ICSI Celebrates Its 20th Anniversary

The International Computer Science Institute was inaugurated twenty years ago this September. ICSI was created to allow for international cooperation and advanced research in computer science to benefit from each other. Over these twenty years, ICSI has made many significant contributions to the field of computer science, and has hosted over 500 outstanding researchers with its international programs.

ICSI began in 1985 with a conversation between Professor Norbert Szyperski, director of the German Federal Laboratory for Computer Science (GMD) and Ron Kay, a senior manager at IBM's San Jose Research Laboratory. They set out to create an international research center for computer science which could serve as a "scientist's playground." With the help of UC Berkeley Professors Domenico Ferrari and Richard Karp, ICSI soon found itself a home three blocks from the Berkeley campus. The German visitor program was quickly augmented by similar visitor programs with both Switzerland and Italy, and later programs from Finland and Spain as well.

A testament to ICSI's continued long–term success is how many of its original group leaders, researchers, and trustees have stayed on board since its inception. Many of ICSI's founders are still actively involved here. Three of the four original research groups are still guided by their original leaders; moreover, these group leaders have trained new generations of researchers that are paving new ground at ICSI.


The AI group was founded by ICSI's first director, Jerry Feldman, who is still an active member of the group. As the work done in the group, formerly called Applications, changes domains, the group is thematically tied together by Feldman's original work on neural computing. This work focused around neural connectionist models, particularly in regards to language modeling, spawned the L0 project in 1990 which developed into the still active Neural Theory of Language (NTL) project.

Feldman's work on connectionist models of language in NTL was supported by graduate student Srini Narayanan, who would go on to enjoy a postdoc appointment at ICSI and a research position at partner institution SRI before he returned as a senior researcher and eventual AI group leader.

While Narayanan continues to employ neural modeling of language acquisition and use in the NTL project, he also takes techniques central to neural connectionist work and applies them to new domains. He is developing interactive intelligent question–answering systems for use in rural India. Also, working with doctoral student Joe Makin, the connectionist techniques have expanded beyond neural modeling to modeling biological processes such as the blood clotting mechanism.

The AI group's focus expanded significantly when prominent linguists Charles J. Fillmore and Paul Kay retired from their positions in the UCB Linguistics department and took up residence at ICSI. Fillmore's work with former graduate student and current PI Collin Baker has produced FrameNet, a rich online database of lexical semantics. Kay has worked extensively with fellow ICSI linguist and former graduate student Terry Regier on linguistic patterns derived from vision, including color naming and lateralization of the Whorfian hypothesis.

What started with Jerry Feldman's investigation into "brain–like computing" has evolved with the improvement of understanding of real neural activity. The AI group has moved past trying to simulate neural actions, towards recreating the results of actual neural activity. This work has been readily embraced by three generations of researchers at ICSI and continues to expand; much of the NTL group's work is explained in Feldman's recent book From Molecule to Metaphor.


Just as the AI group retains Feldman, the Algorithms group is still led by one of its founders, Turing Award winner Richard Karp. Considering Karp's joint appointments in the Computer Science, Mathematics, Bioengineering, and Operations Research Departments on campus, it shouldn't be surprising that the Algorithms group has addressed problems from a wide array of disciplines.

Karp has been involved with ICSI since the mid–1980's when he helped to write Berkeley's winning proposal to be the host site for ICSI, and he held a research appointment when ICSI was formed in 1988. By that time, he had already graduated two doctoral students who would play major roles at ICSI, Michael Luby and Ron Shamir. Luby would go on to lead the group (then called Theory of Computing) for a portion of the 90's while Karp worked at the University of Washington. Luby pioneered Digital Fountain, possibly the group's most significant achievement; it is a revolutionary way of encoding large files for transmission over the Internet that has the miraculous property that the receiver only needs to receive some of the packets to decode and receive the entire file.

Ron Shamir has been a long–time collaborator with Karp, and their work illustrates ICSI's ongoing partnership with Tel–Aviv University. ICSI and Tel–Aviv have developed an informal researcher exchange, with Shamir's doctoral students Roded Sharan and Gad Kimmel recently enjoying appointments at ICSI, and with PI Eran Halperin beginning to split his time between ICSI and Tel–Aviv. A major outcome of this collaboration is highly fruitful work on human genetic variation – a field hailed as Science Magazine's Breakthrough of the Year in 2007.

In his time at ICSI, Karp also advised the Ph.D. of Sally Floyd, who would return to ICSI to be a PI in the Networking group. Karp and the Networking group collaborated extensively upon his return with the formation of ACIRI, explained below, during the late 90's and early 2000's. This collaboration underscores the interdisciplinary nature of Karp's work; he explains that "if you peel back the layers and look at the core of a problem, it's almost always a simple, concise algorithm."

It is this willingness to innovatively explore new areas that has led the Algorithms group to confront problems as assorted as congestion control in computer networks, peer–to–peer networking, file encoding and transmission, genetic variation within disease populations, randomized algorithms, game theory, and computational finance. Karp's tradition of innovation and collaboration recently made him the recipient of the prestigious Kyoto prize, described later in this newsletter.


ICSI's work in large distributed computer systems began with the Tenet group under the supervision of Domenico Ferrari. Ferrari, then Chair of the Computer Science department on campus, was another member of UCB's faculty who "immediately loved the idea" of ICSI's creation when Ron Kay initially discussed it with him in the mid–1980's. His hearty support played a major role in ICSI's establishment. Accordingly, Ferrari served as the first deputy director of ICSI, and founded the Tenet group, heading it well into the 90's. Under his direction, the group designed and built two real–time protocol suites. His "very deep belief in the internationality of science" is evidenced by Ferrari's successful efforts to establish one of ICSI's first visitor programs, in particular with Italy.

In 1995 Ferrari retired and relocated to Italy to head an institute at the Università Cattolica. In 1999, Scott Shenker secured generous funding from AT&T that revitalized the group with the creation of ACIRI, the AT&T Center for Internet Research at ICSI. This enabled researchers to work on Internet issues as they saw fit, free from the usual time– and topic– based constraints so often encountered in both academia and industry. The promise of unfettered research allowed a "dream team" of Internet researchers to gather at ICSI, bringing together Shenker, Vern Paxson, Sally Floyd, Mark Handley, and Richard Karp (returning to ICSI from Washington). Among many other research developments, the Networking group gave birth to the eXtensible Open Router Platform XORP, which has recently spun out into a new startup company. After AT&T's funding was discontinued, the group was rechristened as ICIR, the ICSI Center for Internet Research.

The free–thinking innovation of ICSI's Networking group is apparent by the attention they've received. The SIGCOMM Award – the highest honor in the field of networking research – has gone to Shenker in 2002, Ferrari in 2006, and Floyd in 2007. Paxson received SIGCOMM's Test of Time award in 2006, and the ACM's Grace Murray Hopper award in 2007. All are ACM Fellows. The Networking group has contributed to the fields of Internet architecture, traffic, congestion, measurement, and security. The group's efforts in Internet security, guided by Paxson, were the subject of the last issue of the ICSI newsletter.


ICSI's Speech group is another group that is still being led by its founder, current ICSI director Prof. Nelson Morgan. Initially designated "Realization of Massively Parallel Systems," the group's early endeavors spawned a number of notable hardware innovations. One example is the development of the Ring Array Processor (RAP) in 1989, a high–speed parallel computing processor. The RAP arose out of the group's desire to do large–scale computation with neural nets; no hardware existed that was well–suited to the job, so they invented one. That same problem–solving resourcefulness led to the development of a vector microprocessor system comprised of the microprocessor T0, designed by then–graduate student Krste Asanovic and others, and SPERT–II, designed by James Beck. All while creating innovative hardware, the group was using their new inventions to tackle the multi–faceted problems facing computer speech recognition.

The late 1990's saw the rate of improvement of commercial machines rapidly increase. It reached a point where after taking the requisite years to develop and implement new special–purpose hardware, the resultant hardware may only have been marginally better than mainstream developments. This obviated the group's need to build faster machines to perform their work, and permitted them to refocus primarily on the use of neural nets in speech processing. Accordingly, their name changed from the Realization group to the Speech group.

This shift coincided with the graduation of Asanovic, who would return from MIT a few years later to start the Architecture group (see below). Other early members of the Speech group have also played very long–term roles in ICSI's development. Chuck Wooters (ICSI's first graduate student) returned from his work in industry to enjoy a seven–year stint as a staff researcher working on speaker diarization. David Johnson, a former software engineer for the group's processors, is ICSI's current Systems Manager.

The focus of the group over the last decade has been speech recognition, speaker recognition, and speaker diarization: what is being said, who is saying it, and when is each person speaking. Senior researchers in the ICSI Speech Group, including Morgan and new member Dilek Hakkani–Tur, hold a number of speech–recognition patents, one of which may be implemented on your current cell phone.

New groups

Krste Asanovic recently returned to ICSI in 2005 to revive hardware development by forming the Architecture group. His research focus is energy–efficient high–performance computing, building upon his early work with the Realization group and later projects at MIT. Asanovic is also closely involved with UCB's Par Lab, which is investigating ways to simplify the challenging task of designing parallel software for the new wave of multicore architecture. He is also exploring the use of silicon photonics to provide sufficient memory bandwidth to feed the higher demands of future manycore systems.

Another research area that has seen a recent revival at ICSI is computer vision. Work in this area began in the 1980's as part of the AI group with Steve Omohundro. Omohundro, a staff scientist at ICSI for its first six years, synthesized work done from a number of different disciplines. Utilizing machine–learning techniques, he co–designed computer vision methods to augment speech recognition. In 2008, a decade after ICSI wrapped up its initial efforts in computer vision, Prof. Trevor Darrell joined ICSI and the UC EECS faculty, starting a Vision group. Specific research areas include object and activity recognition, with application to human–robot interaction, visual surveillance, and multimodal human computer interfaces.

… and beyond!

ICSI continues to benefit from the wide array of past and present visitors, not only those who have remained to guide its day–to–day activities, but also those who are present in other capacities. ICSI founder Ron Kay is still a frequent visitor and familiar face at ICSI talks and conferences. Hervé Bourlard, ICSI's very first postdoctoral researcher and current Board member, is another example of ICSI reaping long–term benefits from its training of emerging scientists. As Director of IDIAP and IM2, Bourlard oversees the Swiss visitor program and matches ICSI with many promising graduate students and postdocs. Former graduate student and postdoc Dan Jurafsky is still in the Bay Area (on the Stanford faculty) and is a frequent collaborator with the AI and Speech groups. Terry Regier, mentioned above with AI, is currently taking a one–year sabbatical at ICSI.

In its time, ICSI has opened its doors — both physical and electronic — to over two thousand researchers, international visitors, students, and staff members. The rate at which ICSI retains those researchers, and the rate at which they are recognized for their work, speaks volumes about the quality of ICSI's research environment. Having made remarkable achievements in international macro–collaboration, ICSI is also looking to foster interdisciplinary micro–collaboration within its groups. Speaker recognition can benefit from intelligent computer vision; speech recognition can be augmented by a computer that understands the context of a discussion and can intelligently understand distorted speech; networking can benefit from the development of robust algorithms. The more interconnected that science becomes, the more ICSI has to offer because of its unique environment promoting collaboration in innovative, international, open computer science research.