Event(s) on September 2006
||Protein Function Prediction by Information Fusion|
||Prof. Limsoon Wong, School of Computing and School of Medicine, National University of Singapore, Singapore|
||10:30 - 11:30|
| || |
Sequence homology has been widely used to shed light on the functions
of proteins. However, statistical constraints impose limits on
the sensitivity of sequence homology. Recent works introduce
other sources of biological data such as protein interaction
and gene expression profiles to reinforce protein function inference.
In our current work, we infer protein function from multiple
heterogeneous data sources by reformulating these data sources
into functional linkage graphs. Each data source is modeled into
a graph representing a network of inter-protein similarity observed
from that particular source of evidence. These graphs can be
combined to form a more complete graph in a probabilistic manner.
The edges in such functional linkage graphs, which correspond
to binary protein relationship, can be further weighted by topology
using established techniques in graph theory. Functions are inferred
for a protein from its known neighbours in the graph using a
simple but effective weighted averaging technique. Using this
approach, we combined sequence homology from BLAST searches,
protein interactions from GRID (General Repository of Interaction
Data) and co-occurence of protein names in Pubmed abstracts.
Our experiments show that the data sources complement one another
and combine to infer protein function with greater sensitivity
[This talk is based on joint work in progress with H.N.Chua
||Introduction to Biometrics and a Case Study of palmprint Identification|
||Mr. Adams Kong, Department of Electrical and Computer Engineering, Waterloo University, USA|
||11:30 - 12:30|
| || |
Biometrics based on our physiological and/or behavioral characteristics
such as fingerprints, faces and irises for automatically recognizing
individuals has advantages over traditional authentication approaches.
In the first part of this talk, a brief overview of the field
of biometrics covering markets, applications, genetic issues
and privacy concerns is given. In the second part of this talk,
a palmprint identification system based on Competitive Code is
presented. This system employs orientation fields of palmprints
as features and bitwise angular distance for matching. Experimental
results demonstrate that this system can support high speed identification.
Using Competitive Code, the genetically related features in palms
are identified. Finally, some security measures for this system
are also mentioned.
||Interactively Solving Imaging and Vision Problems Using Optimization|
||Prof. Guoping Qiu, School of Computer Science and Information Technology, The University of Nottingham, UK|
||11:30 - 12:30|
| || |
In many imaging and vision applications, it is often very difficult
or maybe even impossible to develop fully automatic solutions.
For example, despite much research effort, a fully automatic
solution to the longstanding image segmentation problem is still
an unattainable goal. Other examples where a fully automatic
solution is difficult include content-based image retrieval (CBIR).
Humans have remarkable abilities in distinguishing different
image regions or separating different classes of objects. Furthermore,
users may have different intentions in different application
scenarios. Therefore, it is both necessary and helpful to incorporate
high-level knowledge and human intentions into the computational
algorithms. Interactive approaches, which provide semi-automatic
solutions, put the users in the computational loop and allow
users to supply constraints to the computational algorithms interactively,
may offer a more realistic solution paradigm for many imaging
and vision problems.
One of the important challenges to developing successful imaging
and vision algorithms is to effectively model high level knowledge
and to incorporate the users intentions in the computational
algorithms. Traditionally, this is often achieved by integrating
statistically learned prior knowledge into numerous computational
algorithms through techniques such as Bayesian Inference. Although
the Bayesian approach has been successfully used in the literature,
the resulting combinatorial optimization problems have been often
solved by inexact and inefficient computational methods.
In this talk, I will present a computationally simple optimization-based
framework for directly incorporating priors (user inputs which
provide both high level knowledge and user intentions) into the
computational algorithms for solving problems such as interactive
content-based image retrieval and interactive foreground background