computation for the corridor bringing social software to social spaces max van kleek research...

Computation for the Corridor

Bringing Social Software to Social Spaces

Max Van KleekResearch Qualifying ExaminationJanuary 2005

motivationwhy public spaces and

circulation routes?

High traffic public spaces harbor thegreatest opportunity for chance encountersthat can lead to:

new acquaintances unplanned meetings

informal collaborations

One of the primary wayspeople build networks of casual acquaintances within their organization

knowledge workers need to collaborate today more than ever before

Knowledge worker : highly skilled participants of an economy where information and its manipulation are the commodity and the activity.

Examples: researchers, engineers, designers, architects product developers, resource planners, legal counselors, financial consultants, teachers, clerks

Contrast with: makers of physical goods or services

P. Drucker (1959)

Knowledge workers equipped with increasingly specialized skillsets, while facing broad challenging problems

Collaborations form around exchanging expertise/ sharing of skillsets; lets workers achieve goals more efficiently Collaborations more like “consulting sessions”: Spontaneous, small, loose-knit and short-term; Among “familiar strangers”

knowledge workers need to collaborate today more than ever before

“Coming Age of Social Transformation” (socioeconomic theory by Peter Drucker, 1959)

knowledge workers need to collaborate today more than ever before

But how do knowledge workers find collaborators?

out of context ! through friends of one’s supervisor casual social acquaintences in line at the cafeteria waiting for the elevator at the water cooler

IBM - “method to this madness” knowledge discovery server (2000) automated “knowledge management”

Steelcase Workplace Index Survey - (2002)977 full-time employees at various “knowledge-driven” companies– spent 50% of day working away from desks– Remaining time was spent “collaborating, and

holding impromptu meetings in secondary spaces, such as hallways, enclaves and water coolers”

– 64% preferred standing, reclining or leaning while engaged in impromptu meetings than meeting at their desks

– wished employers provided a greater range of seating products and meeting spaces that were conducive to such meetings.

Informal meetings occur frequently at work

– R. Kraut, C. Egido “Patterns for Contact and Communication in Scientific Research Collaboration”, ACM CSCW 1988.

Social proximity correlates with physical proximityand likelihood of mutual collaboration

Background

Other recent trends breakdown in sense of community– Overcrowding– Telecommuting– Disjoint workspaces

R. Kraut, C. Egido “Patterns for Contact and Communication in Scientific Research Collaboration”, ACM CSCW 1988.

“Strength of Weak Ties” (sociological theory by Mark Granovetter, 1973):

personal: more opportunities come from one’s “weak ties” than close friends; “personal success” correlated with size of social network

organizational: networks of weak ties facilitate knowledge flow across cliques within organizations; promotes cohesiveness and organizational memory

Meeting new people is important

yet today, organizations remain poorly socially connected

How well do you know your CSAIL colleagues?(In-house validation survey 1 faculty, 1 staff, 8 students, sept 2003)

How many people are you acquainted with on this floor? On neighboring floors? On other floors?a) 10-75%: mean 52%; b) < 10%; c) 0-1%,

How often do you like to share links (not-directly-work-related items) with lab colleagues. How do you do it? Most: Several times a day. A couple: a few times a

week. One: once a month1: Word of mouth; 2: IM; 3: e-mail.

How do you disseminate information to the entire lab? How often?Rarely, email all-ai, “inappropriate”; Paper posters

social software to the rescue?

Instant messagingPortholesIBM Knowledge Discovery ServerAmbient AwarenessShared Media SpacesMontage

Supplement “accidental” f2f interactions with various software

QuickTime™ and aTIFF (Uncompressed) decompressorare needed to see this picture.

But… the desktop is the wrong place for social software

Desktop : Work Context Competes with productivity for focus of attention Information overload Too many distractions already Ties users to their desks

Workers need to take breaks regularly for their own health MIT RSI guidelines: 1-2 mins every 15 minutes

5-10 minutes every two hours Provide a good opportunity for social activity

Social software belongs in social spaces.

applications

k:info: billboard/screen saver

(user)

knowledge sources andrecommenders

I think the User wants

I think the User wants

BreakingNews #32

It’s Monday.

Give user a break!

You are all wrong. She

wants to know about

free

cameras, microphones motion sensors

other perceptual inputs

Online sources

interaction history displayschedule

k:info, the “smart” billboard

serendipity:making k:infomore social

skinni:an ‘information kiosk’

architecture

ontogen: an ontology language for metaglue

interfaces

distinctive touch:identifying usersby gesture

QuickTime™ and aTIFF (LZW) decompressor

are needed to see this picture.

>> min(times)

1.0290 0.4060 1.5410 0.5180 0.3540 1.1390 1.39901.0310 2.2330 1.0930

>> mean(times)

1.3309 0.9380 2.4711 0.9367 0.5284 1.4328 1.5674 1.1913 2.4735 1.3355

>> max(times)

2.0140 1.5030 2.9120 1.3070 0.6900 1.8130 1.9180 1.3330 2.8690 1.7490

>> std(times) 1.2539 0.3313 1.1927 4.3841 2.1382 0.3453 2.2351 0.3209 0.7660 1.6609

>> mean(mean(times))

1.4206 >> vstd(times)

0.6467

< 5 seconds

QuickTime™ and aTIFF (LZW) decompressor

are needed to see this picture.

QuickTime™ and aTIFF (LZW) decompressor

are needed to see this picture.

QuickTime™ and aTIFF (LZW) decompressor

are needed to see this picture.

QuickTime™ and aTIFF (LZW) decompressor

are needed to see this picture.

QuickTime™ and aTIFF (LZW) decompressor

are needed to see this picture.

QuickTime™ and aTIFF (LZW) decompressor

are needed to see this picture.

trainingdoodles

featureextraction

build amodel/train

classifier

building a dt classifier

trainedclassifier

“max!”

feature extractiondean rubine: specifying gestures by example13 unistroke features: 11 geometric, 2 time

dt extractor rubine extr description

distfv f8 total euclidean distance traversed by a stroke

bboxfv2 f3 dimension of bounding box of a stroke

f4fv f4 stroke bounding box aspect ratio

startendfv f5 euclidean distance between start and end pts of a stroke

f1fv f1,f2,f6,f7 sine and cosine of start and end angles of a stroke

f9fv f9 sum of angle traversed by the stroke

f10fv f10 sum of absolute angle traversed by the stroke (“curviness”)

f11fv f11 sum of squared angles traversed by stroke (“jagginess”)

f12fv f12 max instantaneous velocity within a stroke

f13fv f13 total time duration of a stroke

training set: hiragana character set

train set:45 classes10 ex each1-5 strokes

test set:45 classes5 ex each

strokewise ML -

simplest multistrokegenerative model

represent each class as separate sequencesof states, each representinga stroke.

Each state thus has an associated parametric distribution overthe values of that stroke’sfeature vector

strictly encodes our previous assumption that strokes of the same class always arrive in the same order... otherwise, we’d need an HMM.

strokewise ML - 1 gaussian per stroke

features l-o-o performance test performance

distfv 0.5422 0.3467

bboxfv2 0.8067 0.6756

f4fv 0.5156 0.4267

startendfv 0.6067 0.4400

f1fv 0.8889 0.7511

f9fv 0.6511 0.5822

f10fv 0.5467 0.4222

f11fv 0.4311 0.3333

f12fv 0.2756 0.1289

f13fv 0.6378 0.4178

performance with individual features

strokewise ML - 1 gaussian per stroke

features l-o-o performance test performance

f9fv, bboxfv2, startendfv 0.9733 0.9156

f1fv, startendfv, f9fv, distfv

0.9644 0.8889

f1fv, f9fv, distfv, startendfv

0.9778 0.9644

all combined: distfv, bboxfv2, f4fv, startendfv, f1fv, f9fv, f10fv, f12fv, f13fv

0.9711 0.8800

performance with multiple features

fisher linear discriminant - (1-dimensional)OVA (one-versus-all):train C FLD binary classifiers on the fvsevaluate each one on the test point +1 if it gets the label right, 0 otherwise / (C*N)

features l-o-o performance test performance

distfv 0.9422 0.8311

bboxfv2 0.9356 0.8711

f4fv 0.8444 0.7556

startendfv 0.9600 0.880

f1fv 0.9244 0.9022

f9fv 0.8711 0.7644

f10fv 0.8467 0.5778

f11fv 0.7867 0.5556

f12fv 0.8467 0.7067

f13fv 0.9333 0.8400

fisher linear discriminant -

combined features (warning: figures are a bit misleading; we’ll describe why in the next section)

features l-o-o performance test performance

bbox2, f12fv, startendfv 0.9822 0.8489

f1fv, f9fv, startendfv 0.9533 0.9733

f1fv, f12fv, distfv, startendfv, bboxfv2

0.9867 0.9156

f1fv, f9fv, f11fv, distfv, startendfv

0.9778 0.9644

support vector machines -

OSU SVM Toolkit for Matlab [ http://www.ece.osu.edu/~maj/osu_svm/ ]

features linear k test perf. quad k test perf

distfv 0.9790 0.9790

bboxfv2 0.9789 0.9789

f4fv 0.9784 0.9784

startendfv 0.9778 0.9778

f1fv 0.9831 0.9831

f9fv 0.9778 0.9778

f10fv 0.9778 0.9778

f11fv 0.9778 0.9778

f12fv 0.9774 0.9775

f13fv 0.9778 0.9778

all combined 0.9923 0.9923

training took too long - no l-o-o

comparisonk-nearest-neighbors - simple, most sensitive to choice of feature extractors

sequential ML - simple to estimate, strictly requires stroke order

fisher linear discriminant (1d) - performed well

support vector machines (lin, quad kernel) - outperformed other methods, took significant training time

rejection

knn - greedily chooses k nearest neighborsstrokewise ML - chooses largest log likelihood

choose thresholds empirically using l-o-ovalidation (in theory, tricky in practice -soft thresholds difficult to manage)

FLD and SVMs - gauge ‘specificity’ of discriminantsby measuring performance as follows:

+1 iff all C FLDs/SVMs are correct0 otherwise=> “strict criterion”

Speech-basedinteraction

Searching for specific items via existing SKINNI GUI too slow;

Allow users to simply ask the kiosk to find what they are looking for:

“Where is Patrick Winston’s Office?”“How do I get to the Kiva Seminar Room?”“When is the Theory of Computation Seminar?”

Challenges:

Speech Understood

Speech

Error

Speech Overall

Touchscreen

Best 3 s 10s 3s 4 s

Worst 9s 25s 25s 19s

Avg 5.22s 16.78s 9.11s 7.33s

S.D. 0.92s 5.13s 6.24s 3.18s

What causes fragmented / localized awareness anda breakdown in sense of community?

Lack of regular interpersonal contact

Such as when working remotely

Inadequate channels of social communication

Overcrowding

QuickTime™ and aTIFF (Uncompressed) decompressorare needed to see this picture.

Recent trends in “creative workplace” designs have increased allocation to transient spaces and avenues

Moving away from static office-block layout from 1950’s that maximized personal territory

QuickTime™ and aTIFF (Uncompressed) decompressorare needed to see this picture.

Re-designing the workplace

QuickTime™ and aTIFF (Uncompressed) decompressorare needed to see this picture.QuickTime™ and aTIFF (Uncompressed) decompressorare needed to see this picture.

.. to semi-private arrangements with an emphasis onsmall-group meeting areas for spontaneous meetings,lounges, tea kitchens, and intersections for major circulation routes throughout the building

Ok-net

Extends physical architecture by providing digital services to these spaces to improve social well-connectedness

information dissemination

n-way social communications

augment face-to-face encounters

awareness across time + physical distance