just as computers are disappearing into clothingand walls, the otherness of communications tech-nology might disappear as well, leaving us withorganizations that are not only more efficient, butthat also better balance our formal, informal, andpersonal lives. Assimilation into the Borg Collectivemight be inevitable, but we can still make it a morehuman place to live. �

AcknowledgmentsI thank my collaborators—Joost Bonsen, Jared

Curhan, David Lazar, Carl Marci, M.C. Martin,and Joe Paradiso—and my current and former stu-dents—Sumit Basu, Ron Caneel, TanzeemChoudhury, Wen Dong, Nathan Eagle, Jon Gips,Anmol Madan, and Mike Sung—for all the hardwork and creativity they have added to this pro-ject. Thanks also to Deb Roy, Judith Donath, RozPicard, and Tracy Heibeck for insightful commentsand feedback. Parts of this article have appearedon Edge.org and in Proc. IEEE Int’l Conf.Developmental Learning.

March 2005 39

Figure 4. Display of group dynamics between professor(s9) and students during an experiment to study how agroup functions. Each student in the seminar received aGroupMedia system, which analyzed the class member’sinteractions on the basis of activity level, group interest,and turn-taking patterns. Circle size reflects speakingtime; the width of the link lines reflects the probabilitythat the person will concede the floor. The shadingwithin a circle reflects that person’s interest level, thedarker the shading, the higher the interest. Thicker circle borders denote groups.

s4

s5

s7

s9

s3

s6

s8

s1

s2

40 Computer

References1. A. Pentland, “Social Dynamics: Signals and Behav-

ior,” Proc. Int’l Conf. Developmental Learning, IEEEPress, 2004; http://hd.media.mit.edu.

2. C. Nass and S. Brave, Voice Activated: How PeopleAre Wired for Speech and How Computers WillSpeak with Us, MIT Press, 2004.

3. N. Ambady and R. Rosenthal, “Thin Slices of Expres-sive Behavior as Predictors of Interpersonal Conse-quences: A Meta-Analysis,” Psychological Bull., vol.111, no. 2, 1992, pp. 256-274.

4. R. Brown, Group Polarization in Social Psychology,2nd ed., Free Press, 1986.

5. M. Gladwell, The Tipping Point: How Little ThingsCan Make a Big Difference, Little Brown, 2000.

6. J. Jaffe et al., “Rhythms of Dialogue in EarlyInfancy,” Monographs of the Soc. for Research inChild Development, vol. 66, no. 2, 2001.

7. T. Chartrand and J. Bargh, “The Chameleon Effect:The Perception-Behavior Link and Social Interac-tion,” J. Personality and Social Psychology, vol. 76,no. 6, 1999, pp. 893-910.

8. T. Choudhury, “Sensing and Modeling Human Net-works,” PhD dissertation, Dept. Media Arts and Sci-ences, MIT, 2003; http://hd.media.mit.edu.

9. M. Laibowitz and J. Paradiso, “The UberBadge Pro-ject,” 2004; www.media.mit.edu/resenv/projects.html.

10. A. Madan, R. Caneel, and A. Pentland, “GroupMe-dia: Distributed Multimodal Interfaces,” 2004;http://hd.media.mit.edu.

11. N. Eagle and A. Pentland, “Social Serendipity: Prox-imity Sensing and Cueing,” 2004; http://hd.media.mit.edu.

12. N. Eagle and A. Pentland, “Social Network Com-puting,” LNCS 2864, Springer-Verlag, 1999, pp.289-296; http://hd.media.mit.edu.

Alex (Sandy) Pentland is the Toshiba Professor ofMedia Arts and Sciences at MIT and the formeracademic head of the MIT Media Lab. His workencompasses wearable computing, communica-tions technology for developing countries, human-machine interfaces, artificial intelligence, andmachine perception. A cofounder of the IEEEComputer Society’s Wearable Information SystemsTechnical Committee and the IEEE ComputationalIntelligence Society’s Autonomous Mental Devel-opment Technical Committee, Pentland has receivednumerous awards in the arts, engineering, and sci-ences. Contact him at pentland@media.mit.edu.