<<

NEWSLETTER OF THE DEPARTMENT OF AT COLUMBIA UNIVERSITY VOL.11 NO.1 SPRING 2015

ealth and predisposi- statistics, and signal process- tion to disease strongly ing are needed to detect the CUCS Welcomes Hdepend on the genetic subtle variations and statistical material carried deep within the patterns that reveal traits and cells of every individual. Increas- predisposition to disease.” Two New Faculty ingly this genetic information, unique to every individual, is be- Erlich is well-positioned for coming public. For Yaniv Erlich, untangling the complex genetic it’s two parts of the complex underpinnings of human biol- relationship between genes and ogy. An initial love of math and a health, one increasingly under- friend’s chance remark that biol- stood through quantitative analy- ogy entailed a lot of mathemat- sis and computer algorithms. ics piqued his interest, and he went on to study both biology “Computational methods are and genetics, two sciences in- necessary at every step in creasingly awash in data and in examining the genome,” says need of algorithms for inferring YANIV ERLICH Erlich, assistant professor of information from that data. He ASSISTANT PROFESSOR computer science. “The strings approached both subjects from OF COMPUTER SCIENCE of nucleotides (A, T, G, and C) a computational perspective. that form each person’s unique Dissecting the Complex genome are billions of letters At MIT’s Whitehead Institute of Relationships of Genes, long. It’s not even possible to Biomedical Research, where he Health, and Privacy look at these long sequences headed a research lab from 2010 without a computer. As more through 2014, he had the chance genetic data is collected, con- to develop and then apply new cepts from machine learning, computational tools to genetics.

CS@CU SPRING 2015 1 Cover Story (continued) Student Awards

Results were impressive. onsider the challenge Blei’s interest in data science One method sequenced tens of the modern-day and machine learning was rein- Columbia Engineering Team of thousands of samples at Cresearcher: Potentially forced by the work of Herbert a time. Another harnessed millions of pages of information Robbins, a former professor signal processing and statisti- dating back hundreds of years of mathematical statistics at Finds Thousands of cal learning to extract genetic are available to be read from a Columbia. “Robbins developed information from short tandem computer screen. How does a an algorithm that allows the repeats, or STRs, a fast-mutat- simple Internet search deliver machine learning community to Secret Keys in Android Apps ing fragment of DNA so small appropriate findings? scale up algorithms to massive it had been mostly neglected data,” says Blei, who predicts by the research community. The answer is topic modeling, that one of the next break- Both methods contributed new DAVID BLEI a mathematical approach to throughs in modern machine information about how genes PROFESSOR OF COMPUTER uncovering the hidden themes in learning and statistics will SCIENCE AND STATISTICS collections of documents. David operate at the molecular level. involve observational data, or In a paper presented—and about the content in Google Blei, professor of computer that which is observed but not But Erlich was also interested Embracing the science and statistics, led the awarded the prestigious Ken Play, including a critical security Science of Teaching collected as part of a carefully Sevcik Outstanding Student problem: developers often store in how genes affect health groundbreaking research that re- controlled experiment. and traits. Is the relation- Topics to Machines sulted in the development of the Paper Award—at the ACM their secret keys in their apps ship linear where mutations Latent Dirichlet Allocation (LDA) “Previously, machine learn- SIGMETRICS conference on software, similar to usernames/ predictably sum to a trait, or model, a topic model capable of ing has focused on predic- June 18, Jason Nieh, professor passwords info, and these can is the relationship nonlinear, exploiting hidden topics and se- tion; observational data has of computer science at Columbia be then used by anyone to with mutations interacting genetic information donated mantic themes among countless been difficult to understand Engineering, and PhD candidate maliciously steal user data or unpredictably with one an- by research participants and lines in billions of documents and exploring that data was Nicolas Viennot reported that resources from service providers other? The answer required a cross-reference it to online data via language processing. By a fuzzily defined activity,” Blei they have discovered a crucial such as Amazon and Facebook. The researchers used PlayDrone to recover app sources and, in the process, uncovered crucial population-scale analysis, one sources. Using only Internet co-developing the LDA, Blei ef- explains. “Now, massive collec- Professor of Computer security problem in Google Play, These vulnerabilities can affect Science Jason Nieh security flaws. too large to be constructed searches with no actual DNA, fectively influenced the research tions of observational data are the official Android app store. users even if they are not actively using traditional data collection Erlich and his research team and development of topic mod- everywhere—in government, running the Android apps. Nieh “Google Play has more than notes that even “Top Develop- methods. Here, Erlich and his were able to correlate the els and machine learning. industry, natural science, and one million apps and over 50 Whitehead colleagues came social science—and practitio- ers,” designated by the Google donated DNA to a surname in “We were fortunate that LDA billion app downloads, but no Other findings of the research up with an innovative solution. 13% of the U.S. population, a ners need to be able to quickly Play team as the best developers include: has become influential for ana- one reviews what gets put on Google Play, included these Using genealogical data already result that surprised even Erlich. explore, understand, visualize, into Google Play—anyone can • Showing that roughly a uploaded to social media sites, lyzing all kinds of digital informa- and summarize them. But to do vulnerabilities in their apps. “Our study highlighted current tion, including text, images, get a $25 account and upload quarter of all Google Play free they created a genealogical tree so, we need new statistical and whatever they want. Very little “We’ve been working closely apps are clones: these apps of 13 million individuals dating gaps in genetic privacy as we user data and recommendation machine learning tools that can enter to the brave new era of systems, social networks, and is known about what’s there at with Google, Amazon, Face- are duplicative of other apps back to the 15th century. Such a help reveal what such data sets an aggregate level,” says Nieh, book, and other service already in Google Play. deep genealogy reveals clusters ubiquitous genetic information,” survey data,” says Blei. “Even say about the real world.” who is also a member of the providers to identify and notify • Identifying a performance of genetic variation, with some explains Erlich. “However, we in population genetics, LDA-like ideas were developed indepen- Prior to joining Columbia Engi- University’s Institute for Data customers at risk, and make problem resulting in very slow tied to longevity and some to must remember that sharing PhD Candidate Sciences and Engineering’s the Google Play store a safer app purchases in Google Play: rare disorders. (And offering as genetic information is crucial to dently and are used to uncover neering in July of 2014, Blei was Nicolas Viennot understanding the hereditary patterns in collections of indi- an associate professor in the Cybersecurity Center. “Given place,” says Viennot. “Google this has since been fixed. a side bonus an intriguing view the huge popularity of Google is now using our techniques to into human migration.) Looking basis of devastating disorders. viduals’ gene sequences.” Department of Computer Sci- • A list of the top 10 most ence at Princeton University. He Play and the potential risks to proactively scan apps for these highly rated apps and the top at how mutations ripple through We were pleased to see that Blei, whose challenges and is a faculty affiliate of Colum- millions of users, we thought it problems to prevent this from 10 worst rated apps in Google populations, researchers can our work has helped facilitate rewards both come from solv- bia’s Institute for Data Sciences was important to take a close happening again in the future.” Play that included surprises measure the frequency of a discussions and procedures to ing large-scale machine learning and Engineering. Blei is a re- look at Google Play content.” such as an app that, while the certain trait, and thus evaluate better share genetic information problems, appreciates the In fact, Nieh adds, developers in ways that respect partici- cipient of the 2013 ACM-Infosys worst rated, still had more the genetic contribution. With interconnectivity of disciplines Nieh and Viennot’s paper is the are already receiving notifications Foundation Award in Comput- than a million downloads: it a larger tree, even more will pants’ preferences.” within his field. first to make a large-scale mea- from Google to fix their apps and ing Sciences and is a winner purports to be a scale that become possible. Erlich hopes that ensuring better surement of the huge Google remove the secret keys. “Columbia has an extremely of an NSF CAREER Award; Play marketplace. To do this, they measures the weight of an The promise is great but it all safeguarding of genetic informa- strong faculty in machine learn- an Alfred P. Sloan Fellowship, Nieh and Viennot expect Play- object placed on the touch- tion will encourage more people developed PlayDrone, a tool that means nothing if there isn’t ing and statistics, and across and the NSF Presidential Early uses various hacking techniques Drone to lay a foundation for screen of an Android device, genetic material to work with. to contribute their DNA, speed- other departments, such as in Career Award for Scientists and new kinds of analysis of Android but instead displays a random ing the day when personalized to circumvent Google security to And that requires trust. medicine and neuroscience. Engineers. His accolades also successfully download Google apps. “Big data is increasingly number for the weight. medicine becomes a reality for This makes it an ideal university include the Office of Naval Re- important and Android apps People today are more wary Play apps and recover their Good news for the hundreds all. For those most at risk of rare for my group to pursue our in- search Young Investigator Award are just one form of interesting about revealing personal data, sources. PlayDrone scales by of thousands of developers genetic diseases, the bigger terdisciplinary research agenda,” and the New York Academy of data,” Nieh observes. “Our work and they are right to be so. simply adding more servers and who upload content to Google danger may be in not contribut- he says. “I like working on inter- Sciences Blavatnik Award for makes it possible to analyze Erlich himself is one of the first is fast enough to crawl Google Play and even more so for the ing genetic information. esting applied problems about Young Scientists. Android apps at large scale in to flag the ethical complexities Play on a daily basis, download- millions of users who download BSc, Tel-Aviv University, 2006; data—in the sciences, social new ways, and we expect that involving genetics and privacy. A BSc, Brown University, 1997; ing more than 1.1 million Android the content! PhD, Watson School of Biological Sciences, sciences, and humanities—and PhD, University of California Berkeley, 2004 apps and decompiling over PlayDrone will be a useful tool to paper he spearheaded, released Cold Spring Harbor Laboratory, 2010 better understand Android apps in January 2013, caused a stir letting real-world problems drive 880,000 free applications. the methodological advances and improve the quality of appli- by showing how easy it is to By Holly Evarts that we develop.” By Dave Meyers Nieh and Viennot discovered cation content in Google Play.” take apparently anonymized By Linda Crane all kinds of new information

2 CS@CU SPRING 2015 CS@CU SPRING 2015 3 Feature Articles

A small line becomes a page of A dozen years ago (in 2001) random characters. The original Dan Boneh working with other Chasing Real Security code becomes bit by bit multi- people discovered how to build plication matrices. And this is crypto systems using bilinear An interview with Allison Bishop Lewko one source of inefficiency. maps on elliptic curves (where Actually for this first candidate the set of elements used to it’s not the original software itself satisfy an equation consists of being put through the obfusca- points on a curve rather than tion, but the software trans- integers). Elliptic curves give formed into a circuit (a bunch of you a little more structure than a AND, OR, NOT gates), and this is single function; some things are Indistinguishability obfuscation is achieved when a software circuit is obfuscated in such a way that two different obfuscations are indistinguishable. can take a software program and this added requirement that the matrices and multiplying them. another source of inefficiency. easy to compute but there are put it through the mathemati- encrypted software still has to things that are hard to compute. cal structure provided by these work, that it still performs the Is having two matrices Indistinguishability The same hard problem has dif- multilinear maps and encrypt function it was written to do. So for every slot a way to obfuscation is said to ferent levels of structure. software to the machine level I can give you code and you can add randomness? provide a richer structure Think of a group having expo- while preserving its functionality. run it without a secret key but No. The choice of what matrix for cryptographic methods. nents. Instead of telling you the Anyone can take this obfuscated you can’t reverse-engineer the to use is deterministic based What does that mean? process for getting an exponent software and run it without a code to learn how it works. on the input. But we do add All rests on out, I can give you one expo- secret key and without being The bar is much higher for randomness, which is a very im- trying to build up structures nent, which to some extent rep- able to reverse-engineer the obfuscation because obfus- portant part of any cryptography where certain things are fast to resents partial information, and Assistant Professor code. To see what’s happening cated software must remain a scheme since it’s randomness compute, and certain things are each piece of partial information Allison Bishop Lewko in the code requires solving the functional, executable program. that makes it hard to make out hard. The more you can build a lets you do one thing but not hard problem of breaking these Protecting and using at the what’s going on. For this current richer structure of things that another. The same hard prob- Encrypting software so it performs multilinear maps. same time is what is hard. It’s candidate obfuscation scheme, are fast, the more applications lem can be used to generate a function without the code be- This process of encrypting easy to protect software if you we add randomness to the ma- there are that can take advan- encryption keys with different ing intelligible was long thought by software so it still executes, never have to use it again. trices themselves (by multiply- tage of this encryption. What is levels of functionality, and it all many to be impossible. But in 2013, ing them by random matrices), needed is a huge mathematical exist within the same system. Each bit is represented by one of two matrices, with 0 and 1 determining which matrix to use. which we call indistinguishability These concepts of encryption The selected matrices are then multiplied together. Multiplying two matrices not meant to be researchers—unexpectedly and in a and to representations we cre- structure with a rich variety of surprising way—demonstrated how obfuscation, was something and decryption, however, do live This is the beginning of identity- multiplied prevents the software from being correctly reassembled. we didn’t believe was out there ate of the matrices—like these data problems, where some are based security where you give to obfuscate software in a mathemati- very close together and they multilinear maps. cally rigorous way using multilinear even two or three years ago. interact. Encryption is the tool easy, some are hard. different people particular levels maps. This breakthrough, combined Now there is this first plausible used to accomplish obfuscation, But we do it very carefully. In RSA the hard problem is of access to encrypted data. with almost equally heady progress candidate indistinguishabilty ob- and obfuscation can be used to Among all these matrices, there factoring large numbers. One Maybe my hospital sees a great in fully , is fuscator using multilinear maps. build an encryption. are lots of relationships that way—multiplying two prime deal of my data, but my insur- upending the field of cryptography And it’s rigorous. Today there have to be maintained. Adding numbers—is easy; that’s the ance company sees only the and giving researchers for the first are programs that purport to How so? Do you use the randomness has to be done in encryption. Going back—undo- procedures that were done. time an idea of what real, provable same methods to encrypt a precise way so it’s enough to RSA is a single function where going one way is easy and one way is hard. security looks like. The implications obfuscate software. They add ing the factorization—is hard The structure of bilinear maps data to obfuscate code? hide things but not so much to It requires generating both a private and public key at the same time. are immense, but two large problems dead code, change the vari- unless you have the key. But allows for more kinds of keys remain: creating efficiency and provid- ables, things you imagine you Not exactly, though the general change software behavior. it’s hard in the sense that no with more levels of functional- ing a clear mathematical foundation. would get if you gave a sixth- concept of substituting some- There is this tension. The more one has yet figured out how to ity than you would get with a For these reasons, five researchers, grader the assignment to make thing familiar with something functionality to preserve, the factorize large numbers quickly. simple one-way function. It’s including many of those responsible your software still work but hard seemingly random exists in both. more there is to hide, and the That nobody knows how to fac- not much of a stretch to intui- for the recent advances, have formed to read. It is very unprincipled harder it is to balance this ten- tor numbers isn’t proven. We’re tively understand that more is the Center for Encrypted Functionality In the well-known RSA encryp- and people can break it easily. tion method, each character of sion. It’s not an obvious thing always vulnerable to having possible with trilinear maps or with funding provided by NSF. Allison someone come out of nowhere Bishop Lewko is part of this center. In Indistinguishability obfuscation the output text is replaced by to figure out and it’s one reason with multilinear maps. with a new method. this interview she talks about the rap- is something else entirely. a number that gets multiplied why this concept of obfuscation For a dozen years it was an open idly evolving security landscape, the didn’t exist until recently. RSA is what we’ve had since How does obfuscation by itself a number of times (the problem on how to extend bilin- center’s mandate, and her own role. public key) to get a random- Amit Sahai, director of the the 80s. It works reasonably ear maps to multilinear maps in differ from encryption? Richer structures for hard problems expand the applications for cryptography. looking number that only the Center, compares obfuscated well, assuming no one builds a the realm of algebraic geometry, What happened to make Both obfuscation and encryption secret, private key can unlock software to a jigsaw puzzle, quantum computer. but elliptic curves didn’t seem to secure software suddenly protect software, but the nature and make sense of. where the puzzle pieces have to But RSA and other similar public have a more computationally effi- seem possible? of the security guarantee is dif- fit together the right way before key encryption schemes don’t cient structure on which to build. ferent. Encryption tells you your When you obfuscate software, The short answer is that in it’s possible to learn anything work for more complicated During this time, Craig Gentry data or software is safe because each bit of the original program 2013, six researchers—Sanjam about the software. The puzzle policies where you want some (now at IBM) was working on it is unintelligible to anyone who gets associated with some Garg, Craig Gentry, Shai Halevi, pieces are the matrix encoded people to see some data and fully homomorphic encryption, doesn’t have the secret key to number of slots. For every slot Mariana Raykova , Amit Sahai, in the multilinear map, and the some other people to see other which allows someone to per- decrypt it. The software is pro- there are two matrices, and —published a added randomness makes it data. RSA has one hard prob- form computations on encrypted tected but at a cost of not being every input bit determines what paper showing for the first time that much harder to understand lem and allows for one key that data (the results will also be usable. The software is either matrix to use. that it is possible to obfuscate how these pieces fit together. reads everything. There’s no encrypted). Gentry was using encrypted and protected, or it is The result is a huge chain of ma- software in a mathematically nuance. If you know the factors, lattices, a type of mathematical In lattice-based cryptography, encrypted data can be envisioned as points at some random decrypted and readable. trices. To compute the program What does obfuscated distance from a point on a lattice (shown here in two dimensions, though lattices can rigorous way using a new type you know the factors and you structure from number theory, an have hundreds). On decryption, each blue point is attracted to the nearest lattice point. Obfuscation is a guarantee to on an input bit stream requires software look like? of mathematical structure called selecting the corresponding can read everything. entirely different mathematical Homomorphic operations amplify the random displacements. If the noise exceeds a thresh- multilinear maps. This means I protect software, but there is You would not want to see that. world than algebraic geometry. old, some points gravitate to the wrong lattice point, leading to the wrong decryption.

4 CS@CU SPRING 2015 CS@CU SPRING 2015 5 Feature Articles (continued)

Lattices provide a rich family of Is that the purpose of break this process that takes a lattice-based cryptography, and structures, and lattice schemes the Center for Encryption circuit to another circuit. We have the landscape of easy and hard hide things by adding noise. As Functionalities? to have this tantalizing carrot of a problems. How Things Coil operations go on, noise gradually Yes, the center’s main thrusts challenge to gain interest. Every time something is builds up and that’s what makes will be to increase efficiency And on this front we have made broken, it means someone homomorphic encryption hard. and provide a provable math- great progress. We have a came up with an algorithm You might think immediately ematical foundation. Many of somewhat long, highly techni- for something that we didn’t that these lattices schemes the same people responsible cal proof that takes an abstract previously know. Bilinear maps serve as a multilinear map for the big advances in indis- obfuscation for all circuits, were first used to break things; but there isn’t a sharp divide tinguishability obfuscation and and reduces it to a sufficiently then Dan Boneh said if it’s a fast between what’s easy and hard; fully homomorphic encryption simple problem that others can algorithm that can break things, When one sends an email from as a predictive tool for engineer- there is a falloff. With fully are working at the center, and study and try to break. let’s re-appropriate it and use it Boston to Beijing, it travels ing mechanics of thin filaments, homomorphic encryption having building on previous work. to encrypt things. Every new al- through submarine optical ca- rods, and pipes. It’s exciting to That’s the first piece. What we bles that someone had to install think that this computer model this gradual quality, it wasn’t at We’ll be examining every aspect gorithm is a potential new tool. don’t yet have is the second at some point. The positioning can serve both creative and all clear how to use it in a way of the current recipe to increase piece, which is instantiating the The fact that a particular candi- of these cables can generate engineering enterprises.” to build something where mul- efficiency. There are many av- simple problem as a hard prob- date can be broken doesn’t make intriguing coiling patterns that tiplying three things is easy but enues to explore. Maybe there lem in multilinear maps. it implausible that other good can also cause problems if, for Grinspun’s simulation technol- multiplying four things is hard. is a more efficient way of imple- candidates exist. We’ve also instance, they are tangled or ogy, Discrete Elastic Rods, was menting multilinear maps that Have there been any attempts The solution came in 2013 with made progress on working with kinked. The deployment of a originally developed to animate doesn’t involve matrices. Maybe to “break” indistinguishability the publication of Candidate Mul- different kinds of computational rodlike structure onto a moving hair and fur in film and graph- Discrete elastic rods, used for hair simulation here, are also being used to predict only certain parts of a program obfuscation? tilinear Maps from Ideal Lattices, models. In June, we will present substrate is commonly found ics applications, licensed and the coiling of undersea communication cables. have to be encrypted. We’re still which describes this candidate Yes, in the past few months Indistinguishability Obfusca- in a variety of engineering ap- used in Photoshop for realistic early in the process. I’ve been so far describing. The some attacks did beat back on tion for Turing Machines with plications, from the fabrication paintbrushes, and by Weta authors were able to make a The center also has a social some problems contained in Unbounded Memory at STOC, of nanotube serpentines to Digital for use in films such as Cables, being long and slender, the variables that impact the sharp threshold between easy agenda, to both educate others our proof that we thought might which replaces circuits with a the laying of submarine cables The Hobbit and Planet of the were ideal candidates for study. deployment of such cables, and hard problems within lattice on the rapid developments in be hard, but these attacks did Turing machine to conceptually and pipelines, and engineers Apes series. Reis, who is an But could the technology we we can better balance consid- problems. They did it by taking cryptography and bring together not translate into breaks on the show that we can do obfuscation for years have been interested experimental mechanician at built at Columbia Engineering for erations such as expense (the something like a fully homomor- those working on the problem. candidate obfuscation scheme. for varying input sizes. in predicting the mechanics of MIT, was studying how buckling visually striking film and special length of the cable deployed, phic scheme and giving out a While we wished the problems Things are still new; what we’ve filamentary structures and the of thin elastic structures can be effects be sufficiently accurate the amount of time to de- What is your role? partial secret key; one that lets had been hard, the attacks help done this year at the center is coiling process. turned on its head: buckling is to agree with Reis’s hard and ploy the cable), signal quality you do this one thing at level 10 My role is to really narrow down us narrow and refine the class of motivating work in this field. normally feared by engineers as precise experimental data?” (tangled cables can be more but not at level 11. These partial problems we should be focusing A team led by Eitan Grinspun, a potential failure of a design, prone to interference), and the our understanding of what People will be trying to build With support from the National keys, which are really weak se- we’re assuming when we say on, and they are a reminder that the best multilinear maps for associate professor of computer but what if it could be used as resilience of the connection cret keys, add more nuance, but this line between what is hard science at Columbia Engineer- a functional component of a Science Foundation, Reis and (taut cables are more prone to indistinguishability obfuscation the next 20 years regardless Grinspun recruited doctoral stu- it’s extremely tricky. If you give is secure. If someone today and what is not hard is really of what happens with these ing, and Pedro Reis, associate design? The two researchers damage due to external factors, out a weak key, you may make very much still in flux. professor of mechanical engi- decided to investigate how dents Khalid Jawed (MIT) and such as seismic activity.” breaks RSA, we know they’ve particular candidates. Fang Da (Columbia Engineer- things easy that you don’t want factored big numbers, but what The attackers didn’t just break neering and civil and environ- cables are deployed, both at to be easy. You want it to break The ongoing cryptanalysis is an mental engineering at MIT, has the nanoscale, in stretchable ing) to study cable deployment “Translating computer tools from would it take to break indistin- our problem, they broke a important feedback loop that in detail. In their PNAS article, computers and validating them a little but not too badly. This is guishability obfuscation? landscape of problems for cur- been collaborating on a project electronics, and the macroscale, another source of inefficiency. points to targets that attackers that, in exploring these issues, such as the deployment of the researchers describe how against precision model experi- Indistinguishability obfuscation rent multilinear map candidates, will go after to try to break the bridges engineering mechanics internet communication cables seemingly benign decisions, ments has provided a novel tool is still this very complicated, making them fragile to use. How soon before obfuscated scheme, and it will help drive (Reis’s group) and computer on the ocean floor. such as the diameter of a spool, for engineering mechanics to amorphous thing of a system software becomes a reality? Were you surprised that the direction of cryptography. graphics (Grinspun’s group). or the speed at which a cable tackle the design and analysis of with all these different possible A long time. Right now, indis- the problem was broken? It’s really an exciting time to be The researchers combined “This has been a wonderful is deployed, can dramatically other rodlike structures, which avenues of attack. It’s difficult tinguishability obfuscation is a researcher in this field. precision model experiments example of two seemingly un- affect the way that the cable are common in nature and tech- to comprehensively study the Yes, very surprised, though not a proof of concept. Before it with computer simulations and related fields coming together lies on the ground. They created nology,” Reis continues. candidate scheme in its current super-surprised because this can have practical application, examined the mechanics of to address a practical problem a map of the different patterns form and understand what whole thing is very new. There to introducing powerful and “As we move to the next stage, it should have two features: it About The Center for Encrypted coiling, discovering in particular that can arise, from a wiggling makes it secure and to know are always breaks you can’t we would like to pursue engi- should be fast, and we should Functionalities (CEF) that the natural curvature of novel computational tools that meandering mode to steady exactly when we are vulnerable imagine. were not previously available coiling and on to alternating neering problems that combine have a good understanding of The Center’s primary mission is to transform the rod dramatically affects the the mechanics of slender fila- and when we are not. program obfuscation from an art to a rigorous why it is secure, more than Does it shake your faith in in- coiling process. Their study is in our engineering community,” loops, as the spool diameter or mathematical discipline. It supports direct Reis notes. deployment speed are varied. ments with additional ingre- “this uses math and it looks like I did my PhD in topics to expand distinguishability obfuscation? research, organizes retreats and workshops published in the September 29 dients, such as drag, contact, the set of things that we can to bring together researchers, and engages in The researchers also identified we don’t know how to break it.” I have no faith to be shaken. I Early Online edition of Proceed- The collaboration between and friction,” adds Grinspun. provably reduce to nice compu- outreach and educational efforts. factors that have relatively little Indistinguishability obfuscation simply want to know the answer. ings of the National Academy of Grinspun and Reis began when “We are looking, for example, tational assumptions, which is impact on the deployment, currently doesn’t meet either Sciences (PNAS). Reis invited Grinspun to visit at locomotion of bacteria, tying very much what is at the heart of We learn something from this among them the height from criteria. It is extremely inefficient his lab at MIT. “We wondered of shoelaces, and hair blowing where this new obfuscation stuff process either way. If the cur- “This was a fun, fruitful col- which a cable is dropped. to what you want to run in real if our seemingly distant worlds in the wind.” is going. We need to take this rent multilinear map candidates laboration,” says Grinspun. “We life—running an instance of a could be bridged by a common “These findings have practical huge, complicated scheme down turn out to be irretrievably bro- did something totally new and This work is funded by a National Science very simple obfuscated function vision,” says Grinspun. “We impacts on our everyday lives,” to a simple computational prob- ken—and we don’t know that different: we took a computer Foundation MoM-IDR Collaborative grant would take seconds or minutes both wanted to understand Reis adds. “Take, for instance, under CMMI (1129894). lem to make it easy for people to they are—we still learn some- algorithm that we had designed instead of microseconds —and how physical objects move by an email that travels along a study. Saying “factor two num- thing amazing. We learn about for Hollywood, and, by teaming we don’t have a good answer for looking at how their geometry, transoceanic communication bers” is a lot easier than saying these mathematical structures. up with Reis’s group, discovered why we believe it’s hard to break. or shape, affects their motion. cable. By better understanding By Holly Evarts

We learn about the structure of GRINSPUN AND EITAN BREANNAN SMITH, JEAN-MARIE AUBRY, DANNY M. KAUFMAN, RASMUS TAMSTORF, IMAGE CREDIT: that this same algorithm served

6 CS@CU SPRING 2015 CS@CU SPRING 2015 7 Feature Articles (continued)

With mathematical models pro- ones being eyeball ISPs such viding a more realistic picture as Time Warner, Comcast, and Net neutrality is all good and fine; of the Internet, the authors AT&T that specialize in the last- use game-theory economics, mile content delivery to paying the real problem is elsewhere. specifically Shapley values, to consumers, and the transit ISPs understand the optimal arrange- such as Level 3 and Cogent that UPDATE: Since this article was posted in November 2014, the FCC voted ment under which each ISP is own the global infrastructure incentivized to contribute to the and provide transit services for (2/26/15) to enforce net neutrality rules that ban paid prioritization. overall good. This methodology other ISPs. Where eyeball ISPs showed how in the early Inter- serve people, transit ISPs serve net, the flow of traffic (mainly the content providers and earn emails and files) was roughly revenue by delivering content symmetrical. A packet originat- to consumers on behalf of the As the debate over net neutrality ing at ISP-A and handed off to content providers. Since transit rages, two papers co-authored by ISP-B for delivery would be bal- ISPs don’t have direct access to Vishal Misra have been recently anced by a packet moving in the consumers, they arrange with reprinted in a special IEEE issue on opposite direction. ISPs often the eyeball ISPs for the last-mile net neutrality. Both papers examine entered into no-cost agree- delivery of content to customers. the economics of the Internet. The ments to carry one another’s ISPs are cut off from the grow- first (from 2008) looks at how profit- traffic, each figuring that the ing revenue source happen- motivated decisions by the ISPs work amount of traffic it carried for ing outside the Internet. The The Internet has evolved from a simple, symmetric network where light email and web traffic against customers’ preferences for another ISP would be matched fast Internet speeds at low prices; it resulting imbalance in profit flowed between academics and researchers and the only revenues were from selling bandwidth, by that other ISP carrying its to an asymmetric one where traffic flows from content providers to consumers, generating also predicted the rise of paid peering, sharing reduces the investment own traffic. Network neutrality massive revenues for content providers. Associate Professor which plays a central role in today’s incentives for ISPs to build out prevailed naturally since ISPs, Vishal Misra net neutrality debate. The second the network. (from 2012) shows how the introduc- compensated for bandwidth tion of a public-option ISP works to used, did not differentiate one According to Misra, “In a sumers, there is another way. Content providers like Netflix align the interests of both customers packet from another. The more coalition where one participant Charge higher prices by creating are in a much weaker position and ISPs, in the process achieving a packets of any kind, the more doesn’t feel adequately compen- a premium class of service than the eyeball ISPs. Content free, open Internet without regulation. profits for all ISPs, an economic sated or at least compensated with faster speeds. Creating a providers need ISPs much more situation that aligned nicely with to the same degree as other two-class system is currently than the ISPs need them. If Understanding the econom- customers’ interest in having an participants, this gives rise to prohibited by network neutral- Netflix were to disappear, other ics of the Internet is crucial expanding supply of bandwidth. conflict. Mathematically, we call ity; however the FCC proposed streaming services would rush to understanding how best the situation being outside the in May new rules that would to fill in the gap. For the ISPs, it to maintain a free and open Internet economics have core, where some participants allow such a two-tier Internet, matters little whether it’s Ama- Internet—the stated goal of net The Internet is operated by thousands of interconnected ISPs; each requires the cooperation changed considerably in recent of other ISPs to provide Internet services. have the incentive to leave the effectively ending at least one zon, Hulu, or another service neutrality. years with the rise of behemoth coalition since they will be better aspect of net neutrality: pro- (and worryingly, services run by for-profit content providers such off without it. In schoolyard par- hibiting the discriminations on the eyeball ISPs themselves) The connection between as Facebook, Google, Amazon, lance, it is the equivalent of tak- delivering different contents on providing streaming services. Internet economics and net School of Computing, National This prediction made in 2008 is and Netflix. Their modeling ing your ball and going home.“ the network. neutrality is not well under- University of Singapore) along now coming true and is central shows the appearance of these If ISPs don’t need Netflix, nei- with collaborators Dah-Ming stood, partly because net to today’s heated debate on net for-profit content providers did With an imbalance in the direc- ther do customers. Customers Chiu and John Lui in Hong Kong neutrality is considered more neutrality. This paper (one of two things: it changed the Inter- tion of traffic and no mechanism An imbalance of power unhappy with Netflix’s service set out in 2008 to realistically of a legal and regulatory issue, two co-authored by Misra) was net from a symmetric network for appropriate compensation, can simply choose another such model the way Internet service The eyeball ISPs, however, are and partly because the Internet, republished by the IEEE in a to an asymmetric one where the previous no-cost (or zero- service. They can’t, however, providers (ISPs) interact with in a power position because, a vast maze of inter-connected special net-neutrality issue that the vast preponderance of traf- dollar) bilateral arrangements normally choose a different ISP. one another, looking particu- unlike transit ISPs, eyeball ISPs networks, is hugely complex. appeared in June 2014. fic flows from content providers broke down and were replaced larly at the pricing policies used have essentially no competi- The monopolistic power of Economists themselves when to customers. And it introduced by paid-peering arrangements by ISPs to compensate one tion. Misra cites data released the eyeball ISPs may soon be analyzing Internet economics From symmetry a new revenue stream, one out- where ISPs pay one another to another for transit services. by FCC chairman Tom Wheeler made stronger. The proposed fail to take into account the to asymmetry side the Internet and generated carry one another’s traffic. Each Their findings are described showing that more than 19 Comcast takeover of Time complexity and instead rely on by advertising, online merchan- ISP adopts its pricing policies in On Cooperative Settlement One main goal of the paper was percent of US residents have Warner would create a single abstractions that simplify traffic dizing, or payments for gaming, to maximize profit, and these Between Content, Transit and to more realistically model the no broadband provider offering company that controls more flow, routing decisions, and the streaming video, and financial pricing policies play a role in how Eyeball Internet Service Provid- interactions among ISPs. Econo- 25Mbps service, and another than 40% of the US high-speed impact of congestion, and as and other services. (Revenues ISPs cooperate with one another, ers, which used sophisticated mists have typically analyzed 55 percent have only one such Internet service market. Anti- a consequence fail to account from online video services or don’t cooperate. Profit-seeking mathematical modeling to ac- and modeled the Internet as a provider. The ISPs that provide trust legislation would seem for how the profit-driven deci- alone grew 175% from $1.86 to and cost-reduction objectives curately depict the complex in- two-sided market, with consum- decent-speed service, for all in- to prohibit such a takeover, but sions negatively affect Internet $5.12 billion between 2010 and often induce selfish behaviors teractions among different ISPs. ers on one side and content pro- tents and purposes are monop- both companies argue that they service for customers. 2013. Source: FCC NPRM on in routing—ISPs will avoid links Originally published in 2008, the viders on the other, and a single olies. While wireless Internet don’t compete. In this they are the Open Internet.) considered too expensive for To bring a more networking paper obtained new insights ISP sitting in the middle serving (including Wi-Fi) is often cited correct. The markets in which example—thus contributing to angle to the economic analysis, into ISPs’ routing decisions them both. The reality even in The ISPs themselves were by ISPs as competition, the Time Warner and Comcast oper- Internet inefficiencies. Vishal Misra with his colleague and the settlement (payment) 2008 was far more complex, changing and becoming more comparison is not fair. Unwired ate do not overlap. Seemingly Dan Rubenstein and their among different ISPs, showing with packets traveling through specialized. The paper subdivides Paid-peering is one ISP strategy Internet is not fast as the wired by design, Comcast operates in student Richard T. B. Ma (now the inevitability of paid peering several ISPs before reaching the ISPs into different but overlap- to gain profits. For the eyeball service that is available only San Francisco and Time Warner Assistant Professor with the and the incentives supporting it. intended destination. ping conceptual types, the main ISPs that control access to con- through the eyeball ISPs. in Los Angeles, while Time War-

8 CS@CU SPRING 2015 CS@CU SPRING 2015 9 Feature Articles (continued) ner operates in New York City It would have been relatively over-utilized to the point they Netflix-Comcast deal, Netflix’s are fast, all with zero network and Comcast in Philadelphia. inexpensive for Comcast to cannot accept additional traffic content is being moved along neutrality regulations. Zero. Comcast Chairman and CEO add capacity. But why should and start dropping packets. The a private connection and never The pattern holds in the US. Brian L. Roberts uses this argu- it? Monopolies such as Com- interesting point is that blocked reaches the global Internet. (Tom Susan Crawford uses the ment to support his company’s cast have no real incentive to interchanges occur only in Wheeler, chairman of the FCC, example of Santa Monica’s takeover of Time Warner, but upgrade their networks. There those markets monopolized or has acknowledged that these 10-Gbps broadband service Misra sees it differently. “Time is in fact an incentive to not up- dominated by a single eyeball interconnection arrangements City Net, which charges busi- Warner and Comcast claim grade since a limited commod- ISP and nowhere else. Internet do not fall within the scope of nesses a third of what a private they are not killing competition ity commands a higher price exchange points, which are es- any net-neutrality protections.) operator would charge. (A because they don’t compete in than a bountiful one. By limiting sentially line cards on routers, According to Misra, net neu- second, important benefit of the same markets. That is not bandwidth, Comcast can force are not expensive; at a cost trality loses its meaning and municipal broadband is ensuring an argument; that is a question. Netflix and other providers to of $30,000, they represent a becomes irrelevant when ISPs Internet access to low-income Why aren’t they competing?” pay more or opt into the pre- rounding error to Comcast’s and content providers arrange families and those in rural areas, mium class. daily profits. Adding new inter- Occupying a position of power private pathways that avoid the populations not well-served by changes would easily accommo- and knowing customers are While the eyeball ISPs such as global links. “Technically you for-profit companies.) date increased data traffic from stuck, the eyeball ISPs can and Comcast and Time Warner are restrict fast lanes on a public Netflix or any other content pro- So why aren’t other US cities do play hardball with content monopolistic, the transit ISPs highway, but you are freely vider, but for profit reasons such doing more to promote mu- providers. This was effectively are not and must compete with allowing private highways for upgrades are resisted. (For more nicipal broadband? Because the illustrated in the recent Netflix- one another. Any transit ISP paying content providers. The A backdoor around net neutrality. Netflix paid Comcast to connect directly to its servers, bypassing about the perspective from the global transit where net neutrality is currently enforced (and bypassing also the transit ISPs). powerful ISPs and their lobby- vs-Comcast standoff when that charges higher prices for effect is much the same.” transit side, see blog.level3. ists have persuaded politicians Comcast demanded Netflix pay the same global connectivity com/open-internet/observations- In the face of the monopolistic in 19 states to enact legislation additional charges (above what as another ISP is going to lose internet-middleman.) power of ISPs and their stiff to prohibit public-funded mu- Netflix was already paying for business. As reported by Vox. regulated by net neutrality as public-option ISP changes the resistance to regulation (which nicipal networks, this despite bandwidth). When Netflix ini- com (from data collected from Besides charging more for fast well as a scenario that intro- equation. Under a scenario that they are increasingly able to the 1996 Telecommunications tially refused, Comcast custom- DrPeering.com), average transit Internet lanes, ISPs have other duces a public-option ISP. includes a public-option ISP, prices have fallen by a factor of avoid in any case) and with profits become tied to market Act, which prevents states from ers with Netflix service started ways to extract revenues from The mathematical models they 1000 since 1998, a far different perverse incentives to not share, thus incentivizing ISPs blocking any entity looking to reporting download speeds so content providers. What Netflix built more realistically simu- story than what is happening increase bandwidth, the hope to increase bandwidth to make provide telecommunications slow that some customers quit paid for in its deal with Comcast late actual traffic flow on the in the world of monopolistic for maintaining a free and open themselves more attractive services. The same ISPs that Netflix. These speed problems was not a fast lane in the Inter- Internet, taking into account the eyeball ISPs. Internet would seem to be a to customers. If the existing resist regulations needed to en- seemed to resolve themselves net, but a special arrangement complexities of TCP, specifi- lost cause. ISP employs some non-neutral force net neutrality apparently right around the time Netflix If customers notice the eyeball whereby Comcast connects cally how TCP traffic behaves in strategies, consumers have the have no problem with legislation agreed to Comcast’s demands. ISPs’ failure to upgrade ser- directly to Netflix’s servers to congestion. The previous mod- A little competition option to move to a network- or regulation when it suits their (Netflix has since signed a vice through slower download speed up content delivery. It els used by economists had would help neutral public ISP. purposes or maintains their similar deal with Time Warner speeds, transit ISPs notice it is important to note that this assumed open-loop traffic (i.e., monopolistic status. so that Netflix now has special through blocked or over-utilized arrangement is not currently cov- Susan Crawford, a professor the bandwidth demand of the According to Crawford “(lack of) arrangements with the four interchange points. Transit ISPs ered under conventional net neu- of Law who served as Presi- flows ignore the level of con- competition for high-capacity ISPs like any business entity will major ISPs, including AT&T have long noticed that certain trality, which bans fast lanes over dent Barack Obama’s Special gestion) moving through what Internet access is a big issue always seek and find profits in and Verizon.) interface points are regularly the Internet backbone. In the Assistant for Science, Technol- is called in queueing theory an here in the US. A public-option any scenario. The goal of insert- ogy, and Innovation Policy and M/M/1 queue. This seemingly in the form of municipal fiber ing a public-option ISP is not to author of Captive Audience, The sophisticated mathematical is something that I have long deny profits to ISPs but to align Telecom Industry and Monopoly model view ignores that TCP is advocated as a possible solution the profit motives of broadband Power in the New Gilded Age closed loop (meaning it reacts to the issue. Vishal’s work vali- ISPs with customer interests. and The Responsive City, has to congestion) and that TCP dates technically what a lot of By forcing ISPs to compete for extensively researched the is- packets move through a com- us have been arguing for from a customer subscriptions on the sue from a legal and consumer plex network of routers (packets policy perspective.” merits of better service and perspective. She has been a from the same origin might take abundant bandwidth, a munici- This is borne out by real-world vocal proponent of municipal different paths). Closed- and pal ISP does just that. examples in those cities that broadband, arguing that compe- open-loop traffic flows behave have municipal broadband. In It is a view that the FCC is now tition will keep prices in line and very differently under conges- Stockholm where the city laid coming around to embrace. improve services. It’s an argu- tion, and economic models fiber (raising money through In remarks made just recently ment that makes sense and it that fail to account for this and municipal bonds), customers (October 1), FCC Chairman forms the basis for US anti-trust other differences often arrive at have 100 Mbps Internet at Wheeler cited the importance legislation. But will it work? incorrect conclusions about the one quarter the price of what of competition and local choice quality of service delivered by According to Ma and Misra, customers pay in the US for a to increase broadband and the network. it will. In The Public Option: A connection that is 1/10th the Internet speed. More than net- Non-regulatory Alternative to Using more sophisticated speed. And the city makes neutrality, as the mathematical Network Neutrality (originally modeling, Ma and Misra are money. In the UK market, con- modeling by Ma and Misra published 2012) they again use able to show that even under tent providers can choose from shows, it is competition that game-theory economics to tightly regulated net-neutrality, at least four ISPs that compete will ensure both a free and open model how various participants monopolies increase their prof- for their business. It’s a market Internet and that the US re- would act in different scenarios, its when bandwidth is scarce. where there is plenty of con- mains competitive in the global, SOURCE: Netflix.GRAPHIC : reprinted from The Washington Post. Published April 24, 2014. including a monopolistic one However, the introduction of a tent, prices are low, and speeds knowledge-based economy. By Linda Crane

10 CS@CU SPRING 2015 CS@CU SPRING 2015 11 Feature Articles (continued) “ The goal is to build a system for automatically generating a comprehensive description of Automatically Describing Disasters a disaster, one that includes objective, factual An interview with Kathy McKeown information—what is the specific sequence of events of what happened when—with compelling, emotional first-person accounts of people impacted by the disaster.”

seeing other women follow tion they need, both as the same time it could be you. It boards and wherever else we to tie this back to the hurricane One student is now designing a It does seem like an soon after, what’s most helpful disaster unfolds or months or wasn’t possible to ignore what can discover personal accounts. itself even when two articles most-reportable-event classifier interesting problem. for women wanting to advance years later. was happening after that, and For users—and the intent is for don’t share a common lan- but in a way that is not disaster- How will you solve it? in computer science? It might be emergency respond- rescue efforts accelerated. anyone to be able to use the guage. There’s going to be more specific. We don’t want to have We don’t know, at least not Just having other women around paraphrasing also as journal- to enumerate it for different now. Clearly geographic and ers or relief workers needing to What was the inspiration? system—we envision currently makes a big difference. Women know where to direct their ef- some type of browser-type in- ists and writers, to avoid being events, to have to explicitly state temporal information is impor- can give one another advice and forts first; it might be journalists For us it was Hurricane Sandy, terface. It will probably be visual repetitive after days of writing that the most reportable event tant, but large-scale disasters together support issues that wanting to report current facts since many of us were here in and may be laid out by location about the same topic, change for a hurricane is flooding, and can encompass enormous women in particular care about. or researching past similar disas- New York or nearby when Sandy where things happened. Click- up their sentences. It makes that it’s something different for a areas and the effects go on for Having a woman in a senior ters. It might be urban planners struck. One student, whose ing on one location will present it harder for language tools tornado or a mass shooting. weeks. We need to find clues in Henry and Gertrude position is especially helpful. wanting to compare how two family on the Jersey Shore descriptions and give a timeline that are looking for the same language to draw the connec- Rothschild Professor of was hard-hit, was particularly phrases and words. What are the hard problems? Computer Science When I was department chair, different but similar neighbor- about what happened at that tions between different events. motivated to look at what could Timelines that summarize Kathy McKeown several undergraduate women hoods fared under the disaster, location at different times, and Determining semantic related- We’re just getting started. And be done to help people. changes over time, that describe approached me about starting maybe discovering why one each sub-event will be accom- ness is obviously the key, but that’s the fun of research, the how something that happened a Women in Computer Science escaped heavy damage while But more importantly, Sandy as panied by a personal account. we’re going to need to build new challenge of the difficult. You A deck of playing cards honors 54 no- today is different from what we group. As a woman, I under- the other did not. Or it might be an event has interesting char- language tools and approaches start with a question or a goal table women in computing. Produced Newsblaster is already finding knew yesterday. In general that’s by Duke University and Everwise (in stood the need for such a group someone who lived through the acteristics. It was large scale that don’t rely on the explicit without an easy answer and articles that cover the same a hard problem. On one hand, conjunction with CRA-W and Anita and in my position it was easier event coming back in years after and played out over multiple presence of shared terms. you work toward it. Borg Institute Wikipedia Project), the for me to present the case. to remember what it was like. days and multiple areas, and it event. Will you be building on everything is different, but most There are other problems, of first card decks were distributed at triggered other events. Being top of Newsblaster? How will “Describing Disas- things that are different aren’t Of course, getting women into It’s not a huge grant—it’s course. The amount of data sur- this year’s Grace Hopper conference. able to provide a description of Yes, after all, Newsblaster repre- ters” find personal stories? important. So how do we know engineering and computer sci- enough for me and two gradu- rounding a large-scale disaster (A Kickstarter campaign is raising an event at this magnitude is sents 11 years of experience in That’s one question, but the what’s new and different that is ence requires making sure girls ate students for three years, but is enormous. Somehow we’ll money for a second printing.) hard and poses an interesting auto-generating summaries and more interesting question important and should be includ- in school remain interested in it’s a very appealing one, and it’s have to wade through massive One of the women to be honored with problem. The project is meant it contains years of data, though is how do you recognize a ed? Some things that happen on these subjects as they approach one that allows undergraduates amounts of data to find just her own card is Kathy McKeown, to cover any type of disaster— we will modernize it to include good, compelling story people later days may be connected to college, and I think one way to work with us and do re- what we need. Validation is an- a computer scientist working in the earthquakes, terror attacks, social media, which News- would want to hear? Not a lot the initial event, but not neces- field of natural language processing. that is done is by showing them search. Already we have seven other issue. How do we know floods, mass shootings—where blaster doesn’t currently do. We of people have looked at this. sarily. How do we tell? The first female full-time professor in how stimulating and gratifying students working on this, the if we’ve developed something the impact is long-lasting and also need to also expand the While there is work on what Being able to connect sub- Columbia’s school of engineering, she it can be to design and build two funded graduate students that is correct? We’ll need met- generates sub-events. scope of how Newsblaster uses makes scientific writing good, events triggered by an initial was also the first woman to serve as something that helps others. and five undergraduates. rics to evaluate how well the natural language processing. recognizing what makes a story event can be hard for a program department chair. Currently the Henry system is working. and Gertrude Rothschild Professor Talking of interesting work It makes sense that journalists How will it work? Currently it relies on common compelling is new research. to do automatically. Programs Another issue will be reining in of Computer Science, she is also the that helps others, you recently and emergency responders The underlying system will tap language between articles to We’re starting by investigating a can’t see the correlation neces- everything we will want to do. inaugural director of Columbia’s mul- received an NSF grant for would want an objective into streaming social media both find articles of the same number of theories drawn from sarily. We’re going to have to tidisciplinary Data Science Institute. There is so much opportunity “Describing Disasters and the account of a disaster or event. and news to collect informa- event and then to produce sum- linguistics and from literature insert more intelligence for this And now she is the seven of clubs. with this project, particularly for Ensuing Personal Toll.” What Why include personal stories? tion, using techniques of natural maries. I’m simplifying here, but on what type of structure or to happen. multidisciplinary studies. We is the work to be funded? Because it’s the personal language processing and arti- Newsblaster works by extract- features are typically found Here’s one example. There was Congratulations on being can easily see pulling in journal- The goal is to build a system stories that let others under- ficial intelligence to find those ing nouns and other important in narratives. We’ll be looking a crane hanging over mid-town included among the Notable ism students, and those work- for automatically generating a stand fully what is going on. articles and stories pertinent to words from articles and then especially at the theory of the Manhattan after Hurricane Women of Science card deck. ing in visualization. To be able comprehensive description of Numbers tell one side of the a specific disaster and the sub- measuring statistical similar- sociolinguist William Labov of Sandy came in. If we were What does it feel like to see to present the information in a disaster, one that includes story, but emotions tell another. events it spawns. Each type of ity of the vocabulary in these the University of Pennsylvania using our normal expectations your picture on the seven an appealing way will make the objective, factual information— During Katrina, there were disaster is associated with a dis- articles to determine which who has been looking at the of hurricane-related events, we of clubs? system more usable to a wider what is the specific sequence reports of people marooned in tinct vocabulary and we’ll build articles cover the same topic. language that people use when wouldn’t necessarily think of a range of people, and it may gen- It is really exciting to be part of events of what happened the Superdome in horrendous language models to capture this In a disaster covering multiple telling stories. There is often an crane dangling over a city street erate new ways of looking at of this group, to be part of when—with compelling, emo- conditions—reports that were information. days with multiple sub-events, orientation in the beginning that from a skyscraper. This crane the data. We see this project as such a distinguished group tional first-person accounts of initially dismissed by authorities. Obviously we’ll look at estab- there is going to be a lot less tells you something about loca- sub-event spawned its own a start; there is a lot of potential of computer scientists. people impacted by the disas- It wasn’t until reporters went lished news sites for factual common language and vo- tion, or a sequence of compli- sub-event, a terrible traffic jam. for something bigger. ter. We will use natural language there and interviewed people information. To include first-per- cabulary among the articles we cating actions that culminates in Traffic jams happen in New York You started at the Columbia processing techniques to find and put up their stories with son stories, it’s not yet entirely want to capture. A news item an event, which Labov calls the city. How would we know that Engineering school in 1982 the relevant articles and stories, pictures and videos that the rest clear where to look since there about flooding might not refer most reportable event—some- this one traffic jam was a result and were for a time the only and then weave them together of us could actually see the true aren’t well-defined sites for directly to Sandy by name; it thing shocking or involving life of Hurricane Sandy and not full-time female professor in into a single, overall resource so plight and desperation of these this type of content. We will be may describe the flooding only and death for which you tend to some normal everyday event in Columbia’s school of engineer- people can query the informa- people, and you realize at the searching blogs and discussion as “storm-related” but we have get a lot of evaluative material. New York? ing. From your experience

12 CS@CU SPRING 2015 CS@CU SPRING 2015 13 Feature Articles (continued)

Want the attention of the audience? Hand positions far from home base correlated with audience engagement

Introduce a new gesture. Home base

Points far from “home base” correlated with heightened levels of listener attention (here, Romney’s left hand in the first debate).

Correlating Speaker Gestures in But what gestures are most and distinguish left and right From tracking gestures, it was Two corresponded to specific an increasingly necessary task Political Debates with Audience En- effective in getting listeners to hands even when the hands immediately apparent that areas of the brain: the first to as the amount of video contin- gagement Measured via EEG, a paper pay close attention to what a were clasped. The algorithm Romney used his hands more electrodes near the visual ues to explode. Video is hard to presented earlier this month at ACM speaker is saying? also detected gestural features and made more use of extremal cortex, with more activity sug- chunk meaningfully, and video Multimedia, describes the work done the researchers suspected gestures. This was true in both gesting that audience mem- indexing today often relies on by Columbia University researchers To find out, Columbia University beforehand might correlate with debates 1 and 3. (The study did bers were actively watching a a few extracted images and to identify what gestures are most computer science and biomedi- audience attention, specifi- not look at debate 2 where the candidate. The second compo- standard fast-forwarding and effective at getting an audience to cal engineering researchers cally velocity but also change format—candidates stood and nent was generated near the reversing. An algorithm trained pay attention. Using EEG data of study motivated by John Zhang, a in direction and what is called walked while holding a micro- prefrontal cortex, the site of to find extremal speaker ges- participants watching clips of the PhD student of Prof. John R. an invitation feature, where the phone—could bias the ges- executive function and decision- tures might quickly and auto- 2012 presidential debates, research- Kender, set up an experiment Professor of Computer ers found atypical, or extremal ges- hands are spread in an open tures.) Obama, widely criticized making, indicating that audience matically locate video highlights. Science John R. Kender to correlate the level of audi- tures to be the strongest indication manner. (In previous work for his lackluster performance members were thinking and Students of online courses, for ence interest with specific hand of listeners’ engagement. This finding studying teacher gestures, the in debate 1, gestured less and making decisions. (It was not example, could then easily skip not only benefits speakers but may gestures made by a speaker. authors were the first to identify made zero extremal gestures possible to specify a source for to the parts of a lecture needing lead to methods for automatically The researchers asked 20 par- this type of gesture.) in the first debate. If gestures the third component.) the most review. indexing video. ticipants—equally divided along alone decided the debate win- gender and political party lines— Another feature was likeli- ner, Obama lost debate 1. (He Once time-stamped, the EEG More work is planned. The re- to view video clips of the 2012 hood of hand position. Though would improve in the third de- data, averaged across subjects, searchers looked only at correla- When people talk, they often gestures vary greatly from was aligned with gestures in tion, leaving for future work the gesture to add emphasis and presidential debates between bate, probably after coaching.) Obama and Romney. The intent person to person, individuals the video so researchers could task of prediction, where some help clarify what they are say- was to track the candidates’ have their own habits. They may For gauging the level of interest, locate statistically significant EEG data is set aside to see if ing. People who speak profes- researchers chose EEG data correlations between a ges- it’s possible to use gestures to gestures with the level of inter- move their hands up and down sionally—lecturers, teachers, since it is a direct, non-intrusive ture feature (direction change, predict where there is engage- est exhibited by the audience. or left and right, but they tend politicians—know this either to do within a defined area, or measure of brain activity already velocity, and extremal pose) ment. Whether certain words John Zhang instinctively or learn it from Gestures were automatically home base, only occasionally known to capture information and a strong EEG component. are more likely to get audience experience, and it’s supported extracted from the video using moving their hands outside the related to attention. Electrodes Moments of engagement were reaction is another area to by past studies showing that an algorithm that incorporated home base. These non-habitual, were attached to the scalp of defined as a common neural re- explore. In a small step in this particular gestures are used for computer-vision techniques to or extremal, poses became an the 20 study participants to sponse across all subjects. Re- direction, the researchers looked various semantic purposes. identify and track hand motions additional feature. record brain activity while they sponses not shared with other at words with at least 15 occur- watched debate clips. The data subjects might indicate lack of rences, finding that two words, capture was carried out in the engagement (i.e., boredom), as business and (interest) rate, Mapping hand locations, first debate lab of Paul Sajda, a professor each participant would typically were unusual in their ability to in the Biomedical department, focus on different stimuli. draw attention. Though the data who also helped interpret the is insufficient for any definite Extremal gestures turned out to Left Right Both data. (EEG data is not easy conclusions, it does suggest to work with. The signals are be the strongest indication of potentially interesting results. weak and noisy, and the capture listeners’ engagement. No mat- process itself requires an elec- ter how researchers subdivided Nor did researchers look at what the participants—Democrats, the hand was doing, whether Obama trostatically shielded room. The data was also voluminous—46 Republicans, females, males, it was pointing, forming a fist, electrodes for 20 participants all—what really triggered peo- or posed in another manner. tracked for 47 minutes with ple’s attention was something The current work focused only 2000 data points per second— new and different in the way a on hand position, but this was and was reduced using an algo- candidate gestured. enough to show that gesturing can effectively engage and influ- rithm written for the purpose.) This finding that extremal poses Romney ence audiences. While the EEG data showed correlate with audience engage- many patterns of activity, there ment should help speakers stress important points. was a lot of underlying signal, By Linda Crane from which researchers identi- And it may also provide an Each point indicates a hand location (sampling was done every 1 second). Curves indicate the probability of a gesture occurring at a specific distance fied three main components. automatic way to index video, (in pixels) from the candidate’s home base.

14 CS@CU SPRING 2015 CS@CU SPRING 2015 15 Feature Articles (continued) Faculty News & Awards

Reconstruction history of Ashkenazi Jews and The researchers note that their Europeans indicates the former to descend from results also raise a number of a small, “bottleneck” group, which is a mix Professors Nayar Elected Mapping the hypotheses regarding Ashkenazi between Europeans and Middle Easterners. origin, which they have now begun exploring. They are look- to the National Academy DNA Sequence ing at, for example, the identity of the Ashkenazi founders: who of Inventors are the modern populations of Ashkenazi Jews most similar to them in Europe and in the Middle East? When did the admixture process take place and to what extent was that process sex-biased, where men with one origin mated with women from another? Led by Itsik Pe’er, associate (TAGC) in September 2011 Before the TAGC study, data was than in other well-known isolated Pe’er adds, “Equally impor- Nayar, who directs the Com- professor of computer science at with Todd Lencz, an investiga- available for a limited number of populations like the Amish and tant to us is making our data puter Vision Laboratory, has Columbia Engineering, a team of tor at The Feinstein Institute DNA markers (only approximately Hutterites locally, or the Iceland- available to the entire research published over 200 scientific researchers has created a data for Medical Research, director one in every 3000 letters of DNA) ers overseas. The researchers community. We’ve released it articles and has been awarded resource that will improve ge- of the Laboratory of Analytic that are mostly common in Eu- expect that medical insights from to public-access databases and more than 40 patents for his nomic research in the Ashkenazi Genomics at the Zucker Hillside ropeans. The TAGC researchers studies of specific populations fully expect the creativity of the inventions related to digital Jewish population and lead to Hospital, and associate profes- performed high-depth sequenc- will also be relevant to general scientific world to come up with imaging, computer vision, more effective personalized med- sor of molecular medicine and ing of 128 complete genomes of populations as well. additional uses for the data. human-computer interfaces, icine. The team, which includes psychiatry at the Hofstra North Ashkenazi Jewish healthy indi- and robotics. Today, his re- The TAGC team’s findings also What’s especially gratifying is experts from 11 labs in the New Shore-LIJ School of Medicine. viduals. They compared their data the idea that our work will pave search results are widely used York City area and Israel, focused The other TAGC members, who to European samples, and found shed light on the long-debated in cameras for smart phones, origin of Ashkenazi Jews and the way for personalized genom- on the Ashkenazi Jewish popula- are providing expertise in the that Ashkenazi Jewish genomes ics in other populations as well.” industrial vision systems for fac- tion because of its demographic diseases they are studying, are: had significantly more mutations Europeans. The genetic data tory automation, and rendering history of genetic isolation and that had not yet been mapped. indicates that the Ashkenazi Pe’er and his TAGC colleagues engines for computer graphics. - Gil Atzmon, associate professor of Jewish population was founded the resulting abundance of pop- medicine and genetics, Albert Ein- Pe’er and his team analyzed the are already working on their He is the recipient of the pres- ulation-specific mutations and raw data and created a compre- in the late medieval times by a next step: to study specific tigious David Marr Prize in 1990 stein College of Medicine (genetics of small number, effectively only high prevalence of rare genetic longevity and diabetes); hensive catalog of mutations diseases in the Ashkenazi popu- and 1995, the David and Lucile disorders. The Ashkenazi Jewish present in the Ashkenazi Jewish hundreds of individuals, whose lation, including schizophrenia, Packard Fellowship in 1992, Shree Nayar, T.C. Chang Profes- population has played an im- - Lorraine Clark, associate profes- population. descendants expanded rapidly Parkinson’s, Crohn’s, diabetes, the National Young Investigator sor of Computer Science, has portant role in human genetics, sor of clinical pathology and cell while remaining mostly isolated and cancer, as well as other in- Award in 1993, and the NTT Dis- been elected a fellow of the with notable successes in gene biology and co-director, Personal- The TAGC database is already genetically. herited traits such as longevity. tinguished Scientific Achieve- National Academy of Inventors mapping as well as prenatal and ized Genomic Medicine Laboratory, proving useful for clinical genom- They are also working with ad- ment Award in 1994. He was “Our analysis shows that Ashke- (NAI) for demonstrating “a highly cancer screening. The study was Columbia University Medical Center, ics, identifying specific new ditional collaborators, including elected to the National Acad- nazi Jewish medieval founders prolific spirit of innovation in published online on Nature Com- Laurie Ozelius, associate professor mutations for carrier screening. the New York Genome Center, emy of Engineering in 2008 and were ethnically admixed, with creating or facilitating outstand- munications today. at Icahn School of Medicine at Mount Lencz explains: “TAGC advances to sequence approximately 500 the American Academy of Arts origins in Europe and in the Mid- ing inventions that have made a Sinai, and Susan Bressman, chair of the goal of bringing personal more Ashkenazi genomes to and Sciences in 2011 for his dle East, roughly in equal parts,” tangible impact on quality of life, “Our study is the first full DNA neurology at Mount Sinai Beth Israel genomics to the clinic, as it tells make the catalog of mutations pioneering work on computa- says Shai Carmi, a post-doctoral economic development, and the sequence dataset available for (Parkinson’s disease and related the physician whether a mutation even more comprehensive. tional imaging and his seminal Ashkenazi Jewish genomes,” neurological phenotypes); in a patient’s genome is shared scientist who works with Pe’er welfare of society.” He is among and who conducted the analysis. contributions to physics-based says Pe’er, who is also a co-chair by healthy individuals, and can The study was funded by: 170 new fellows who will be - Harry Ostrer, professor of pathology, “TAGC data are more compre- computer vision. In 2006, he - Columbia Engineering: Human Frontier inducted during the NAI’s 4th of the Health Analytics Center at genetics, and pediatrics, Albert Einstein alleviate concerns that it is caus- received the Columbia Great hensive than what was previous- Science Program Annual Conference on March 20, Columbia’s Institute for Data Sci- College of Medicine (radiogenomics, ing disease. Without our work, Teacher Award from the Society ly available, and we believe the - Collaborating labs: NIH, NSF, Glenn Founda- 2015, at the California Institute of ences and Engineering, as well cancers and rare genetic disorders); a patient’s genome sequence of Columbia Graduates. as a member of its Foundations is much harder to interpret, data settle the dispute regarding tion for the Biology of Aging; Rachel and Technology in Pasadena, CA. - Ken Offit, chief of clinical genetics European and Middle Eastern Lewis Rudin Foundation; North Shore-LIJ of Data Science Center. “With and more prone to create false Health System Foundation, Brain & Behavior “It is wonderful to be an this comprehensive catalog of at Memorial Sloan Kettering Cancer alarms. We have eliminated two ancestry in Ashkenazi Jews. In Foundation; US-Israel Binational Science Foundation; New York Crohn’s Disease academic and be recognized mutations present in the Ashke- Center (breast, ovarian, colon and thirds of these false alarms.” addition to illuminating medi- By Holly Evarts Foundation; Edwin and Caroline Levy and as an inventor,” Nayar says. nazi Jewish population, we will prostate cancers, lymphoma); eval Jewish history, our results Joseph and Carol Reich; the Parkinson’s Dis- “I am honored to be a part be able to more effectively map - Inga Peter, associate professor of The TAGC study further enables further pave the way to better ease Foundation; the Sharon Levine Corzine of such a remarkable group disease genes onto the genome genetics and genomic sciences, and more effective discovery of understanding European origins, Cancer Research Fund; and the Andrew Sabin Family Research Fund. of innovators.” and thus gain a better under- Judy Cho, professor of medicine and disease-causing mutations, since millennia before. For example, standing of common disorders. professor of genetics and genomic some genetic factors are observ- our data provides evidence for We see this study serving as a sciences, both at The Mount Sinai able in Ashkenazi individuals but today’s European population vehicle for personalized medi- Hospital (inflammatory bowel dis- essentially absent elsewhere. being genetically descendant By Holly Evarts cine and a model for researchers ease); and Moreover, the demography of primarily from late mid-eastern working with other populations.” the Ashkenazi population, the migrations that took place after - Ariel Darvasi, vice-dean of The Fac- largest isolated population in the last ice age, rather than from To help in his hunt for disease ulty of Life Sciences at The Hebrew the U.S., enables large-scale the first humans to arrive to the genes, Pe’er founded The University of Jerusalem (multiple recruitment of study patients and continent, more than 40,000 Ashkenazi Genome Consortium diseases). hence more genetic discoveries years ago.”

16 CS@CU SPRING 2015 CS@CU SPRING 2015 17 Faculty News & Awards (continued)

look to tightly integrate the Zheng’s system will have to visual and audio components tackle not only the physical phe- Changxi Zheng of simulated environments. It nomena but do so in a computa- Roxana Geambasu will represent a change from tionally efficient way so there is today’s practices where digital no delay between an action and Wins NSF sound is usually created and the resulting sound. “Realistic Scores a “Brilliant 10” recorded separately of the virtual environments require a action and then dropped in great many complex computa- CAREER Award when appropriate. In Zheng’s tions. We will need fast, cost- proposal, computational sound effective algorithms. We need will be automatically generated also to identify what calculations using physics-based simulation can be approximated while still methods and fully synchronized maintaining realism.” with the associated motion. The computational sound is just “What I propose to do about the beginning. Zheng fore- Changxi Zheng is the sound synthesis is analogous sees his proposal as laying the Roxana Geambasu securely erase data. And at the number of awards, including a to the existing image rendering groundwork for building interac- other end, cloud services not prestigious Microsoft Research recipient of a National methods for creating photoreal- tive applications and spurring has been named one only accumulate endless logs of Faculty Fellowship and an NSF Science Foundation (NSF) istic images,” says Zheng. “We new techniques for efficient of Popular Science’s user activity, such as searches, CAREER award, both in 2014, would like to create audiovisual digital production and engineer- site visits, and locations, but also and the first Google Fellowship CAREER Award for his animations and virtual environ- ing design. Education outreach “Brilliant Ten”, the keep them for extended periods in Cloud Computing in 2009. of time, mine them for business She received her BS from Poly- proposal to create ments using computational for disseminating the informa- magazine’s 13th annual methods.” tion learned from the research value, and at times share them technic University of Bucharest realistic, computer- is also part of the proposal, with list of the “brightest with others—all without the (Romania) in 2005; her MS “In addition to the realism, workshops planned for high- user’s knowledge or control. (2007) and PhD (2011) from the Assistant Professor generated sounds for sound that is not tightly coupled Assistant Professor young minds in science University of Washington. Changxi Zheng school students to encourage Roxana Geambasu Geambasu notes that this has with motion loses some of its their interest in STEM fields, and immersive virtual realities. impact and detracts from the and engineering.” become an untenable state In order to arrive at the “Brilliant with college courses for teaching and is working to identify the Ten”, Popular Science combs overall immersive experience,” undergraduate and graduate says Zheng. “Attaining any security and privacy risks inher- through hundreds of nomina- researchers how to build audiovi- ent in current mobile and web tions from around the country type of realism in virtual worlds sual computational technologies. requires synchronizing the audio technology and designs, and and vets the most promising and visual components. What For now, numerous challenges constructing systems to address candidates with experts working The five-year, $500,000 grant Noting that these top 10 re- those problems. Her research in those fields. will fund his proposal “Simu- the user hears should spring exist toward computationally searchers are already changing directly and naturally from what synthesizing realistic sound but spans broad areas of systems lating Nonlinear Audiovisual the world, Popular Science cites research, including cloud and “Popular Science has always Dynamics for Virtual Worlds and the user sees.” if the research is successful, Geambasu for her research that looked optimistically towards it will enable new virtual-world mobile computing, operating Interactive Applications.” It will take new mathematical “teaches the cloud to forget systems, and databases, all with the future, and there’s no more applications with fully synchro- personal data” and builds “soft- promising sign of what that “I am honored by the award and models and new computational nized audiovisual effects, lead a focus on security and privacy. methods. Sound is a physical ware that allows the public to She integrates cryptography, future will bring than the inspired excited about the research it sup- to new ways of creating and see where the information they work of young scientists and ports,” says Zheng, who is co- phenomenon, and creating editing multimedia content, distributed systems, data- sound by computer requires share goes.” base principles, and operating engineers across the country,” directing the Columbia Computer and bring computer-generated says Editor-in-Chief Cliff Ransom. Graphics Group (C2G2) in the Co- understanding and simulat- sound into the future. “It is very exciting and hum- systems techniques and works ing all the motions and forces collaboratively in developing “The ‘Brilliant Ten’ is a celebra- lumbia Vision + Graphics Center. bling to get this honor,” says tion of the best and brightest.” “The algorithms and tools used that go into producing sound. Geambasu. “I am flattered cross-field ideas in order to solve Computational methods will today’s data privacy issues. today for constructing immersive to be included in such an have to replicate everything virtual realities are inherently inspiring group.” from the surface vibrations on “My research aims to forge a By Holly Evarts appearance-oriented, modeling an object that produce the pres- Geambasu’s work is focused new, more privacy-conscious only geometry or visible motion. sure waves we hear as sound, on ensuring data security and future where users are better Sound is usually added manually while taking into account how privacy in an era of cloud com- aware of the implications of and only as an afterthought. So frequency, pitch, and volume puting and ubiquitous mobile their online actions and Web there is a clear gap.” are affected by the object’s size, devices, all technologies upon services and applications are built with users’ privacy in While computer-generated shape, weight, surface textures, which billions of users rely to mind,” Geambasu adds. imagery has made tremendous and countless other variables. access and host sensitive data— progress in recent years to at- The way sound propagates dif- emails, documents, financial She is the second professor at tain high levels of realism, the fers also, depending on whether information, etc.—and which the School to make the maga- same efforts have not yet been sound waves travel through air have become easy targets for zine’s annual list. In 2011, fellow applied to computer-generated or water and what obstructions, theft, espionage, hacking, and Computer Science Associate sound. Zheng is among the first or other sound waves, get in legal attacks. As she has pointed Professor Eitan Grinspun was working in the area of dynamic, the way. It’s a dynamic situation out, the mobile devices we carry named to the “Brilliant 10”. physics-based computational where a slight adjustment to with us everywhere are packed sound for immersive environ- one variable produces nonlinear with confidential information un- Geambasu, who joined Columbia ments, and his proposal will changes in another. der operating systems that never Engineering in 2011, has won a

18 CS@CU SPRING 2015 CS@CU SPRING 2015 19 Faculty News & Awards (continued) Henning Schulzrinne Named Bjarne Stroustrup Awarded AITO Dahl-Nygaard Prize, ACM Fellow and Recognized and Named a Fellow of the Computer History Museum by ITC with Service Award Bjarne Stroustrup, a visiting Stroustrup was influenced by Just recently, he was named professor at Columbia Engineer- the object-oriented model of also a Fellow of the Computer ing, is the senior award winner Simula when he first began History Museum to recognize of the 2015 AITO (Internationale work on C++ in 1979. Then at the impact and importance of pour les Technologies Objets) Bell Labs, he needed a language C++ in computer history. Dahl-Nygaard Prize for the that would provide hardware design, implementation, and access and high performance He is a prolific writer and evolution of the C++ program- for systems programming tasks contributor. His publications The two awards are Henning Schulzrinne has been the immense impact of VoIP and ming language. For this same while also handling complexity. include several books—The C++ named a 2014 ACM (Association related technologies on society, achievement, Stroustrup has Since no such language then Programming Language (Fourth Edition, 2013), A Tour of C++ testament to his lasting for Computing Machinery) Fellow and they follow on Schulzrinne’s Visiting Professor in also been made a Fellow of the existed, he designed one by es- in recognition of “contributions induction into the Internet Hall of Computer Science Computer History Museum. sentially building on top of C to (2013), Programming: Principles achievements and the Bjarne Stroustrup to the design of protocols, Fame in 2013. Previous awards add support for object-oriented and Practice using C++ (2014)— immense impact of VoIP applications, and algorithms for include the New York City This prize, established in 2005 programming, data abstraction, as well as a long list of academic Internet multimedia.” He is one Mayor’s Award for Excellence in and one of the most prestigious and other capabilities. and general-interest publications and related technologies of 47 to be so recognized in 2014 Science and Technology and the in the area of software engineer- that Stroustrup maintains here. ing, is named for Ole-Johan Dahl In 1985, C++ was commercially for contributions to computing. VON Pioneer Award. He is also Stroustrup continues to up- on society an IEEE Fellow. and Kristen Nygaard for their released and spread rapidly, “I’m deeply honored to be work in creating Simula, the first becoming the dominant object- date and add functionality to Julian Clarence Levi C++. Even as new languages Professor of named ACM Fellow. Indirectly, In addition to his research, object-orientated language. oriented programming language Mathematical Methods this also recognizes the contribu- Schulzrinne is active in public in the 1990s and one of the most have been created for the and Computer Science tions of all the IRT lab members policy and in serving the broader “I feel particularly honored by rapidly shifting programming Henning Schulzrinne popular languages ever invented. over the years,” says Schulz- technology community. From this award because I knew landscape, C++ remains widely rinne, in referring to the Internet 2011 until 2014, he was the Ole-Johan Dahl and Kristen Fittingly Stroustrup is the recipi- used, particularly in large-scale Real-Time Lab (IRT) lab, which Chief Technology Officer for Nygaard,” says Stroustrup. ent of many awards, honorary and infrastructure applications he directs and which conducts the Federal Communications “While still a student in the degrees, and other recognitions. such telecommunications, bank- research in the areas of Internet Committee where he guided University of Aarhus, I learned He is a member of the National ing, and embedded systems. and multimedia services. the FCC’s work on technology object-oriented programming Academy of Engineering and and engineering issues. He from Kristen Nygaard through was previously made a Fellow It is the second honor accorded continues to serve as a technical long discussions.” of both the IEEE and the ACM. By Linda Crane Schulzrinne in as many months. advisor to the FCC. Schulzrinne In December he received an is also a past member of the Outstanding Service Award by Board of Governors of the IEEE the Internet Technical Commit- Communications Society and tee (ITC), a joint committee of a current vice chair of ACM the Internet Society and the SIGCOMM. He has served on IEEE Communications Society. the editorial board of several key Steve Bellovin Elected to the The award was given in recogni- publications, chaired important tion of excellent and extended conferences, and published National Cyber Security Hall of Fame service rendered to the fields more than 250 journal and of the ITC, of which Schulzrinne conference papers and more was the founding chair. He than 70 Internet Requests for served from 1994 to 2000 and Comment. Computer Science Professor Repelling the Wily Hacker, and including for the Department of helped shape the agenda of the Steve Bellovin, one of the holds a number of patents on Homeland Security, Election As- committee in its early years. Currently a professor within foremost experts on Internet cryptographic and network sistance Commission, and the Columbia’s Computer Sci- security and privacy, has been protocols. A graduate of Co- National Academies. From 2012 The ITC award, being given for ence department, Schulzrinne elected to the National Cyber lumbia College, he joined the to 2013, he held the post of the first time, acknowledges continues to work on VoIP and both Schulzrinne’s service to Security Hall of Fame for his Columbia Engineering faculty in chief technologist at the Federal other multimedia applications. Trade Commission. the ITC as well as service to the Ongoing projects look to create research on networks, security, 2005 following a distinguished community through his founda- an overall architecture for the and public policy issues. Bel- career at Bell Labs and AT&T In addition to this honor, Bellovin tional work on the key protocols Internet of Things, limit phone lovin joins a crop of top industry Research Labs, where he was is now the Percy K. and Vida L. Session Initiation Protocol (SIP) spam (“robocalls”), make it leaders as the 2014 class of an AT&T Fellow. Bellovin was W. Hudson Professor of Com- and Real-Time Transport Protocol Percy K. and Vida L. W. easier to diagnose network inductees, to be honored at elected a fellow of the National puter Science. The Hudson chair (RTP) that led to the develop- Hudson Professor of problems, and contribute to a Computer Science a banquet on October 30th. Academy of Engineering and was established by Mr. Hudson, ment of Voice-over-IP (VoIP). better understanding between Steve Bellovin Bellovin coauthored one of for many years has served as a a descendent of Henry Hudson, The two awards are testament engineers and public policy. the first books on firewalls, member of specialized tech- who graduated from Columbia to his lasting achievements and Firewalls and Internet Security: nology advisory committees, School of Mines in 1899.

20 CS@CU SPRING 2015 CS@CU SPRING 2015 21 Faculty News & Awards (continued) Department News & Awards

Professor Al PhD student Professor Julia Hirschberg Elected to the Aho received Jonathan Bell Simha Sethu- the Alumni and Professor madhavan Association Gail Kaiser was appointed American Philosophical Society Distinguished received an by FCC as a Faculty Teaching ACM SIGSOFT member of the Award for Distinguished Downloadable Julia Hirschberg, Percy K. and berg. “I am definitely looking leaders, from George Washing- outstanding Jonathan Bell Paper Award Security Vida L. W. Hudson Professor of forward to the APS meetings. ton, John Adams, and Thomas teaching. The award was at the 36th Technology Advisory Committee Computer Science and Depart- Jefferson to John James presented on Class Day, International (DSTAC) on January 27, 2015, Founded by Benjamin Franklin ment Chair, has been elected Audubon, Charles Darwin, Louis May 19, 2014. Alfred will also Conference which will develop interoperabil- receive an honorary degree on Software ity standards for TV navigation to the American Philosophical in 1743, APS’s mission is to Pasteur, Thomas Edison, Marie from the University of Toronto Engineering devices. Society (APS), in recognition promote useful knowledge in Curie, Linus Pauling, Margaret the sciences and humanities during convocation ceremonies (ICSE). The of her contributions to spoken Mead, and Robert Frost. Con- A paper by Dr. through excellence in scholarly in 2015 as a world-leading paper, “Unit Test language processing. She is temporary luminaries include Lisa Wu, PhD research, professional meetings, computer scientist for his Virtualization one of 33 scientists, humanists, Warren Buffet, Juan Carlos of Prof. Gail Kaiser groundbreaking research on with VMVM”, student Andrea social scientists, and leaders in publications, library resources, Spain, Jimmy Carter, and Neil programming languages, describes a Lottarini, Percy K. and Vida L. W. civic and cultural affairs elected and community outreach. Only Armstrong, and Columbia mem- MS student Hudson Professor of to APS this year, including Jill 5,474 members have been compilers and algorithms. technique for significantly Computer Science and bers University President Lee C. Timothy Paine, Lepore, James Levine, Alice elected since its founding. speeding up the software Department Chair Bollinger and Professors Joseph Professor testing process, which allows Professor Julia Hirschberg Waters, and Rem Koolhaas. APS members are nominated Stiglitz and Eric Kandel. Steven Feiner test cases to be executed more Dr. Lisa Wu Martha Kim “I’m delighted and very by current members and have was a keynote frequently during software and Professor honored to be joining such a dis- included a broad range of il- speaker at development, and hence Kenneth Ross SIGGRAPH was selected tinguished group,” says Hirsch- lustrious scholars and national By Holly Evarts allowing bugs to be detected Asia 2014 in and repaired earlier. The as one of Shenzhen, technique was shown to IEEE Micro’s China. He talked decrease the amount of time 12 Top Picks about Augmented Reality (AR) needed to run tests by an from across all and its ability to change our lives. conferences average of 62%, significantly Andrea Lottarini He discussed how AR could reducing the time and computa- in computer augment a world beyond our tional cost of testing software. architecture. Xi Chen Named 2015 Recipient own first-person perspective, The paper is assist us as we work and play EE PhD student titled “Q100: together, and create hybrid user Hyunwoo Nam, The Architecture of the Presburger Award interfaces for multiple display EE Adjunct and Design of a and interaction devices that Professor Doru Database Pro- complement each other. Calin (Director, cessing Unit”, High Perfor- Timothy Paine and describes Xi Chen was named this year’s and computational econom- be tested in time exponential in Professor recipient of the Presburger ics includes the answer to the n1/5—the first significant progress mance Wireless an improved Jonathan Technologies processor for Award. Since 2010, the award long-standing question about in more than a decade.” Gross was has been bestowed by the Eu- the computational complexity and Networks at Alcatel Lucent relational data It is unusual for someone so awarded a ropean Association for Theoreti- of Nash equilibria for two-player Bell Labs), and Professor analytics. young (Chen was born in 1982) 5-year Simons cal Computer Science (EATCS) games, showing PPAD-com- Henning Schulzrinne received to have already written papers Foundation to “a young scientist for pleteness. For classes of mar- the best paper award at IEEE that contributed to many dif- Collaboration outstanding contributions in kets and types of utility functions WCNC 2015 for their paper, ferent aspects of theoretical Grant in support theoretical computer science, widely used in economics, “Intelligent Content Delivery computer science. of his research on computational Prof. Martha Kim documented by a published he settled the complexity of over Wireless via SDN”. The aspects of low-dimensional paper describes how end-to-end Assistant Professor paper or a series of published market equilibria, again showing “I feel honored to receive this topology. The Simons Founda- Xi Chen Internet services can be achieved papers.” The award is named PPAD-completeness. His work award and be included in the tion is a private foundation that after Mojzesz Presburger, who on complexity theory includes a company of the previous win- in mobile networks such as 4G funds research in mathematics using software-defined network- while still a student in 1929 complete dichotomy theorem for ners whose work I know and and the physical sciences. It invented the field of Presburger partition function computation, respect,” said Chen. ing (SDN). They introduced high- was founded by James Simons, level architecture, possible use- arithmetic, whose simplicity—it showing it to be either polynomi- Chen joined the Columbia a well-known mathematician, cases, and experimental results involves only addition of natural al or #P-complete, as well as for Computer Science Department hedge-fund manager, and in Wi-Fi networks to show the numbers—means Gödel’s counting constraint satisfaction philanthropist. Prof. Kenneth Ross problems with complex weights, in 2011. He attended Tsinghua feasibility of their approaches. Incompleteness Theorem does University where he received not apply to it. general concepts that include e.g. counting graph homomor- a BS degree in Physics/ Chen was selected for his phisms. His work on algorithms Math (2003) and a PhD in fundamental contributions in includes a proof that isomor- Computer Science (2007). a variety of areas within theo- phism of strongly regular graphs, retical computer science. “His a well-known hard case for the work in algorithmic game theory graph isomorphism problem, can By Linda Crane

22 CS@CU SPRING 2015 CS@CU SPRING 2015 23 Columbia University in the City of New York Department of Computer Science NON-PROFIT ORG. 1214 Amsterdam Avenue U.S. POSTAGE Mailcode: 0401 PAID New York, NY 10027-7003 NEW YORK, NY PERMIT 3593 ADDRESS SERVICE REQUESTED

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