Transcript Systems - Columbia University

Columbia’s Vision for Tomorrow’s Global
Intelligent Systems
Henning Schulzrinne, Chair
Department of Computer Science
October 13, 2005
Bill Gates/CS Faculty Roundtable
Computer Science Research Highlights
•
Comprehensive research, with areas of focus
•
Leading research in
– Natural language processing
– Mobile and wireless computing
– Designing digital systems
– Network security
•
Growing Research Impact
Columbia CS
Faculty: 35 (32 tenure track, 3 lecturers) + 3 joint
Aho
Gravano
Malkin
Schulzrinne
Allen
Belhumeur
Grinspun
Gross
McKeown
Servedio
Bellovin
Grunschlag
Cannon
Carloni
Edwards
Feiner
Hirschberg
Jebara
Kaiser
Kender
Misra
Nayar
Nieh
Nowick
Shortliffe
Stolfo
Stein
Traub
Ramamoorthi
Unger
Ross
Wozniakowski
Galil
Keromytis
Rubenstein
Yannakakis
Yemini
Columbia CS
Columbia Computer Science Research
UI, NLP,
collab work
graphics,
robotics,
vision
Interacting with
networks,
security, OS,
software eng
Humans
quantum
computing,
crypto,
learning,
algorithms
Interacting with
(5 faculty)
The Physical World
(9)
Computer
Systems
Science Theory
(11)
(8)
databases,
data mining,
machine
learning
Making Sense
of Data
Designing
(7)
Digital Systems
(4)
Columbia CS
CAD, async
circuits,
embedded
systems
Interacting with Humans: Newsblaster
•
•
•
•
Automatic summarization of articles on the same event
Generation of summary sentences
Tracking events across days
Foreign news  English summaries
Columbia CS
Faculty: Kathy McKeown
Interacting with Humans: Newsblaster
Research Findings
•
Quality of facts gathered significantly better
– With Newsblaster than with no summaries
•
User satisfaction higher
– With Newsblaster sentence summaries than Google style 1sentence summaries
• Summaries contributed important facts
– With Newsblaster than Google summaries
•
Full multi-document summarization more powerful than
documents alone or single sentence summarization
Columbia CS
Interacting with Humans: Detecting Deceptive Speech
•
•
•
Problem:
– Can we detect deception from spoken language cues only?
Method:
– Collect corpus of deceptive/non-deceptive speech
– Extract acoustic, prosodic and lexical features automatically
• E.g., disfluencies, response latency, high pitch range, lower
intensity, laughter, personal pronouns
– Run Machine Learning experiments to create automatic prediction
models and test on held-out data
Results:
– Baselines:
• Best general human performance in literature ranges from
criminals (65% accuracy) down to parole officers (40%)
• Majority class, our data (predict truth): 61%
• Mean human performance, our data: 60%
– Our (automatic) results: 69%
Columbia CS
Faculty: Julia Hirschberg
Interacting with Humans: Learning to Match Authors
Entity Resolution of Anonymized Publications
7 Teams: UMass, Maryland, Fair-Isaac, Illinois, Rutgers, CMU, Columbia
Key
1 - Permutational Text Kernels
2 - Permutational Clustering
3 - SVM
Error rate
1
3
2
Columbia
Source: 2005 KDD Challenge
Columbia CS
Faculty: Tony Jebara
Windows XP
Systems: Distributed Channel Allocation in
Mobile Mesh Networks
Channel Allocation Protocol
TCP/IP
MCL*
NDIS**
DevCon
802.11card A 802.11card B
CEPSR
research
building
Multi-radio mesh node
Results
• Channel scarcity  need automated
channel allocation in 802.11 mesh
networks
• Allocates radios by self-stabilizing
algorithm based on graph coloring
• First self-organizing mechanism &
implementation
• Network self-organizes in seconds
• Network throughput improvement of
20-100% cf. static channel allocation
Collaborators: Victor Bahl and Jitendra Padhye @ MSR
Columbia CS
Faculty: Misra/Rubenstein
Systems: Evolution of VoIP
“how can I make it
stop ringing?”
long-distance calling,
ca. 1930
“amazing – the
phone rings”
1996-2000
“does it do
call transfer?”
going beyond
the black phone
catching up
with the digital PBX
2000-2003
Columbia CS
2004-
Faculty: Henning Schulzrinne
Systems: Creating new services for VoIP
•
•
•
•
Old telecom model:
– Programmers create mass-market
applications
– new service each decade
Our (web) model:
– Users and administrators create
universe of tailored applications
Incorporate human context:
– location, mood, actions, …
“FrontPage for service creation”
– Based on presence, location, privacy
preferences
– Learn based on user actions
Columbia CS
Systems: Self-healing Software
• Problem: zero-day attacks
• Approach: Enable systems to react and self-heal in response
to unanticipated attacks and failures, via:
– Coordinated access control in large-scale systems
– Block-level system reconfiguration
– Self-healing software systems
– Application communities: enable large numbers of
identical applications to collaboratively monitor their
health and share alerts
• Prototypes: worms, software survivability
Columbia CS
Faculty: Angelos Keromytis, Sal Stolfo
Systems: Developing Profiles of Attackers
Surveillance
detected at
site A
Common
sources of
scans for all
three sites
Surveillance
detected at
site B
Site A
Site B
Site C
Profile and signature
generation for defense
Surveillance
detected at
site C
•Worms use hit lists to reduce spread time
•Gathered in stealth
•Collaborative and distributed intrusion detection
•Leverage header and payload anomaly
Columbia CS
Faculty: Sal Stolfo
Theory: Leveraging Cryptography
•
Traditional cryptography  provable security of
protocols, but assumes a clean, controlled model
– Key exposure causes more security breaches than
cryptanalysis
• smartcards, PDAs can easily leak keys
•
 Expand theoretical foundations to capture provable
security against strong, realistic attackers, including:
–
–
–
–
Key exposure
Key tampering
Security against side channel attacks (power, timing analysis)
Security in an Internet-like setting
• when attacker can coordinate across several protocols
Columbia CS
Faculty: Tal Malkin
Columbia’s Growing Computer Science Research Impact
government
Industrial
12
research gifts
10
8
Funding
(M$)
6
4
2
Year
Technology impact through
• start-ups
• security, network management, thin clients, VoIP, …
• standardization
•VoIP, security, …
• education
Columbia CS
5
0
4
/0
4
0
3
/0
3
0
2
/0
2
0
1
/0
1
0
0
/0
0
9
9
/0
9
/9
8
9
8
9
7
/9
7
/9
9
6
9
5
/9
6
0
Conclusion
•
•
Broad-based research motivated by real problems
Breaking new ground in several key areas, e.g.:
–
–
–
–
•
Natural language processing
New network services and models
Network security
Graphics & vision
Columbia has a growing impact on computer science as
demonstrated in successfully bringing new technology to
the field
Columbia CS