Press release from ACM thatannounced the award.

San Diego, August 6, 2007—ACM SIGGRAPH will award its 2007 Significant NewResearcher Award to Ravi Ramamoorthi for hisseminal contributions to thevisual appearance of objects.Dr. Ramamoorthi, an associateprofessor of computer scienceat Columbia University, hasdeveloped mathematical andcomputational models that haveled to a deeper understanding ofvisual representation by digitallyrecreating or rendering complexnatural appearance. Some ofhis models have already beenadopted by industry.

The SIGGRAPH Significant NewResearcher Award is givenannually to a researcher whohas made a recent significantcontribution to the field ofcomputer graphics and is newto the field. Dr. Ramamoorthiwill receive his award at SIGGRAPH 2007, August 5-9, at the San Diego ConventionCenter, San Diego, CA.

Dr. Ramamoorthi’s work com-bines foundational mathematicalanalysis with novel practicalalgorithms to address long-standing problems in graphicsand computer vision. His

(continued on next page)

Graphics & Computer Vision Innovator Wins ACM SIGGRAPHResearch Award

Columbia Professor

Cited for Ground

Breaking Work on

Visual Appearance

of Objects

Columbia University in the City of New YorkDepartment of Computer Science1214 Amsterdam AvenueMailcode: 0401New York, NY 10027-7003

ADDRESS SERVICE REQUESTED

CS@CU NEWSLETTER OF THE DEPARTMENT OF COMPUTER SCIENCE AT COLUMBIA UNIVERSITY WWW.CS.COLUMBIA.EDU

NON-PROFIT ORG.

U.S. POSTAGE

PAID

NEW YORK, NY

PERMIT 3593

CUCS Resources

You can subscribe to the CUCS news mailing list at http://lists.cs.columbia.edu/mailman/listinfo/cucs-news/

CUCS colloquium information is available at http://www.cs.columbia.edu/lectures

Visit the CUCS alumni portal at http://alum.cs.columbia.edu

CS@CU

CS@CU FALL 2007 1

NEWSLETTER OF THE DEPARTMENT OF COMPUTER SCIENCE AT COLUMBIA UNIVERSITY VOL.4 NO.1 FALL 2007

Professor Ramamoorthi and colleagues have been working on methods for interactive imagesynthesis or rendering of realistic photograph-quality scenes. Images like this can be created in3-5 seconds per frame, with varying lighting and view. Note a variety of “photorealistic” effects,including complex natural lighting, intricate soft shadows of the chairs, and glossy reflectionsfrom the bowls on the table. The technique is based on a new mathematical framework ofwavelet triple product integrals, that has broad relevance in a variety of computer graphics,numerical analysis and multimedia applications.

Ravi Ramamoorthi

research explores practicalapplications ranging from real-time photorealistic rendering, toaddressing complex illumination,materials, and shadows incomputer vision. His renownedSIGGRAPH 2001 paper and Ph.D.dissertation used ideas fromsignal processing to establish afirm mathematical frameworkto describe reflection, whichhas led to a deeper theoreticalunderstanding of light transportas it relates to the visualappearance of objects.

Dr. Ramamoorthi received aB.S.degree in engineering as well as M.S. degrees incomputer science and physicsfrom the California Institute of Technology in 1998. He wasawarded a Ph.D. in computerscience from Stanford Universityin 2002, and joined the facultyof Columbia University in Augustof that year. A prolific contributorto ACM SIGGRAPH publicationsand conferences, Ramamoorthihas published his research inseveral other leading internationaljournals on graphics.

About ACM: ACM, the Association forComputing Machinery (www.acm.org),is an educational and scientific societyuniting the world’s computing educators,researchers and professionals to inspiredialogue, share resources and addressthe field’s challenges. ACM strengthensthe profession’s collective voice throughstrong leadership, promotion of thehighest standards, and recognition oftechnical excellence. ACM supports the professional growth of its membersby providing opportunities for life-longlearning, career development, and professional networking.

About ACM SIGGRAPH: The ACMSpecial Interest Group on ComputerGraphics and Interactive Techniques(www.siggraph.org) is an interdiscipli-nary community interested in research,technology, and applications in computergraphics and interactive techniques.Members include researchers, develop-ers and users from the technical, aca-demic, business, and art communities.SIGGRAPH provides information to thecomputer graphics community throughits annual conference, publications andthe SIGGRAPH Video Review.

NIH NewInnovatorDana Pe’er(pro-nouncedpay-er) is lookingforward to buildingher labteam andworking on

the next phase of her research,which seeks to illuminate howa cell’s regulatory networkprocesses signals, and howthis signal processing goeswrong in cancer. As one of theworld’s leading computationalbiologists, Pe’er develops highlysophisticated computational“machine learning” methodsthat analyze genomic data anddetect patterns that underlieinteractions and influencesbetween molecules in a cell.

With the NIH award funding,Pe’er and her team will seekto understand the generalunderlying principles governinghow cells process signals, howmolecular networks compute,and how genetic variationsalter cellular functioning.

Specifically, she wants tounderstand how changes inDNA codes modify a cellsresponse to its internal andexternal cues, which then leadsto changes throughout theentire body. These changes, or malfunctions, can cause anything from autoimmune disease to cancer.

“Cancer is a very individualdisease—unique in how itdevelops in every person,” saidPe’er, who came to Columbialess than a year ago with herhusband, Itsik Pe’er, also acomputational biologist at theUniversity. “Our research isaimed towards personalizedcancer diagnosis and treatment.For each individual patient, we wish to detect the keymutations that cause cancer,understand how these combineto cause the malignant behaviorand pinpoint where and howto target drug intervention,leading to better therapies anddrug development. I’m hopingthis will be my impact.”

Written by Clare OhArticle reprinted with permissionfrom Columbia News

CS@CU FALL 2007 32 CS@CU FALL 2007

Cover Story (continued) Feature Article

Dana Pe'er: DecodingGenetic Variations andRegulatory Networks

Dana Pe’er

(a) We view the cell’s molecular network as a “computer”,receptors sense the environment, signaling pathways integrateall signals and process and appropriate response, these activate transcription factors which execute this response in an orchestrated manner. Our lab aims to understand howmolecular networks function at a global systems scale.

(b) By applying Bayesian network structure learning algorithmsto high-throughput biological data, we successfully recon-structed known interactions and influences in human signalingpathways, as well as suggesting new influences which weexperimentally confirmed.

(c) The primary focus of our lab currently is to understand howchanges in the DNA sequence perturbs the regulatory network,alters signal processing and is manifested in complex cellularphenotypes such as growth rates in cancer tumors.

Dana Pe’er

How did you get interestedin computer science?

I think from a very young age,I always knew that I wanted tomix biology, math and computerscience, maybe when I was in 7th grade. Biology is really beautiful, very descriptive andvery hardcore physics and it boilsdown to math. I felt as thougha lot of patterns in biology needto be pushed forward withmath and computer science.

When I finished my Army training, there was this wholegenome revolution and DNAwas being sequenced and suddenly we began looking atmolecules. There were somenew technologies coming out,microarray technologies, whichreally allowed you to take snap-shots of all the RNA in the cell atonce. I really wanted to try andunderstand how cells compute.

We’ve got the parts list andsuddenly people don’t knowwhat to do with it. Microarrayswere generating these hugedata sheets for the activity often thousand genes across a hundred conditions and biologistswere used to looking at five or six numbers and analyzingthem. But what are you goingto do with a matrix of ten

thousand by a hundred? Areyou going to stare at that andfind a pattern? It really felt likesomething computer sciencecould help with.

What motivates your research?

I’m motivated by biologicalquestions. I’m motivated byquestions that look at systemsas a whole, and how theyfunction in a more perceptual,principled way. Beyond justdrawing a wiring diagram ofwhat the network is, I want tounderstand some fundamentalprinciples of how it works,how different changes in theDNA really propagate, how it’s designed.

Do you have a method orapproach to your research?

The place that I have a realadvantage is to go places, thatwithout new state of the artcomputation, you couldn’t go. I take Bayesian networks and make some adaptations to them, or take regularizedregression and make someadaptation to it. I didn’t inventit, but I do make non-trivialchanges to sophisticatedmachinery. The ability to takestate of the art algorithms,know how to use them properlyand know how to make non-trivial changes to them tomake them fit the biology,which is very different thanwhat they’ve been specificallydesigned for, is something that most bioinformaticiansdon’t have. So when I choosea problem, I try and play to my relative strength.

Tell me about your currentresearch challenges.

My current favorite project iswhere we study how moleculesinteract in yeast. I work inyeast because you can reallyask simple questions in yeastand actually get at them. OnceI develop the concepts inyeast, I’ve gone from yeast tohuman three times. One of the things biology has failed tostudy properly is combinatorics.They always like looking at asingle set. So I designed theexperiment where you havemultiple combinatorial changesto the DNA. You measure howthat changes the regulation andhow that changes the fitness.

There’s also an element of activelearning. I developed my owntechnology to do active learningin these biological networks.That’s a project that’s close tofinishing right now. Makingrandom changes in the DNAhas been easy but makingfine, exact sophisticatedchanges to the DNA has beenpretty much impossible. We’vedeveloped a new technology to be able to do that.

I wanted (my research) to berelevant to the world. I try andalways have an arm in the labthat takes these things and triesto apply them to somethingreal, something meaningful,which in my case is cancerbecause it is a disease whereregulation goes wrong.

How do you see your fieldevolving?

You have to work at multiplelevels. I’ve been criticizedbecause in biology, it’s so complex you have to focus. Ifyou look only in the small scale,then you’re never going to seethe big picture, you’re nevergoing to see how it all comestogether. You’re not going tounderstand everything by looking at small scale. On theother hand if you take too manysteps back, and look at it tooabstractly, you get the entiremess without understandingwhat little bits and pieces of itare doing. Therefore getting thetwo perspectives and constantlyjumping back and forth betweenthe bird’s eye view and themicroscopic view, really allowsme to think of new ideas.

Young Investigator Awards

CS@CU FALL 2007 54 CS@CU FALL 2007

Faculty Spotlight

Ravi Ramamoorthi

How did you get interestedin computer science?

I think the first time I gothooked into computers waswhen I saw a CASIO PB-110pocket computer in 3rd grade.Later, when I was in the 6thgrade, my father bought me aZX Spectrum home computerthat I got totally into. I tried towrite a number of games (in2D in BASIC), inspired by the(certainly simple by today’sstandards) various commercialgames available on it... in manyways, that’s when I really gotinterested in graphics.

What motivated you to focuson your current research?

When I was at Stanford, work-ing on my PhD in computergraphics, I thought renderingor image synthesis (creatingrealistic images on a computer)was really fundamental to thefield—it’s based on a sound theory of simulating physical lighttransport, and is basic to actuallymaking images (or creatinggraphics). Over the last fiveyears, I realized that computergraphics image synthesis isincreasingly using high-dimen-sional datasets to tabulate visual

appearance (a broad term todescribe the way objects orpeople “look“, fundamental tographics image synthesis). In theprocess, functions like lighting,reflectance properties of materi-als, spatially and time-varyingappearance are often measuredor tabulated. So, a major focus ison the appropriate mathematicalrepresentations and computa-tional models for visual appearance, with relevance to a broad range of practicalapplications. It was this body ofwork that the ACM SIGGRAPHAwards Committee recognized.

Do you have a method orapproach that you like to usein your research?

My goal has been to look forfundamental results, be theymathematical or computational,often focusing on the correctmathematical representationsand signal-processing tools. Inthis way, I believe one can layfirm scientific foundations forthe field, while impacting a widerange of practical applications.Besides the immediate impactin graphics and vision, my workhas for instance also settledlong-standing questions inradiative transport in physics (I had a paper published in theJournal of the Optical Societyof America in 2001).

I think a good paper and aresearch question would makean important insight and funda-mental long-term contributionto our scientific understanding,coupled with a useful practicalresult or implementation. Ofcourse, this is a difficult goalto strive for.

What do you consider themost interesting challengesof your current research?

I think the mathematical andcomputational challenge is howbest to represent, manipulate,edit, render with, and compresscomplex visual appearancedatasets. At its core it is adata representation and signal-processing problem. Someideas draw from techniques inother fields; some ideas (like ourwavelet triple product algorithmfrom SIGGRAPH 2004) areessentially new and potentiallyapplicable to other domains.Because of the fundamentalnature of the questions posedby this research, it has reallybeen an interesting journeyover the last several years.

From a practical perspective, thechallenge in visual appearance ishow do you acquire completelyrealistic models of lighting,reflectance, spatial variation etcfrom the real world and be ableto display them interactively soyou can really feel you are in areal environment. This is some-thing very hard to achieve; evenin offline computer graphicsimagery and movies, faces forinstance often look fake in somesubtle way. More broadly, howcan you obtain completely realis-tic simulations of the visual world(including motion and objectgeometry)? Will there come atime when a high-school studentcan create a completely realisticvirtual world in one afternoon?

Do you have a favorite projectthat you’ve worked on?

I’m proud of a number of contri-butions, mostly those that haveintroduced some fundamentallynew mathematical idea or com-putational insight. My SIGGRAPH2001 papers and PhD thesisintroduced the notion of reflec-tion from a surface as a sphericalconvolution, that brought a hostof new mathematical tools intothe field that have now becomewidespread. I really like the SIG-GRAPH 2004 paper on wavelettriple products, as it was a newand deep mathematical question,that hadn’t really been addressedpreviously even in the numericalliterature (and may have furtherimpact in that domain). This year,we’ve made a number of quitesignificant theoretical advancesbuilding on previous results,which really show how rapidlythe scientific foundations haveadvanced over the last five years.

How do you see your subfield evolving?

Since our goal at some level isto simulate the real world fully,graphics is constantly drawingfrom and giving back to a num-ber of different fields, includingphysical simulation, studies of human and animal motion,geometric representations inmechanics, and lighting simula-tion, as well as associatedhardware and numerical meth-ods. It’s an exciting time to bein the field, since graphics nowinfluences or has the potentialto impact almost all the otherareas of computer science andmore broadly many scienceand engineering disciplines,besides drawing from them(both ideas and researchers).

Our roving reporter Sean White interviewedProfessor Dana Pe’er and Professor Ravi

Ramamoorthi, to learn more about theirresearch and motivations.

CS@CU FALL 2007 76 CS@CU FALL 2007

Featured Articles Message from the ChairHenning

Schulzrinne,Professor & Chair

Thank you for reading our fall2007 edition of the ColumbiaComputer Science newsletter!

The Department is proud tointroduce two new additions toour faculty, strengthening coreareas of research and instructionin our department. We willintroduce both in the next edition of our newsletter inmore detail, but here’s a shortpreview. Simha Sethumadhavan(Ph.D., The University of Texasat Austin) joins the departmentin January 2008 as an assistantprofessor. His research interestsare in computer architecturesthat scale to Exaflops andbeyond, and software supportfor features found in thesearchitectures. He was part ofthe successful Tera-OP, TRIPSprocessor team and his mostsignificant contribution was thedesign and implementation ofthe first-ever fully partitionedprimary memory system. Hisdissertation proposes methodsto scale the hardware used for disambiguating memoryreferences which has been akey impediment to exploitinginstruction-level parallelism for several decades now. His other research interestsinclude application-specificarchitectures for emergingworkloads, implementation ofdistributed systems and novelcomputing technologies.

Junfeng Yang will receive hisPh.D. in Computer Sciencefrom Stanford University inJanuary 2008 and will join usnext summer. He currentlyworks at Microsoft Research.During his graduate studies, hebuilt tools that detected seriousdata-loss errors in real-worldstorage systems and securityvulnerabilities in the Linux kernel. His research interestsspan the areas of operatingsystems, security, and softwareengineering, focusing on practical error detection inlarge software systems.

Our MS program, both in itstraditional on-campus form andthrough our distance learningprogram (CVN), has long beena core part of the department,offering advanced computerscience education with theencouragement to pursueresearch projects. This year,we have welcomed our largestincoming MS class in years,with 114 new students.

For all of our students, I believethat being able to convincinglyexpress their technical ideas inwriting and public speaking isa core competency for aspiringcomputing professionals. This goes significantly beyondproducing grammatically correctEnglish, a skill that the AmericanLanguage Program at ColumbiaUniversity teaches to our

international students. In thepast, our Women in ComputerScience (WICS) group organizedseminars on writing topics forthe department; commonly, PhDadvisors spend a fair amount oftheir time editing the papers oftheir students. For my students,I have attempted to collectguidelines* and common mis-takes, but all of these attemptsseem insufficiently systematicand are hard to sustain. Thisyear, with the help of colleaguesat UCSB, we organized a multi-week workshop, taught byJanet Kayfetz. Students wrotediaries and technical articles, and presented practice talks. In smallgroups, students critiqued eachother’s writing samples. Thefeedback has been sufficientlypositive that we intend to repeatthis experiment next year. Weare also looking at other optionsthat turn aspiring researchersinto competent writers. I wouldbe particularly interested in hearing what other departmentsare doing to address this issue.

As always, I look forward tohearing from you as our readers.

*http://www.cs.columbia.edu/~hgs/etc/writing-style.html

Much of the work I do involvespublic policy, especially thingslike privacy, cryptography, andwiretapping. These are complexissues even from a policy perspective. However, what is possible—and what therisks are—is very dependenton technology. Lawyers, evensmart ones—and most of the ones I deal with are verysmart —are often unaware ofimportant technical issues.

As technologists, we have thesame right, qualification—andresponsibility—to speak out onsuch issues. However, we haveno more rights, qualifications,or responsibilities than anyoneelse if the issue is purely policy. Where we differ fromthe general population is onthe technical side. There, wehave special knowledge thatmost people don’t have; thatmakes us more qualified to

speak on certain points—butonly when they’re technicalpoints, not policy.

I confess to ambivalence aboutwiretapping. It’s inherently veryintrusive (hence the need for“minimization”, the eliminationfrom the record of all calls byothers who share the samephone line); it also providessome of the very best evidenceagainst people I really don’t like. The current US wiretaplaw was actually a pretty goodcompromise, in that it limitedwiretapping to serious crimeswhere other investigative techniques have failed or aretoo risky, but recent changes to the law, such as CALEA (theCommunications Assistance to Law Enforcement Act), aremore dangerous.

CALEA mandates that all phoneswitches contain built-in wiretap

interfaces. It was designed withan eye towards a certain set ofbenefits, without regard to thequite-considerable risks. That’swhere computer scientistshave value to add in the debate,above and beyond that of J.Random Citizen: we can warnpolicymakers what could gowrong. And go wrong thingsdid. Few computer security professionals were surprised tolearn that similar features on aGreek cellphone switch weresubverted by parties unknown;as a result, more than 100Greek citizens, including theprime minister, had theirphones tapped.

We’ve had better luck with voting machines. Though manystates rushed to adopt direct-recording electronic votingmachines in the wake of the2000 presidential election vote-counting fiasco, they are now

realizing that risks inherent insuch machines. As result, they’renow switching to machines thatproduce an auditable paper trail,just as most computer scientistshad suggested.

All that said, some lawyersreally do get it. A Court ofAppeals opinion striking downthe Communications DecencyAct showed that the judgesreally understood the end-to-end principle, a fairly subtleargument in systems design. I mentioned this to a friendwho had been at the hearing;he said that the judges wereasking questions based on it in real-time.

So—when lawyers come calling,don’t run away. You can have aneffect—and it’s fun.

Meddling with Lawyers By Steven M. Bellovin

In Lord of the Rings, Gildor, an Elf, advisedFrodo, “Do not meddle in the affairs ofwizards, for they are subtle and quick toanger.” Many people feel the same aboutlawyers. That said, I work with lawyers alot, to our mutual benefit.

your idea of the month. Youdidn’t marry it after all.

Lesson 2: if you are going toinvest lots of time building acomplex system, build one thatcan be used to explore manydifferent interesting scenarios,both because the problem thatmotivated you to build it mighthave been a dud and becauseit could lead you to lots ofother interesting problems!

Fast forward several months:I’m working in my office, tryingto get a particularly recalcitrantrouter (the router runs a Linuxdistribution for embeddeddevices called OpenWrt andcontrols the Roomba—or so thetheory goes) to do somethingstunningly simple, wonderinghow I got myself in this mess.By this point it’s clear the previ-ous work wasn’t going to unfoldin the expansive way I hadhoped and that it didn’t fit withthe Roombas either. So I had, aweek earlier, written it up to fileaway for some future time. I didso with the hope it might proveuseful, or perhaps spontaneouslymetamorphosize itself into a techreport. Meanwhile I was stuckwith misbehaving robots andonly a partially conceived plan ofhow I would get them to work.What precisely I would usethem for when I accomplishedthis was something I wasactively avoiding thinking about.Then I get an email from Vishal:“does anyone in the grouphave results appropriate for theMAMA workshop—work inprogress or results that may notbe quite conference worthy?” I pause for a moment, I reallyneed to keep working on theserobots and don’t really havetime to spare as I’m off to asummer internship soon. ThenI shake my head and run forthe file cabinet as an oldRabbinic saying runs throughmy head “if not now, when?”

Lesson 3: write up all yourresearch even if it’s not clearwhen or where you’ll publish it.Jump at the opportunity if anappropriate venue arrives. Whoknows where it will take you?

One month later, San Diego: Iwalk into the dining room of acrowded conference. I’ve justfinished my workshop talk, ormaybe I’ve still to do it. Doesn’treally matter the point is I’mhere, I don’t know anyone, andI need to sit somewhere forlunch—sort of like the first dayof high school actually. Ah, thereare Dan and Vishal. Maybe, I’llgo sit with them. The table’scrowded but there should beenough room for one more. I sitdown, introduce myself, listento the conversation. I eat yogurtand lettuce. Oftentimes that’sall you get when you’re bothstrictly vegetarian and kosher—I don’t like attracting attentionthat way, but it seems inevitableif I don’t want to starve. I try tomake small talk, am not terriblysuccessful, then:

Suman: …we’re looking for more student demos at Mobicom.

Vishal: oh Josh has a robotsystem… Roombas… networks. He could demo.

Me: pause, thinking (I’m startingmy internship this Monday andfinishing two weeks before thedemo, I’ve only got one partiallybuilt robot and no demo applica-tion. This could be a particularlyembarrassing way of committingprofessional suicide if I can’tkeep things on track.), then Ihear those darned Rabbis again.Yeah, definitely I’d love to! I’m pretty sure I can throwsomething together.

Suman: excellent send me an abstract.

Lesson 4: eat lunch with youradvisors (but don’t be a pest),listen to your Rabbis (at leastthe imaginary ones in yourhead—please feel free to substitute in Priest, Imam,Guru, Inner Spirit or whateverfits your belief system).

Monday, August 27th, Columbia:I’ve just returned to campus. A bit burnt out, finished myinternship (60-70 hours/week) onFriday, I’m very tempted to goto the beach, but for now theonly waves I’ll get to examineare ones produced by the sin

and cosine functions. I hadadvised two students over thesummer and came in to theoffice on one or two weekends.Not nearly as much had gottendone as I’d hoped—seems to bea given in life, eh? So I sat downwith a determined look on myface, ready to somehow forgethe disparate pieces of work,code, and robot I had left behindat the summer’s beginning. AndI did, but not with quite the vigorfor which I had hoped. It wasn’tthat I failed to work and produceand plod forward, it was that Iwas plodding not springing for-ward. I was frankly a bit worriedabout this, I needed to movefaster, what was I going to do?Nothing. I was tired. But I couldkeep going at a plod, and I did.

Lesson 5: keep working, nomatter how slow progress is.

Towards the middle of the weekbefore the demo, I began topick up speed, but it still wasn’tquite enough. Friday afternoonI was done, toasted, cooked,exhausted. I had a demo. I knewit wasn’t great. I could demo theequipment and the software, butthe application I’d come up withwasn’t implemented, nor wereits constituent parts even closeto being integrated. I showedthe demo to Vishal. He said itwould be fine, but maybe Icould find a way to get theapplication working. Instead ofa complete console GUI, whynot just have the robots spinand beep when they passed

messages? It was a great idea.“Maybe, I just don’t think it’s feasible in the time left, I’llsee if I can do something but I doubt it…”, I said. I wasdepressed. I couldn’t work onthe Sabbath, and I was flyingout for the demo on Mondaymorning, early. Plus I hadbought tickets to Farm Aid withmy wife for Sunday and didn’twant miss it. I needed to sleep.

Morning: I’m sleeping, but I’mthinking, about the demo.Could it be done? Haltingly,“yes, maybe, no yes it could”. I started trying to look at thecode as I recollected it in myhead. Yup, I’d do this, and that.This piece could be hookedinto that code base. Damn,that would be great! Whatabout the beeping? And thedancing? I could probably dothat in another 2 hours, needto modify some lower-levelstuff, implement some newroutines. How about the visualrecognition? Maybe that way?Yeah, perhaps. Perhaps I couldemulate downscaling thetransmission… all told—30hours? 25? 57? Doable though.Maybe. No, I could do it. Bynow I’m awake and energized.I’m ready to work as long andhard as I need to, only oneproblem—10 more hours ofthe Sabbath left. I’ve rarelywanted to break it this much,but I reluctantly decide to wait.

10.5 hours later: I’m working andI keep working with occasional

Like all good, and not so good,stories, this one has a beginning.Late fall of last year I was meet-ing with my advisors, VishalMisra and Dan Rubenstein. Wewere discussing a particularDelay Tolerant Networking (DTN)technique on which I had beenworking. We knew we’d like toapply this technique to someproblem, test it, and hopefullydemonstrate it’s efficacy—otherwise no one would bepersuaded to care about whatI discovered. (As it later turnedout, what I had discoveredwas, at least for the timebeing, a dead end). The conver-sation went something alongthese lines:

Vishal: We should test this.

Me: I agree, I could throwtogether a simulation.

Vishal and Dan: Simulationwould be a nice start, but it’s a bit passé.

Me: Well I have some experi-ence with robots…

Dan: How about Roombas?People like them!

Vishal: The robotic vacuumcleaners.

Me: They are quite neat, um but how do they fit theproblem?

Dan: We could use them in…

Me: I’m not convinced, thatdoesn’t really seem to utilizethe techniques we’ve beenworking on.

Dan: Okay how about…

(Loop a couple of times)

Me: Yeah I could see that, I’mstill not sure it’s the best fitwith this technique, but if youthink it’s a good idea, I’ll startlooking into it.

As it turned out, my reservationsweren’t entirely unfounded; theRoombas never really did fit thatproblem nicely. But more impor-tantly, it turns out my advisorshad given me the opportunityto build a system that fit manyother problems nicely!

Lesson 1: listen carefully toyour advisors and/or colleaguesand don’t get too attached to

By Joshua Reich

Several weeks ago I traveled to Montreal tocompete in the joint ACM Mobicom/Mobihocstudent demo competition. Of course I didn’t go alone, my robots came too, and I think I can say that we were all thrilled to go (it wastheir first time on an airplane after all). It was a long and winding journey that took us thereand from this journey, I have learned many valuable lessons (hopefully not all obvious),which I will now attempt to convey to you.

CS@CU FALL 2007 98 CS@CU FALL 2007

Featured Articles

Adventures in Mobile Networking: What I Learned From Building and Demoing a Testbed of NetworkedRoombas.

10 CS@CU FALL 2007 CS@CU FALL 2007 11

breaks for food and sleep until5PM the next day. Then I go to Farm Aid with the wife. I’mfeeling sick (sore throat) andtired, but Willie Nelson rarelycomes to New York and he’sgetting old. The Rabbis remindme of their favorite saying. Wereturn at 11PM having had awonderful evening—Willie, TheAllman Bros., Dave Matthews,Counting Crows (although I carefor them less). Linda’s happy, soam I. Work for another couple,pack, watch some BattlestarGalactica (great show), sleep for3. 6AM off to the airport robotsin tow, application (mostly)working (and it’s really cool).

Lesson 6: sometimes anunreasonable deadline can be helpful. Push yourself.

Lesson 7: solicit as much feed-back as you can get and listento it, but maybe after sleeping.

Lesson 8: You can trade sleepfor religious observance, DaveMatthews, and a happyspouse.

Montreal: was great. I spentmost of the time from arrival

through Tuesday afternoon inmy room fixing last minuteproblems, testing, and retesting,remotely compiling changes onColumbia’s servers and slappingthem onto my robots, whichwould then break unexpectedly.Those of you in research knowhow this one goes. The after-noon before the demo, I tookoff, walked around and justsavored being in Montreal. Itwas a wonderful, sun-filledafternoon. I won’t forget it any-time soon. I found a veggierestaurant and had dinner.Then I grabbed some OJ andyogurt (no lettuce this time) forthe next day. Back to work.

Lesson 8: You went somewhere,look around. Enjoy it, otherwisewhat’s the point?

Morning of, I got downstairsearly, but not as early as Ishould have. Some of the demospaces were already filled, andnone were as commodious asI’d been expecting. Also thelighting was terrible. I wanderedabout and tried to figure outwhere people would be enteringthe space, where they were

most likely to look, and whichspaces were obscured. I alsoneeded at least two walls tokeep my robots from wanderingoff (I only had two IR barriers).I almost took two tables, acrossfrom one another against awall. But I decided that wouldbe impolite and didn’t do it(someone else actually did dothis later on). Instead I took asomewhat less well-lit, off-to-the-side space. But it was inthe first row towards the entryway and in front of a second setof doors that might be opened(in the end they weren’t) and ithad the walls I needed. I tooka little time to see what otherfolks were doing, many of themhad business cards, handouts,etc. I didn’t. Stupid.

Lesson 9: Get there early, reallyearly. Scout the space, find anadvantageous location (maybedon’t worry about being a hog,maybe do—I haven’t decided).

Lesson 10: Have handoutswith your name and contactinfo at the ready.

While it was still relatively quiet,I tested one of my robots in the

space itself (the whole bunchwere over 100 lbs and I’d ratherdebugging occur in the privacyof my room—which was up thestairs and down the hall). Thevisual detection component wastoasted by the poor lighting. Iwas worried. Without this piece,the demo would be really lame.What would the robots alwaysrecognize that I could whip upon the spot? The direct beam ofa flashlight was the answer thatsprung to mind. I got one fromthe janitors, reprogrammed the robots, and while it wasn’tideal, it worked.

Lesson 11: Test your demo in the space beforehand, beprepared to improvise.

3PM: Show time. I still wasn’tready though. Another lastminute bug had popped up andhere I was frantically typing onthe terminal. When people comeby, I apologize, explain there isa glitch and that they shouldplease come back in a couple ofminutes, or enjoy my beautifulposter. 3:12PM: it’s closeenough, I’ll just wing it. The next2.5 hours are full of telling spec-

tators about my demo, shininga flashlight at robots, sweatingwhen they stop moving. I’mreplacing batteries and rebootinghardware (when it busts, whichit does often) all the whileanswering questions and takingeach opportunity to show folkshow the system works. After all,battery replacement is a niceopportunity to discuss lifetime,power-draw, performance, andcomponent integration; whilesoftware bugs open into a nicediscussion about the softwaredevelopment process inOpenWrt, issues with debuggingand the multi-threaded, socket-based, distributed software thatmakes our ‘bots tick. When allelse failed, I’d let the robots spinaround in crazy circles and giveanimated presentations in frontof the poster board—it seems tome that attendees actually foundthese very valuable. I also hadtwo units sitting on the table thatpeople could pick up and fiddlewith. And sometimes thingsworked smoothly. Those weregreat 5-minute spans. I didn’tget to use the bathroom or drinkanything, nor did I really get tosee much of the other demosand I completely forgot to takeout my camera. But while noone else from my group was atthe conference, my friend Yufangwas, and he told me about theother demos, gave me feedback,and moral support. If I had onlytold him about the camera!

Lesson 13: Be ready on time,but if you aren’t don’t panic.People will look at your poster,and look at your demo, theyalso look at you—be a carsalesman, but with out thegreasiness. Try to turn everyproblem into an opportunity.

Lesson 14: Go to the bathroombeforehand, have some waterhandy, bring a friend, and don’tforget to take out your camera!

So you must be wonderingwhat the demo actually did. Myresearch explores the challengesfaced by mobile, networkedsystems. This includes a spec-trum of scenarios: from thosein which nodes can manage toretain fairly robust connectivity,

to situations where nodes arehighly disconnected (in this latter situation, nodes can onlypropagate information by storingand forwarding one another’smessages). The purpose of mydemo was two-fold: firstly, itwas to showcase the mobile ad-hoc delay-tolerant networkingtestbed we’ve built (we call it aRoomba MADNeT), secondly, todemonstrate a proof-of-conceptapplication utilizing this testbed.The proof-of-concept applicationwe demoed was an opportunisticsurveillance application. Imaginea national park sometime in thefuture—a very large area, someservice robots moving around,some vehicles with mountednetwork-enabled data-gatheringequipment, perhaps evensome units worn by hikers and rangers. All of these unitscould take pictures of grizzliesand other interesting animals,but wouldn’t be in control oftheir own mobility patterns,nor would they generally bewithin transmission range ofeach other. How would thesenodes get their data to inter-ested researchers? Our demoapplication solved this problemby leveraging a data propagationtechnique we had previouslydeveloped, and you’ve alreadyheard the story of how it was

implemented (see our upcomingMC2R paper for more details).The aforementioned lightingproblem prevented our nodesfrom recognizing their targetaccurately, hence our substitutionof the flashlight for our originaltarget, a red Roomba. Whenone node saw the flashlight itwould demonstrate this by stop-ping and beeping. Other nodesdid likewise when receivingmessages. As in the nationalpark example this happenedonly gradually and dependenton nodal movement patterns,since our nodes only rarely werewithin range of one another(we dialed down the broadcaststrength in software). In thisway, onlookers could see thepropagation of informationacross the network and interactwith it at the same time. Itwas actually quite interestingto watch, and, along with luck,was probably the reason asteady stream of visitors cameto our demo area.

By the time things wrapped upI felt wrung-out, exhilarated,and ready for dinner. ThankfullyI had Yufang (by this point therobots were sleepy, grumpyand poor company). We packedup the equipment and wentout to celebrate. What were

we celebrating? Having done agood job? Having survived? I’mstill not sure. The next day Ireturned to New York, thinkingthe next couple of weekswould provide for a moderatedegree of downtime. Then Igot a call from Vishal (it wasmoments from sundown andI’d be off email for 3 dayswhilst observing the Jewishnew year): “Had I checked myemail?” “Yeah, I saw that I wasone of the four finalists, I wastruly thrilled.” “No, you won.”“I won? The competition?”“Yeah, you won the entire thing,best demo”. I was a bit over-whelmed. It was a great startto the new year. In addition tobeing tremendously gratifying,this experience seems to bepropelling me towards otherexciting adventures, researchopportunities, and problemsI’m eager to begin solving.Who knows where they willlead me? I did miss the award presentation dinner though, soperhaps I’ll close by thanking allthe people who are supportingme through this continuingjourney —my advisors Dan andVishal, my colleagues in theDistributed Network Analysislab and the department, therobots, Yufang, and of course,my family and wife.

In an attempt to do somethingthat I had never done before, I spent a day modeling for an ad agency in Taiwan thatwas doing a promotion forMcDonald’s. This was my wifeNina’s fault, of course: she hadvisited Leo Burnett (the agency)to try to promote her work andone thing led to another. Theywere looking for a westernmodel for a promotion and discovered she had an Americanhusband. She sent off a fewpictures of me and since theydidn’t know any better, theygave me the one-day job.

In fact, this whole episode canbe attributed to Nina’s network-ing skill. The fully story goes likethis: Nina went to Italy to attenda children’s book conference.There, she met the head of thesociety of children’s book illustra-tors in Taiwan, who invited her

to give a lecture. Nina accepted,and at her lecture in July, shemet an illustrator who hadworked for Leo Burnett, the adagency. This illustrator referredNina to the right people at LeoBurnett, who happened to noticethat Nina had an unusual (i.e.,western) last name and askedif her husband was a westernersince they were looking for awestern model.

They first brought me in for a “fitting” at 11:30. It was thenthat I learned what they wantedme for: they wanted an “eskimo”(actually, an Alaskan) for a posterthey were doing for McDonald’sstores in Taiwan in September,2006. Great. It’s 100F degreesoutside and they want me todress up in a really warm jacket,hat, and scarf. The fitting wentfairly quickly (they just wantedto make sure the colors they

A Day in the Life of a Model Professor

had chosen were reasonableand that I looked something likewhat they were looking for), soNina and I went to visit somefriends until 6:00 PM, when theyneeded me back at the photostudio for the actual shoot.

There’s actually quite a lot ofirony here: having grown up inMinnesota, which is routinelycolder than Alaska in the winter,I’m actually a reasonable stand-in for an Alaskan. I certainlyhave the experience of wearingreally warm jackets like the oneI’m pictured in.

They sat me down, appliedmakeup (!) and then startedshooting. I guess I did OK, butthey kept wanting me to smilebigger, which is hard when you can’t hear anything fromthe tight hat and you’re boilingunder the hot lights. Theyturned the air conditioner up fullblast and pointed a fan at myfeet and it was still very hot.

After four hours, they finallygot what they wanted. Therewere about eight other peoplein the studio all working on thisincluding the photographer, the hard-working stylist, whowas constantly retouching mymakeup, fussing with the crazyhood she had sewn onto thejacket, an art director or two,somebody from McDonald’s,and a bunch of assistants.

The hood was a good story:they wanted a furry hood, butthe jackets they could find didn’thave a hood that was furryenough. So they got a minkstole and wrapped it around the hood on my head. I thinkthe stylist quickly sewed it in place, but we had to keepmoving it around.

It was an interesting experience,but I wouldn’t want to make a living being a model. My face

ached for days from trying tosmile wider.

The “fish” campaign for whichthis photoshoot was done turnedout to be fairly substantial. Notonly was I on a poster, but alsoon large banners at McDonaldsacross Taiwan, and theMcDonald’s Taiwan website.(http://www.mcdonalds.com.tw/)

They put up the posters, banners, etc. on September1st, 2006. Nina and I went fora walk through downtownChunli to visit the McDonald’sthere. We found two that hadmy photo up: a large one at amajor intersection that had twobanners outside and my posterinside, and a smaller one thatjust had the banner. As wewere visiting, some Taiwanesehigh school girls looked at meand asked Nina, “Is that him?”

If anybody from McDonald’s inTaiwan (or elsewhere) reads this,know that I enjoyed workingwith such a professional group,which was a big part of why I wanted to do this. Plus, itmakes for a good story.

12 CS@CU FALL 2007 CS@CU FALL 2007 13

By Stephen Edwards

In an attempt to do something that I hadnever done before, I spent a day modelingfor an ad agency in Taiwan that was doinga promotion for McDonald’s.

Featured Articles

“My face ached for days from trying to

smile wider.”

CS@CU FALL 2007 1514 CS@CU FALL 2007

Professor Peter Allen, ComputerScience, Dennis Fowler,Surgery, and Professor Nabil Simaan, MechanicalEngineering, have been awardeda 3 year NIH ExploratoryResearch Grant to develop an“Image Guided in vivo ToolingPlatform for Minimal AccessSurgery”. This project focuseson the development of a newlyconceived insertable roboticeffector platform and the integration of that platform witha recently developed insertable,remotely controlled camera system to be used for minimalaccess surgery. The project willinvolve the actual design andconstruction of the platform fortools and the integration of theimaging platform (insertablecamera system) with the toolsinto a fully functional imageguided system for minimalaccess surgery. This may alsoinclude the addition of varioussensors on the tools, so thatthe resultant data stream fromboth the imaging platform andthe tools can be processed to control the intervention. The overall aim is to develop a disruptive technology thatincludes an insertable imagesource, a wide range of surgical tools, and a computerto integrate the function of all components.

Ph.D. student Hila Becker wasselected as a finalist in the2007 Anita Borg MemorialScholarship competition. TheAnita Borg Institute for Womenand Technology and Googlesponsor the 2007 Google AnitaBorg Memorial Scholarship. TheGoogle Anita Borg MemorialScholarship was established tohonor the legacy of Anita Borgand her efforts to encouragewomen to pursue careers in computer science and technology. Finalists receive a cash award. For the 2006-2007academic year, the institutereceived over 250 applicationsfrom students at 115 differentuniversities across the country.Eligible students must be goinginto their final year of study at a US university or college.

Selection criteria includes academic performance, lettersof recommendation, answers to short essay questions andinterviews with members of thereview committee. After threerounds of review, the committeeselected 50 finalists.

Ph.D. student Matei Ciocarlie

won the best student paperaward at the 2007 World HapticsConference in Tsukuba, Japanfor his paper “Soft Finger Modelwith Adaptive Contact Geometryfor Grasping and ManipulationTasks” (co-authored withClaire Lackner and ProfessorPeter Allen). The prize carriesan award of $1,000.

Professor Stephen Edwards

was featured in an EE Times article about a ‘Wild and CrazyIdea’ session at the 2007 DesignAutomation Conference (DAC)held in San Francisco in June.

Professor Steven Edwards wasawarded an NSF grant (joint withProfessor Edward Lee of UCBerkeley) on “Precision TimedArchitectures”. This project proposes to reintroduce timingpredictability as a first-classproperty of embedded processorarchitectures.

Professor Eitan Grinspun wonan NSF CAREER award, titled“Multiresolution Foundationsfor Physics-Based ComputerAnimation and InteractiveEngineering Design”. This projectwill develop new computer simulation techniques that helpengineers, scientists and artiststo understand the behavior ofcomplex mechanical systems.Software that helps to developintuition will help engineers toproduce better designs, spurscientists to poise more likelyhypotheses, and give artistsbetter control over the processof computer animation.

Professor Jonathan Gross’snewest book, CombinatorialMethods with ComputerApplications, was published inlate October 2007. The topicsinclude recurrences, sums,

binomial coefficients, permuta-tions, number-theoretic tech-niques, graph enumeration,and combinatorial designs. Thecover features the painting“Peaceable Kingdom”, by theAmerican artist Edward Hicks.

Professor Julia Hirshberg waselected to the University Senate.

Professors Gail Kaiser andJason Nieh were awarded agrant from the NSF to study“Autonomic Mechanisms forReducing System Downtime dueto Maintenance and Upgrades”.The project investigates anddevelops autonomic mechanismsfor reducing system downtimedue to software maintenanceand upgrades. The projectaddresses operating systemupgrades and also applicationupgrades, focusing on stand-alone binary-executable applica-tions. The main goal is to lessenthe possibility that patches andupdates will “break” expectedfunctionality of the environmentthat worked fine together withthe old version—overall maxi-mizing availability and reliabilityboth during and after mainte-nance while imposing littlemanagement overhead.

Professor John Kender

received a highly-competitiveNSF-Information and IntelligentSystems Division grant titled“Analysis and Display ofSemantics of Structured andUnstructured Videos”. This sup-plements an earlier award fromthe Disruptive Technology Office.This project explores three newrelated approaches to making theindexing and retrieval of videosmore efficient, meaningful, andhumanly navigable, even whenthe videos have little editor-imposed structure. The first is theexploration and refinement of anovel, highly efficient machinelearning technique for data-richdomains, which selects small and fast subsets of multimediafeatures that are most indicativeof a given high-level concept.Speed-ups of three decimalorders of magnitude are possible.The second is the development

of new methods and tools forrefining user concepts anddomain ontologies for videoretrieval, based on statisticalanalyses of their collocation andtemporal behavior. The goals are the determination of videosynonyms and hypernyms, theverification of temporal shot patterns such as repetition andalternation, and the exploitationof a newly recognized power-law decay of the recurrence of content. The third is the demonstration of a customizableuser interface, the first of itskind, to navigate a library ofvideos of unedited and relativelyunstructured student presenta-tions, using visual, speech, facial,auditory, textual, and other fea-tures. These features are shownto be more accurately and quicklyderived using the results of thefirst investigation, and more com-pactly and saliently presentedusing the results of the second.

Professor Angelos Keromytis

was awarded a 3-year grant fromthe NSF CyberTrust program to investigate the use of a newcommunication paradigm againstnetwork denial-of-service attacks.This project will investigate anew communication paradigm,named PacketSpread, whichmakes feasible the use of capability-like mechanisms on the current Internet, withoutrequiring architectural modifica-tions to networks or hosts.Professor Keromytis was alsoawarded a different 2-year grantby the NSF CyberTrust program,where he will investigate the useof new security policy modelsand mechanisms that integratevarious protection and reactionelements in an enterprise net-work in a coherent and effectivedefense system.

Professor Angelos Keromytis

won a 5-year grant with theUniversity of Pennsylvania andGeorgia Tech, titled “Foundationaland Systems Support forQuantitative Trust Management”.The goal of this project, fundedby the Office of Naval Research(ONR) under the Multi-UniversityResearch Initiative (MURI) pro-

Department News & Awardsgram is to develop QuantitativeTrust Management as the basisfor a scalable decentralizedapproach to dynamic, mission-based access control (MBAC).Dynamic trust managementtechniques will address theinabilities of current capabilitiesto maintain security policies atthe operational tempo requiredfor network-centric warface, toscale to emerging nation-statethreats, and to manage hetero-geneous computing and networkelements supporting Service-Oriented Architectures (SOA).

Professors Angelos Keromytis,Jason Nieh and Sal Stolfo wona 5-year MURI grant, funded bythe Air Force, titled “AutonomicRecovery of Enterprise-wideSystems After Attack or Failurewith Forward Correction”, jointwith GMU and Penn StateUniversity. This project willdevelop autonomic recovery andregeneration mechanisms thatwill enable commodity systemsto detect attacks, corruptions,and failures, then self-regenerateto a known good state, forboth program and data, whileincreasing the reliability andsecurity of the software to be more resistant and less vulnerable to attack.

Professor Tal Malkin receiveda grant from Mitsubishi ElectricResearch Laboratories (MERL) tostudy privacy-preserving learning.

Professor Vishal Misra waselected to the board of directorsof ACM SIGMETRICS. SIGMET-RICS is the ACM SpecialInterest Group (SIG) for thecomputer/communication sys-tem performance community.

Professor Shree Nayar’s work oncomputational photography wasprominently featured in an April2007 edition of Science News.

Associate professor Jason

Nieh was awarded tenure.Congratulations, Jason!

In Spring 2006, ProfessorSteven Nowick was invited byDuane Armstrong, a manager at NASA Goddard Space Flight

Center (Greenbelt, MD), to collaborate in developing a newadvanced measurement chip forspace applications. The projectis exploring the potential benefitof clockless (i.e. asynchronous)digital circuitry for time-of-flightlaser measurement. Over thepast year, using one of Prof.Nowick’s asynchronous digitalcircuit styles for controllers(called “burst-mode”), and hisautomated computer-aideddesign (CAD) tool “Minimalist”,they have successfully designeda working chip. The chip is beingincorporated by NASA into several prototype applications,where it will be evaluated foruse in future space missions.

Ph.D. student Janak Parekh

spoke on behalf of doctoralcandidates at Convocation onMay 14 as part of Columbia’s2007 graduation exercises. Thisis the Convocation for doctoralcandidates in the schools of Architecture, Business,Engineering, Journalism, Law, Nursing, Physicians andSurgeons, Public Health, SocialWork, and Teachers College.Vice President Lee Goldman

and a faculty member alsospoke at this celebration.

Professor Dana Pe’er is among29 researchers selected frommore than 2,100 applicants toreceive the 2007 NationalInstitutes of Health (NIH)Director’s New InnovatorsAwards. The awards recognizethe promising work of scien-tists who are in the early yearsof their career and who havenot previously received an NIHregular research, or R01, grant.(See the article in this editionof the newsletter.)

Professor Ravi Ramamoorthi

was named the winner of the2007 Significant New ResearcherAward by ACM SIGGRAPH, theACM special interest group oncomputer graphics. The citationreads: “The Significant NewResearcher Award is awardedannually to a researcher who hasmade a recent significant contri-bution to the field of computer

graphics and is new to the field.The intent is to recognize peoplevery early in their career whohave already made a notable con-tribution and are likely to makemore. ACM SIGGRAPH is proudto recognize Ravi Ramamoorthifor his groundbreaking work onmathematical representationsand computational models for the visual appearance ofobjects”. (See the article in this edition of the newsletter.)

Professor Ravi Ramamoorthi

was named a winner of the2007 Office of Naval ResearchYoung Investigator Award, one of only two awardees inComputer Science.

Professor Ravi Ramamoorthi

received an NSF grant titled“Theory and Algorithms for High Quality Real-TimeRendering and Lighting/MaterialDesign in Computer Graphics”.The proposed research is ondeveloping the theoretical foun-dations and next generationpractical algorithms for highquality real-time rendering andlighting/material design.

Research scientist Owen

Rambow of the Center forComputational Learning Systems(CCLS) received an NSF granttitled “Email, Social Networks,and Organizations: InvestigatingHow We Use Language toCreate and Navigate Social andOrganizational Relations”. Theproject studies three aspects ofhuman linguistic communication:the language used in the communication (for example,whether formal or informal), thetopology of the social network ofthe communicators (for example,whether the the speaker isembedded in a single tight-knitgroup), and the roles the commu-nicators occupy in an organization(for example, whether the speak-er is an upper-level manager, or an administrative assistant).

Ph.D. student Joshua Reich

won the student demo contestat 2007 ACM Mobicom/hoc forhis Roombanet system. (Vacuumcleaners roaming around the

courtyard have been a familiarsight in the CS department for some time.) His projectwas titled “MadNETt: DTNs on Roombas”. Mobicom is the flagship wireless conference andthe 2007 conference was jointwith Mobihoc. (See the articlein this edition of the newsletter.)

Professor Henning Schulzrinne

was elected Vice Chair of ACMSIGCOMM. SIGCOMM is theACM Special Interest Group(SIG) on Data Communications;it is ACM’s professional forumfor the discussion of topics inthe field of communications andcomputer networks, includingtechnical design and engineer-ing, regulation and operations,and the social implications ofcomputer networking.

Professors Rocco Servedio

and Tal Malkin received an NSFCyberTrust grant on cross-lever-aging cryptography and learningtheory. The project proposes adetailed study of the connectionsbetween cryptography and learn-ing. Very roughly speaking, cryp-tography is about manipulatingand encoding information so thatit is difficult to reconstruct the initial information, while learningtheory is about efficiently extract-ing information from someunknown object. This dualitymeans that ideas and resultsfrom each area can potentially beleveraged to make progress inthe other area. The first main goalof the project is to obtain newcryptographic results based onthe presumed hardness of vari-ous problems in computationallearning theory. Work along theselines will include constructing andapplying cryptographic primitivessuch as public-key cryptosystemsand pseudorandom generatorsfrom learning problems that arewidely believed to be hard, andexploring the average-case learn-ability of well-studied conceptclasses such as decision treesand DNF formulas. The secondmain goal of the project is toobtain new learning results viacryptography. The PIs will work todevelop privacy-preserving learn-ing algorithms; to establish com-

16 CS@CU FALL 2007 CS@CU FALL 2007 17

putational hardness of learningvarious Boolean function classesusing tools from cryptography; toobtain computational separationsbetween pairs of well-studiedlearning models; and to explorethe foundational assumptions ofwhat are the minimal hardnessassumptions required to provehardness of learning.

Ph.D. student Yingbo Song

was named as a SEAS “GreatTA” for the Fall 2006 semester.Congratulations, Yingbo!

Professors Sal Stolfo, Angelos

Keromytis and Shlomo

Hershkop won a 2-year grantfrom the Department ofHomeland Security (DHS), aspart of a larger project with the Institute for InformationInfrastructure Protection (I3P).

The project, titled “HumanBehavior, Insider Threat, andAwareness”, will focus on inves-tigating, developing, and experi-mentally evaluating methods andmodels for detecting maliciousinsider activity and behavior ona host computer system.

Ph.D. student Sean White,M.S. student Dominic Marino

and Professor Steve Feiner

won the Best Note award atCHI 2007 for their short paper,“Designing a Mobile UserInterface for AutomatedSpecies Identification”. CHI2007 is the top conference inhuman-computer interaction,and was held in San Diego,April 28-May 3, 2007.

Former PhDStudent Montek

Singh (Ph.D.,2001) has beenpromoted withtenure to asso-ciate professorof computer

science at the University ofNorth Carolina, Chapel Hill. Hisresearch is in high-performanceasynchronous and mixed-timingdigital circuits and architectures.

Larry Leftin and colleagues at Honeywell were recentlyawarded a patent in controltheory: “Methods and systemsfor implementing an iteratedextended Kalman filter within a navigation system” (UnitedStates Patent 20070213933).This patent applies to all navigation systems, especiallythose whose hardware exhibitsnon-linearities.

Eric Siegel

(Ph.D., 1997)has spent thelast four yearsfounding a datamining and pre-dictive analyticsboutique firm

in San Francisco. The company,Prediction Impact, Inc. (www.predictionimpact.com), providesconsulting services to its clientsby forming specialized teamsfrom its network of senior consultants. The firm also offerscommercial training programs;more information about these isavailable from the website. Onthe personal side, Eric is marriedto Maria de Fatima Callou, a fash-ion designer he met five yearsago while visit Jacques Robin(Ph.D. 1995) in Brazil. “Thanks fornetworking me to true love!” Ericproclaims to the CS department,and, especially, Dr. Robin.

Department News & Awards (continued)

Alumni News

Recent & Upcoming PhD Defenses Ricardo BarattoTHINC: A Virtual and Remote DisplayArchitecture for DesktopComputing

Continuing advances in the performance and ubiquity of networks have enabled the proliferation of technologies thatextend our computing environmentbeyond the boundaries of a single computer. As one of thesetechnologies, remote display hasenjoyed increased popularity inrecent years due to the numberof benefits associated with it.However, its mass acceptance hasbeen hampered by the inability ofcurrent technology to provide ahigh fidelity visual and interactiveexperience for end users acrossthe vast spectrum of graphicaland multimedia applicationscommonly found on traditionaldesktop computers. This dissertation introducesTHINC, a virtual display architec-ture for high performance remotedisplay on local and wide areanetwork environments. THINC vir-tualizes the display by introducinga simple device driver that looksand behaves like traditional,hardware-specific drivers. On topof the driver, THINC implementsnovel optimizations for efficientlytranslating high level applicationdisplay requests into a set of simple, low-level commands, and techniques to prioritize andguarantee that only relevant com-mands are delivered to clients.We have compared THINC’s performance to a set of current,state of the art commercialremote display systems. Ourresults demonstrate that THINCcan deliver good interactive performance even when usingclients located around the world,and provides much fasterresponse time than other systemsin WAN environments. Building on THINC’s core archi-tecture, we present two systemsthat address the need for multi-media and mobile user support in remote display systems. First,we extend our virtualizationapproach to audio devices toprovide audio capture and play-

back, and leverage existing video acceleration interfaces toefficiently deliver synchronizedaudio/video desktop content.Second, we present pTHINC, a PDA remote display solutionthat leverages server computingpower to run full-function desk-top environments and only sendsimple screen updates to thePDA for display. pTHINC usesserver-side scaling of displayupdates to provide high-fidelityand seamless mobility across abroad range of different clients,screen sizes, and orientations.pTHINC also leverages existingPDA control buttons to improvesystem usability and maximizeavailable screen resolution forapplication display. Finally, this dissertation demon-strates how THINC’s virtual display approach can be used for desktop computing systemsbeyond remote display. First, weshow how display virtualizationcan be used to encapsulate thestate of a desktop session anddecouple it from the underlyinghardware and operating systeminstance, allowing it to be trans-ferred across computers. Second,we present a system that lever-ages THINC’s inherent ability toredirect display output to trans-parently record all desktop con-tent. Coupled with a text capturemechanism, this system enablesusers to playback, browse, and search the recorded data,allowing them full visual access to any information that has everbeen displayed on their desktop.

Arvid BessenOn theComplexity of Classical and QuantumAlgorithms for NumericalProblems

in Quantum Mechanics

Our understanding of complexquantum mechanical processesis limited by our inability to solve the equations that govern it except for simple cases.Numerical simulation of quantumsystems is therefore essential tounderstand, design and improvequantum systems. We would like

to understand the computationalcomplexity of quantum mechanicalproblems and investigate theeffect the introduction of quantumalgorithms has on these problems. In this thesis we explore differentcomputational problems fromquantum mechanics, study theircomputational complexity and tryto find ways to realize the bestpossible algorithm. We investi-gate the reasons behind theimproved performance of Shor’squantum factorization algorithmand we study the quantum phaseestimation algorithm used byShor’s algorithm. We apply ourtechniques to the Sturm-Liouvilleeigenvalue problem and provelower bounds for it. We study the importance ofdestructive interference in quan-tum computing and how it makesquantum algorithms more power-ful than classical stochasticcomputation. In order to do thatwe consider the complexity classStoqMA of algorithms that onlyuse classical gates, but are givena quantum “witness” and showthat there exists a completeproblem for this class, which we call the stoquastic localHamiltonian problem. Finally we study a computationalproblem from lattice quantumchromodynamics (QCD). In mostpopular algorithms that treatproblems in QCD the (gauged)Dirac matrix has to be invertednumerous times. We study anewly proposed multigrid method,adaptive smoothed aggregationthat has promise to overcomethese difficulties. We show thatwhile classical CG’s convergencebecomes worse as the matrixbecomes almost singular, adaptive smoothed aggregationwill still perform well.

Hanhua FengScheduling:From Optimalityto Configurability

We begin bystudying optimalscheduling

policies in a single-server systemand optimal dispatching policiesin systems with multiple parallelqueues and servers. With different

assumptions of system character-istics (e.g. workload patterns) andperformance requirements (e.g.mean response time and meanslowdown), the optimal schedulingpolicies differ vastly. As an example, in a system with a singlequeue, we show that a schedulingpolicy can be the best schedulingpolicy for one kind of service-time distributions and at thesame time the worst for another,and that a scheduling policy canbe the best policy in terms ofmean slowdown but the worst interms of mean response time. Since there is no such thing asubiquitous optimality, we put ourefforts into two directions. First,inspired by the observation thatthe variability of service time is adominant factor on the optimalityof scheduling policies, we setconstraints on service-time vari-ability and investigate the bestand worst scenarios in terms ofmean response time for severalscheduling policies on a single-server queue. Second, we proposea configurable policy that can be tuned with a parameter. Thisconfigurable policy is not onlyanalyzed in both deterministicand stochastic configurations,but also implemented in Linux asa kernel scheduler. We justify byexperiments our claim that theubiquitous optimality can now beachieved by configurability.

Elena FilatovaUnsupervisedRelationLearning or Event-FocusedQuestion-Answering

and Domain Modelling

In this thesis, we investigate theproblem of identifying, within a text, relations that captureinformation important for event-focused document collections.The presented solutions workwith events of various granularityand show how to use these rela-tions to improve the performanceof a number of natural languageprocessing applications. For a set of related event-focused documents, we intro-

CS@CU FALL 2007 1918 CS@CU FALL 2007

Recent & Upcoming PhD Defenses (continued)types of speaker state: paralin-guistic state, in particular emo-tion; pragmatic state, in particu-lar questioning; and the state ofspoken language proficiency ofnon-native English speakers. Paralinguistics. Intonational fea-tures describing pitch contourshape were found to discriminateemotion in terms of positive andnegative affect. A procedure isdescribed for clustering groups oflisteners according to perceptualemotion ratings that foster furtherunderstanding of the relationshipbetween acoustic-prosodic cuesand emotion perception. Pragmatics. Student questions ina corpus of one-on-one tutorialdialogs were found to be sig-naled primarily by phrase-finalrising intonation, an importantcue used in conjunction with lexico-pragmatic cues to differ-entiate the high rate of observeddeclarative questions from prop-er declaratives. The automaticclassification of question formand function is explored. Proficiency. Intonational featuresincluding syllable prominence,pitch accent, and boundary toneswere found to correlate with language proficiency assessmentscores at a strength equal to thatof traditional fluency metrics. The combination of all prosodicfeatures further increased correlation strength, indicatingthat suprasegmental informationencodes different aspects ofcommunicative competence.

Michael E.Locasto IntegrityPostures forSoftware Self-Defense

Software current-ly lacks the capability to respondintelligently and automatically tohacking attempts in a way thatpreserves both its availability and integrity. Unfortunately, mostcurrent software protection tech-niques typically abort a processafter an intrusion attempt or afault violates the integrity of anapplication’s data or executionartifacts, thereby transforming an

arbitrary exploit or failure into a self-induced denial of serviceattack—resulting in a compromiseof both the application’s integrityand its availability. Althoughmany consider this approachsafe, it is unappealing becausesystems remain susceptible tothe original fault upon restart andrisk losing accumulated state. What is needed, then, are mech-anisms that preserve both theinternal and external integritypostures of a software applica-tion so that it can continue toprovide service in the face of a new attack or similar futureinstances of that attack.Preserving external integrity can prevent the recurrence of aparticular fault or vulnerability,and it can be accomplished byfiltering input that triggers orexercises the underlying error.Preserving internal integrity can be accomplished by makinginformed, targeted changes to a program’s state in case a newattack is able to bypass theexternal integrity monitor. In order to support the mainte-nance of both internal and externalintegrity postures, I introduce the notion of “policy-constrainedspeculative execution”, where thepolicy in question is an “integrityrepair policy” based on my exten-sions to the Clark-Wilson IntegrityModel. Speculatively executingslices of an application, or“microspeculation”, is similar tohardware speculative execution,except that acceptance of a particular slice of execution ispredicated on the integrity repairpolicy rather than the result of abranch conditional. These policiescontain a collection of constraintsthat my system leverages torestore an internal integrity posture while an attack or faultoccurs. Repair policy effectivelycustomizes a software system’sresponse to attack. To demonstrate the feasibility andeffectiveness of this approach toself-healing, I created a runtimeenvironment, STEM, that super-vises an application’s executionand enforces integrity repair policies in both binary-only andsource-available environments.

In order to preserve the externalintegrity posture of a softwaresystem, I combined STEM with anexisting network content anomalydetector to create FLIPS, a sys-tem capable of automatic exploitsignature generation. No systemsupplies perfect security, but we can provide intelligent, well-formed recovery mechanisms andautomatically invoke them. Anintegrity repair model assists inbridging the gap between currentsystems and systems that canautomatically self-heal.

Hassan Malik EfficientAlgorithms forClustering HighDimensionalData usingInterestingPatterns

Recent advances in data miningallow for exploiting patterns as primary means for clustering largecollections of data. In this thesis,we present three advances in pat-tern-based clustering technology,and a related advance in patternfrequency counting. In our firstcontribution, we analyze numerousdeficiencies with traditional pat-tern significance measures suchas support and confidence, andpropose a web image clusteringalgorithm that uses an objectiveinterestingness measure to identifysignificant patterns, yielding meas-urably better clustering quality. In our second contribution, weintroduce the notion of closedinteresting itemsets, and showthat these itemsets provide sig-nificant dimensionality reductionover frequent and closed frequentitemsets. We propose GPHC, asub-linearly scalable global pat-tern-based hierarchical clusteringalgorithm that uses closed inter-esting itemsets, and show thatthis algorithm achieves up to 11%better FScores and up to 5 timesbetter Entropies as compared tostate of the art of agglomerative,partitioning-based, and pattern-based hierarchical clusteringalgorithms on 9 common datasets. Our third contribution addressesproblems associated with using

globally significant patterns forclustering. We propose a pattern-based hierarchical clusteringalgorithm that builds the clusterhierarchy without mining for globally significant patterns. Thisalgorithm allows each instance to“vote” for its representative size-2patterns in a way that ensures aneffective balance between localand global pattern significance,produces more descriptive clusterlabels, and allows a more flexiblesoft clustering scheme. Withresults of experiments performedon 40 standard datasets, we showthat this algorithm almost alwaysoutperforms state of the art hier-archical clustering algorithms andachieves up to 15 times betterEntropies, without requiring anytuning of parameter values, evenon highly correlated datasets. Our final contribution deals withthe problem of finding a datasetrepresentation that both offers agood space-time tradeoff for fastsupport (i.e., frequency) counting,and also automatically identifiestransactions that contain thequery itemset. We compare FPTrees, and Compressed PatriciaTries against several novel vari-ants of vertical bit vectors. Wecompress vertical bit vectorsusing WAH encoding and showthat simple lexicographic orderingmay outperform the Gray coderank-based transaction reorderingscheme in terms of RLE compres-sion. These observations lead us to propose HDO, a novelHamming-distance-based greedytransaction reordering scheme,and aHDO, a linear-time approxi-mation to HDO. We presentresults of experiments performedon 15 common datasets with vary-ing degrees of sparseness, andshow that HDO- reordered, WAHencoded bit vectors may take aslittle as 5% of the uncompressedspace, while aHDO achieves similar compression on sparsedatasets. With results from over109 database and data miningstyle frequency query executions,we show that bitmap-basedapproaches result in up to 102times faster support counting, andthat HDO-WAH encoded bitmapsoffer the best space-time tradeoff.

duce a notion of a shallowsemantic network based on therelations between the importantelements discovered in thesedocuments. This shallow semanticnetwork captures the mostimportant relations among theobjects, people, and other elements that are involved in theevents described in the inputdocument collection. We presentexperimental evidence that sucha relation-based representationof event-focused documents issuperior to techniques that relyon term frequencies for the taskof information selection.For a set of document collectionsdescribing similar events withinthe same domain, we design andimplement a completely automatic,data-driven procedure for inducingdomain templates. These domaintemplates capture facts that are important for most domaininstances. We then devise a pro-cedure for identifying commonal-ities across different subdomains.We experiment with a specialcase of a domain, a biographydomain, and identify commonali-ties across activities used fordescriptions of people belongingto different occupations. We also propose a methodology forcreating domain hierarchies. We apply our methods for identi-fying relations to the question-answering task. We design and implement a two-prongedapproach for answering open-ended event-related questions.The first approach relies on automatically created domaintemplates and is used when theevent mentioned can be identifiedas one of a particular class of events (e.g., earthquakes,presidential elections, etc.). The second, complementaryapproach is based on a shallowsemantic network, which weextract from the documents relevant to the question. We also suggest a formal modelfor efficient information packagingthat is based on mapping theinformation selection task on theset cover complexity problem.Using this mapping, we outlineand implement informationselection algorithms that are

provably optimal polynomial-timeapproximations for informationselection tasks that have a limiton the output size.

Kshitiz Garg Rain in Visionand Graphics

Rain producessharp intensityfluctuations inimages and

videos which severely degradethe performance of outdoor visionsystems. Removing the visualeffects of rain is therefore impor-tant in order to make outdoorvision robust to rain. In contrast, ingraphics, rain effects are desirableas they are often used to conveyscene emotions in movies and inother graphics applications, suchas games, to enhance realism.Developing efficient algorithms forrealistic rendering of rain effects ishowever challenging. This thesis,studies rain from the perspectiveof vision and graphics. We havetwo main goals (1) to understandthe visual appearance of rain and(2) to develop efficient algorithmsboth for handling its effects invision and for its realistic render-ing in graphics. We begin by modeling the visualappearance of rain. We describethe appearance by modeling thephotometric intensities producedby individual raindrops and thespace-time correlation producedby the motion of a large numberof raindrops. We then use theseappearance models to developtwo complementary approachesto handle the effects of rain onvision systems. The first approachuses a post-processing algorithmfor detection and removal of rainfrom videos that have alreadybeen captured. In the secondapproach we use our appearancemodel to show that the uniquephysical properties of rain its smallsize, high velocity and spatial distribution makes its visibilitydepend strongly on cameraparameters. This dependence isused to reduce the visibility ofrain during image acquisition byjudiciously selecting cameraparameters. Conversely, camera

parameters can also be chosento enhance the visibility of rain.This ability can be used to developan inexpensive and portablecamera-based rain gauge thatprovides instantaneous rain-ratemeasurements.In graphics, we have developedtwo models for realistic renderingof rain. In the first model we consider the complex appearancepatterns produced by close byraindrops. We show that thesepatterns are produced due to therapid shape distortions (i.e. oscil-lations) that a raindrop under-goes as it falls and develop aphysics-based model to faithfullyrender these complex appearancepatterns. This model is then usedto develop an efficient algorithmfor rendering rain in images andvideos. In the second work, wefocus on realistic rendering ofmaterial-dependent splashing of raindrops. Here we measureraindrop splashes for 22 differentreal world materials. Thesemeasurements are then used to build a compact stochasticmodel that accurately capturesthe material and inclinationdependent splashing behavior ofraindrops. Additionally, it allowsuser to render physically plausiblesplashes for novel materials thathave not been measured.

Cheoljoo JeongOptimizationTechniques for RobustAsynchronousThresholdNetworks

Digital systems design is chang-ing rapidly as integrated circuittechnology evolves. The arrivalof deep submicron technologybrings about increasing designchallenges in terms of obtainingcircuit reliability, increasing needfor low-power design, and thehandling of system complexity. It is critical to cope with thesechallenges in both circuit designand in automated computer-aided design (CAD) flows. One attractive alternative is touse robust asynchronous circuitswhich gracefully address some

of these problems. Asynchronouscircuits communicate throughhandshaking in a distributed wayrather than by adopting a central-ized control such as the globalclock. Asynchronous design hasbeen the focus of renewed interestand research activity because ofits potential benefits of low powerconsumption, low electromagneticinterference, robustness to delayvariations, and modularity ofdesigns. However, they typicallysuffer from high area and latencyoverhead, as well as lack of CADtools and optimization techniques.Also, careful design is requiredsince hazard-freedom is essentialfor their correct operation. This thesis presents optimizationtechniques for a class of asyn-chronous circuits, called asyn-chronous threshold networks,which is one of the most robustasynchronous circuit styles. Twotechniques are proposed: (i) tech-nology mapping and cell merger,and (ii) relaxation. The technologymapping and cell merger algorithmoptimizes the circuits using a specific technology library whilethe relaxation technique relaxesthe overly-restrictive style of thecircuits. Both of these techniquesoptimize area or delay of the circuits while still fully-preservingtheir robustness properties. Also, the proposed algorithmswere implemented in automatedCAD tools. Experiments wereperformed on industrial circuitsand largest MCNC benchmarkswith significant improvement inboth area and delay.

JacksonLiscombe Prosody andSpeaker State:Paralinguistics,Pragmatics, and Proficiency

Prosody—suprasegmental char-acteristics of speech such aspitch, rhythm, and loudness—isa rich source of information inspoken language and can tell alistener much about the internalstate of a speaker. This thesisexplores the role of prosody inconveying three very different

ACROSS

2 Detects intruders

4 Architect who reuses components

6 Can summarize the newsinto a few sentences

7 Augments reality

9 Director of the programmingsystems laboratory

13 Distinguished faculty teacher

17 Founded cricket website

19 Can see through raindrops

21 Our main weapon on the front line in 1004

24 Uses CS to study disease

25 Knuth prize winner

26 Working with NASA to send chips into space

27 Wrote the book on compilers

29 Wrote the book onCombinatorial Methodswith Computer Applications

30 Wrote the book on firewalls

31 Young investigator award winner

DOWN

1 He’s a quantum computer

3 A good listener

5 Computationally learns, at least theoretically

8 Can teach you to measure biometrics

10 Internet phone pioneer

11 Makes your queries fast

12 Web searcher

14 Likes to be embedded

15 Physical simulator

16 A strong man

18 Lends a hand to surgeons

19 Advised on Microsoftantitrust case

20 Is discriminative and generative

22 Wrote the book on complexity and information

23 Advanced cryptographer

28 Has SMARTS

20 CS@CU FALL 2007 CS@CU FALL 2007 21

Recent & Upcoming PhD Defenses (continued)Cristian SovianiHigh-LevelSynthesis for PacketProcessingPipelines

Packet process-ing pipelines are essential compo-nents of state-of-the-art networkrouters and switches. Theirdesign methodology is a delicatetrade-off between, on one hand,the desire of abstract specifica-tions, short development time, and design maintainability and, onthe other hand, a very aggressiveperformance requirement. The present thesis proposes acoherent design methodology for packet processing pipelines.Following systematically thedesign flow, it starts with the highlevel specification of the pipeline,where a domain specific languageis introduced. The next step is the RTL-synthesis of the highlevel model and the worst casethroughput analysis of the synthesized pipeline, as well as a FIFO sizing technique. Finally, I address some characteristicRTL-level optimization issues. I claim, based on experimentalresults, that my proposed tech-niques can dramatically improvethe design process of thesepipelines, while the resultingperformance matches the expec-tations of a hand-crafted design. The considered pipelines exhibit apseudo-linear topology, which canbe too restrictive in the generalcase. However, especially due toits high performance, such archi-tecture may be suitable to applica-tions outside packet processing, in which case some proposedtechniques can be easily adapted. As I ran my experiments onFPGA platforms, this work has an inherent bias towards thistechnology; however, most resultsare technology independent.

Jia Zeng PartialEvaluation forCode Generationfrom Domain-SpecificLanguages

Partial evaluation has beenapplied to compiler optimizationand generation for decades.Most of the successful partialevaluators have been designedfor general-purpose languages.Our observation is that domain-specific languages are also suit-able targets for partial evaluation.The unusual computational modelsin many DSLs bring challenges aswell as optimization opportunitiesto the compiler. To achieve an interesting level ofoptimization, partial evaluation hasto be specialized to fit the specificparadigm. In this dissertation, wepresent three such specializedpartial evaluation techniquesdesigned for specific languagesthat address a variety of compila-tion concerns. The first algorithmprovides a low-cost solution forsimulating concurrency on a sin-gle-threaded processor. The sec-ond enables a compiler to compilemodest-sized synchronous programs in pieces that involvecommunication cycles. The thirdstatically elaborates recursivefunction calls that enable programmers to dynamically create a system’s concurrentcomponents in a convenient andalgorithmic way. Our goal is todemonstrate the potential of partialevaluation to solve challengingissues in code generation fordomain-specific languages. Naturally, we do not cover all DSLcompilation issues. We hope ourwork will enlighten and encouragefuture research on the applicationof partial evaluation to this area.

Crossword Puzzle

CS Professors

Two new photos are included for this issue’scaption contest. Please send suggestions tonewsletter-editors@lists.cs.columbia.edu.The winners will be announced in the springnewsletter. In case you have trouble makingout the faces, the second image includesProfessors Malkin, Misra, Schulzrinne, Nayar,and Kender.

CS@CU FALL 2007 23

Caption Contest

22 CS@CU FALL 2007

The winner of Spring 2007’s caption contest

for the picture of Steve Bellovin and AngelosKeromytis (shown left) was Michael Rabin,with the caption:

“Have you heard, Zvi Galil gave up Tel Avivand is returning to the CS department!”

PHOTO 1 PHOTO 2