Future directions in
computer science research
John Hopcroft
Department of Computer
Science
Cornell University
Heidelberg Laureate Forum Sept
27, 2013
Time of change
The information age is a revolution
that is changing all aspects of our
lives.
Those individuals, institutions, and
nations who recognize this change
and position themselves for the
future will benefit enormously.
Heidelberg Laureate Forum Sept
27,2013
Computer Science is
changing
Early years
Programming languages
Compilers
Operating systems
Algorithms
Data bases
Emphasis on making computers
useful
Heidelberg Laureate Forum Sept
27,2013
Computer Science is
changing
The future years
Tracking the flow of ideas in scientific literature
Tracking evolution of communities in social networks
Extracting information from unstructured data
sources
Processing massive data sets and streams
Extracting signals from noise
Dealing with high dimensional data and dimension
reduction
The field will become much more application oriented
Heidelberg Laureate Forum Sept
27,2013
Computer Science is changing
Drivers of change
Merging of computing and communication
The wealth of data available in digital form
Networked devices and sensors
Heidelberg Laureate Forum Sept
27,2013
Implications for
Theoretical Computer Science
Need to develop theory to support
the new directions
Update computer science education
Heidelberg Laureate Forum Sept
27,2013
Theory to support new
directions
Large graphs
Spectral analysis
High dimensions and dimension reduction
Clustering
Collaborative filtering
Extracting signal from noise
Sparse vectors
Learning theory
Heidelberg Laureate Forum Sept
27,2013
Sparse vectors
There are a number of situations where
sparse vectors are important.
Tracking the flow of ideas in scientific
literature
Biological applications
Signal processing
Heidelberg Laureate Forum Sept
27,2013
Sparse vectors in biology
plants
Phenotype
Observables
Outward manifestation
Genotype
Internal
code
Heidelberg Laureate Forum Sept
27,2013
Digitization of medical
records
Doctor needs my entire medical
record
Insurance company needs my last
doctor
visit, not my entire medical record
Researcher needs statistical
information but
no identifiable individual
information
Heidelberg Laureate Forum Sept
27,2013
A zero knowledge proof of a statement is a proof
that the statement is true without providing you
any other information.
Heidelberg Laureate Forum Sept
27,2013
Heidelberg Laureate Forum Sept
27,2013
Zero knowledge proof
Graph 3-colorability
Problem is NP-hard - No polynomial
time algorithm unless P=NP
Heidelberg Laureate Forum Sept
27,2013
Zero knowledge proof
Heidelberg Laureate Forum Sept
27,2013
Digitization of medical records
is not the only system
Car and road gps privacy
Supply chains
Transportation systems
Heidelberg Laureate Forum Sept
27,2013
Heidelberg Laureate Forum Sept
27,2013
In the past, sociologists
could study groups of a few
thousand individuals.
Today, with social
networks, we can study
interaction among hundreds
of millions of individuals.
One important activity is
how communities form and
Heidelberg Laureate Forum Sept
27,2013
Future work
Consider communities with more
external edges than internal edges
Find small communities
Track communities over time
Develop appropriate definitions for
communities
Understand the structure of different
types of social networks
Heidelberg Laureate Forum Sept
27,2013
Our view of a community
Colleagues at Cornell
Classmate
s
TCS
Me
Family and friends
More connections
outside than
inside
Heidelberg Laureate Forum Sept
27,2013
Structure of
communities
How many communities is a
person in?
Small, medium, large?
How many seed points are
needed to uniquely specify a
community a person is in?
Which seeds are good seeds?
Etc.
Heidelberg Laureate Forum Sept
27,2013
What types of
communities are there?
How do communities
evolve over time?
Are all social networks
similar?
Heidelberg Laureate Forum Sept
27,2013
Are the underlying graphs for
social networks similar or do we
need different algorithms for
different types of networks?
G(1000,1/2) and G(1000,1/4) are
similar, one is just denser than the
other.
G(2000,1/2) and G(1000,1/2) are
similar, one is just larger than the
other.
Heidelberg Laureate Forum Sept
27,2013
Heidelberg Laureate Forum Sept
27,2013
Heidelberg Laureate Forum Sept
27,2013
TU Berlin Sept 20, 2013
Two G(n,p) graphs are similar
even though they have only
50% of edges in common.
What do we mean
mathematically when we say
two graphs are similar?
Heidelberg Laureate Forum Sept
27,2013
Theory of Large Graphs
Large graphs with billions of vertices
Exact edges present not critical
Invariant to small changes in definition
Must be able to prove basic theorems
Heidelberg Laureate Forum Sept
27,2013
Erds-Renyi
n vertices
each of n2 potential edges is present with
independent probability
N pn (1-p)N-n
n
number
of
vertices
vertex degree
binomial degree distribution
Heidelberg Laureate Forum Sept
27,2013
Heidelberg Laureate Forum Sept
27,2013
Generative models for
graphs
Vertices and edges added at each
unit of
time
Rule to determine where to place
edges
Uniform probability
Preferential attachment - gives rise to
Laureate distributions
Forum Sept
power
lawHeidelberg
degree
27,2013
Preferential attachment gives
rise to the power law degree
distribution common in many
graphs.
Number
of
vertices
Heidelberg Laureate Forum Sept
27,2013
Vertex degree
Protein interactions
2730 proteins in data base
3602 interactions between proteins
Only 899 proteins in components. Where are the 1851
missing proteins?
Science 1999 July 30; 285:751753
Heidelberg Laureate Forum Sept
27,2013
Protein interactions
2730 proteins in data base
3602 interactions between proteins
Science 1999 July 30; 285:751753
Heidelberg Laureate Forum Sept
27,2013
Science Base
What do we mean by science
base?
Example:
High dimensions
Heidelberg Laureate Forum Sept
27,2013
High dimension is fundamentally different
from 2 or 3 dimensional space
Heidelberg Laureate Forum Sept
27,2013
High dimensional data is
inherently unstable.
Given n random points in ddimensional space, essentially all n2
distances are equal.
d
x y xi yi
2
i 1
Heidelberg Laureate Forum Sept
27,2013
High Dimensions
Intuition from two and three dimensions is not
valid for high dimensions.
Volume of cube is
one in all
dimensions.
Volume of
sphere goes to
zero.
Heidelberg Laureate Forum Sept
27,2013
Gaussian distribution
Probability mass concentrated
between dotted lines
Heidelberg Laureate Forum Sept
27,2013
Gaussian in high dimensions
Heidelberg Laureate Forum Sept
27,2013
Two Gaussians
Heidelberg Laureate Forum Sept
27,2013
2 Gaussians with 1000 points each: mu=1.000, sigma=2.000, dim=500
4
3
2
1
0
-1
-2
-3
-4
-4
-3
-2
-1
Heidelberg Laureate Forum Sept
27,2013
2 Gaussians with 1000 points each: mu=1.000, sigma=2.000, dim=500
4
3
2
1
0
-1
-2
-3
-4
-4
-3
-2
-1
Heidelberg Laureate Forum Sept
27,2013
Distance between two random
points from same Gaussian
Points on thin annulus of radiusd
Approximate by a sphere of radius
d
Average distance between two points2d
is
(Place one point at N. Pole, the other point at
random. Almost
surely, the second point will be near the equator.)
Heidelberg Laureate Forum Sept
27,2013
Heidelberg Laureate Forum Sept
27,2013
2d
Heidelberg Laureate Forum Sept
27,2013
Expected distance between
points from two Gaussians
separated by
2d
2d
2
Heidelberg Laureate Forum Sept
27,2013
Can separate points from two
Gaussians if
2 2d
2d 1
1 2
2 2d
2d
L
2d
1 2
2 2d
2 2d
1
4
Heidelberg Laureate Forum Sept
27,2013
Dimension reduction
Project points onto subspace
containing centers of Gaussians.
Reduce dimension from d to k, the
number of Gaussians
Heidelberg Laureate Forum Sept
27,2013
Centers retain separation
Average distance between points
reduced
by
d
k
x1, x2 ,L , xd x1, x2 ,L , xk ,0,L ,0
d xi k xi
Heidelberg Laureate Forum Sept
27,2013
Can separate Gaussians
provided
2 2k
2k
> some constant involving k and
independent of the dimension
Heidelberg Laureate Forum Sept
27,2013
We have just seen what a science
base for high dimensional data
might look like.
For what other areas do we need a
science base?
Heidelberg Laureate Forum Sept
27,2013
Ranking is important
Restaurants, movies, books, web pages
Multi-billion dollar industry
Collaborative filtering
When a customer buys a product, what
else is he
or she likely to buy?
Dimension reduction
Extracting information from large data
sources
Social networks
Heidelberg Laureate Forum Sept
27,2013
This is an exciting time for
computer science.
There is a wealth of data in digital
format, information from sensors,
and social networks to explore.
It is important to develop the
science base to support these
activities.
Heidelberg Laureate Forum Sept
27,2013
Remember that institutions,
nations, and individuals who
position themselves for the future
will benefit immensely.
Thank You!
Heidelberg Laureate Forum Sept
27,2013
