Seminar Home
Fall 2013 Schedule
Winter 2014 Schedule
Summer 2014 Schedule
Archives
|
Date
(
Winter 2007
) |
Category |
Seminar Info |
| 2007/04/30 |
Bioinformatics |
Place: Duff Medical Building, 3775 University Street, Main Amphitheatre 1
Time: 12:00 - 13:00
Speaker:
Mark
Gerstein
Affiliation: Molecular Biophysics & Biochemistry, Computer Science, Yale University
Area:
Analysis and understanding of the human genome
Title: Human Genome Annotation
Abstract: A central problem for 21st century science will be the analysis and
understanding of the human genome. My talk will be concerned with topics
within this area, in particular annotating pseudogenes (protein
fossils) in the genome. I will discuss a comprehensive pseudogene
identification pipeline and storage database we have built. This has enabled
use to identify >10K pseudogenes in the human and mouse genomes and analyze
their distribution with respect to age, protein family, and chromosomal
location. One interesting finding is the large number of ribosomal
pseudogenes in the human genome, with 80 functional ribosomal proteins
giving rise to ~2,000 ribosomal protein pseudogenes.
I will try to inter-relate our studies on pseudogenes with those on tiling
arrays, which enable one to comprehensively probe the activity of intergenic
regions. At the end I will bring these together, trying to assess the
transcriptional activity of pseudogenes.
Throughout I will try to introduce some of the computational algorithms and
approaches that are required for genome annotation and tiling arrays -- i.e.
the construction of annotation pipelines, developing algorithms for optimal
tiling, and refining approaches for scoring microarrays.
|
| 2007/04/30 |
Faculty Candidate Talk |
Place: MC437
Time: 9:30 - 10:30
Speaker:
Christine
Vogel
Affiliation: nstitute for Cellular and Molecular Biology, The University of Texas at Austin
Area:
aspects of domain duplication and combination
Title: The Evolution of Complex Protein Repertoires - and beyond
Abstract: Much of an organism's physiology is determined by the proteins encoded in
its genes. Proteins in turn are composed of smaller structural, functional
and evolutionary units called domains. Duplication, combination and
divergence of these domains have shaped the protein repertoire by forming
protein families and multi-domain proteins.
My talk discusses different aspects of domain duplication and combination
from a genomic, structural and functional perspective; it describes the
relationship between the two processes, and their likely contributions to
the evolution of complex organisms.
I will also briefly describe more recent work which addresses 'faster'
processes that shape the protein repertoire, e.g., a mass-spectrometry-based
method to determine absolute protein abundance and its applications
Biography of Speaker:
-
Since 2005, Christine Vogel is a Post-doctoral Fellow at the
Institute for Cellular and Molecular Biology at the University of Texas at
Austin. In 2004, she received a Ph.D. in Computational and Structural
Biology from the University of Cambridge, MRC Laboratory of Molecular
Biology, U.K. Her research interests lie in the understanding of eukaryotic
complexity, with a focus on development and application of an integrative
and quantitative definition of eukaryotic cell types, combining
morphological and large-scale molecular data for various eukaryotic model
organisms.
|
| 2007/04/17 |
General |
Place: MC437
Time: 9:30 - 10:30
Speaker:
Jernej
Barbič
Affiliation: Carnegie Mellon University
Area:
Graphics
Title: Real-time Deformable Objects: Graphics, Haptics, Sound
Abstract: Real-time deformable objects are an exciting research area in computer
graphics, with applications to computer games, movie industry, CAD/CAM, and
virtual medicine.
Deformable objects are well-understood in solid mechanics, however the
standard simulation algorithms are too slow for interactive simulation with
detailed geometry. How can we support real-time simulation on commodity
workstations, while compromising physical correctness as little as possible?
First, I will present reduced-coordinate nonlinear deformable objects, a
novel class of deformable objects obtained by applying statistical model
reduction to finite element models. The idea is to replace the general
degrees of freedom of a deformable object for a much smaller set of reduced
degrees of freedom, thereby trading accuracy for speed. The reduced degrees
of freedom incorporate geometric and material
information, and are chosen automatically from the first principles of
continuum mechanics.
In addition, I will present a time-critical algorithm for collision
detection, deformable object simulation and contact force computation
between reduced-coordinate deformable (or rigid) objects with detailed
geometry. The algorithm runs at haptic rates (1000 Hz), enabling
applications in interactive path planning, virtual assembly and game
haptics.
Finally, I will present an algorithm for real-time sound synthesis where
both the mechanical vibrations (deformations) that cause sound, and the
sound propagation (wave equation) into the surrounding air for detailed
geometries, are (approximately) simulated at audio rates (44,100 Hz). This
captures effects such as object self-shadowing and diffraction of sound
around corners, and further improves realism of real-time virtual
environments.
PDF version
Biography of Speaker:
-
|
| 2007/04/16 |
General |
Place: MC437
Time: 9:30 - 10:30
Speaker:
Vivek
Kwatra
Affiliation: University of North Carolina
Area:
Vision
Title: Spatio-temporal Textural Modeling for Data-driven Synthesis and Visualization
Abstract: With the increased accessibility of capture devices and techniques, rich
amounts of real-world data is available to us in various forms including
images, video, 3D models, motion capture, weather patterns, etc. On the
other hand, one of the primary goals in computer graphics has been to
synthesize this data from first principles either through simulation or user
interaction. My research goal has been to develop synthesis-friendly models
from spatio-temporal data that not only exploit the richness of real data,
but also afford the controllability of simulation and interaction.
In this talk, I will focus on synthesis methods that treat visual as well as
dynamic data as texture. Such textural modeling is especially conducive for
controllable synthesis because of its locality in space and time. I will
first demonstrate a technique that allows for spatio-temporal extension of
image and video data, and combined with intelligent user interfaces, can be
used for computational photography applications such as smart image
compositing and storyboarding. I will then describe a more flexible texture
model that can be used to augment fluid simulation with appearance and shape
textures to generate complex fluid effects such as ripples and foam in
turbulent flow and patterns in the crust of flowing lava. I will also show
how we can incorporate physical and geometric characteristics of the flow
into the synthesis process in a temporally coherent manner. This technique,
besides adding to the visual realism of the simulated fluid, also provides a
handy visualization tool that lays bare significant features on the surface
of the flowing fluid, which may not be apparent otherwise.
Biography of Speaker:
-
|
| 2007/04/16 |
Bioinformatics |
Place: Duff Medical Building, 3775 University Street, Main Amphitheatre 1
Time: 13:30 - 14:30
Speaker:
Dr. Daniel
Figeys
Affiliation: The Ottawa Institute of Systems Biology, BMI, University of Ottawa
Area:
Analyzing specific portions of the proteome (sub-proteome)
Title: Probing sub-proteomes using affinity purification
Abstract: The concentration range of protein present in biological samples remains a serious challenge for proteomic technologies. The proteome is defined as the ensemble of the proteins in a sample; however, the reality is that a good portion of the proteome remains invisible because of detection and processing limitations in proteomic technologies. We have developed technologies to analyze specific portions of the proteome (sub-proteome) that combine affinity purification, mass spectrometry, and bioinformatics. In this presentation, we will provide some examples of sub-proteome studies. First, we will report our results on mapping protein-protein interaction for over 330 human genes using immunopurification coupled to mass spectrometry and bioinformatics. Using this approach, 2235 human proteins were observed to participate in 6463 interactions with the bait proteins. A suite of bioinformatic approaches were used to assess the validity of the results. Second, we will discuss the mapping of protein polyubiquitination sites using affinity purification coupled to the proteome reactor and mass spectrometry and will discuss the potential applications of this approach.
Biography of Speaker:
-
|
| 2007/04/12 |
Faculty Candidate Talk |
Place: MC437
Time: 9:30 - 10:30
Speaker:
Martin
Isenburg
Affiliation: University California, Berkeley
Area:
HCI
Title: Streaming Geometry Processing
Abstract: Modern technology enables the creation of digital 3D models that represent
objects or processes for scientific or engineering applications with
incredible detail. Operating on these large data sets is difficult because
they cannot be completely loaded into the main memory of common desktop PCs.
We have developed new streaming representations for geometric data sets that
we use as input to new streaming algorithms that can process data sets
without first loading them into memory. The key insight is to keep the data
in streams and document, for example, when all triangles around a vertex or
all points in a particular spatial region have arrived with "finalization
tags". These tags allow us to complete operations, output
results, and de-allocate data structures.
We have designed streaming simplification, compression, triangulation, and
extraction algorithms that can operate on data sets much larger than the
available memory. An added benefit of streaming is the ability to pipe
several stream modules together and avoid storing temporary results to disk.
I present two example processing pipelines: One extracts, simplifies, and
compresses iso-surfaces from a volume grid. The other generates elevation
contours from densely sampled terrain points via a temporarily constructed
and then simplified triangulation.
Biography of Speaker:
-
Martin Isenburg is currently a postdoc at UC Berkeley. He received a Ph.D.
in Computer Science from UNC Chapel Hill in 2004 and a M.Sc. in Computer
Science from UBC Vancouver in 1999. He has published over 25 papers
including three SIGGRAPH, five Visualization, and five journal publications.
His dissertation is on compressing and streaming of polygon meshes, both
in-core and out-of-core. Recently has worked on streaming other processing
tasks for large geometric data sets. His SIGGRAPH'06 paper on Streaming
Computation of Delaunay Triangulations is his showcase example for an
extremely scalable streaming algorithm.
|
| 2007/04/05 |
General |
Place: MC437
Time: 9:30 - 10:30
Speaker:
Paul
Kry
Affiliation: Université René Descartes, Paris (France)
Title: Measurement, Models, and Simulation for Computer Animation
Abstract: Motion capture has become popular for computer animation because it directly
captures the subtleties of human motion. However, reusing captured motion
becomes difficult when contact is involved. In addition to motion, the
compliance with which a character makes contact also reveals important
aspects of the movement's purpose. In this talk, I will present a new
technique called Interaction Capture and Synthesis, for capturing and
reusing these contact phenomena. Using hands and grasping as an example, we
capture contact forces at the same time as motion, at a high rate, and use
both to estimate a nominal reference trajectory and joint compliance. Unlike
traditional methods, our method estimates joint compliance without the need
for motorized perturbation devices. New interactions can then be
synthesized by physically based simulation using a novel position-based
linear complementarity problem formulation that includes friction, breaking
contact, and the compliant coupling of contacts at different fingers. I
will discuss validation and show examples of interaction synthesis. I will
also briefly describe related problems in character deformation, contact
sounds, and locomotion control.
Biography of Speaker:
-
Paul Kry is a postdoc with the EVASION group at INRIA Rhone Alpes and the
LNRS at the Université René Descartes Paris. He received his PhD from the
University of British Columbia, and conducted his research at both UBC and
Rutgers while completing his thesis with Dinesh Pai.
|
| 2007/04/03 |
Math |
Place: MC103
Time: 16:00 - 17:00
Speaker:
Jonathan
Farley
Affiliation: University of the West Indies
Title: Distributive Lattices of Small Width: A problem from Stanley's Enumerative Combinatorics
Abstract: In Richard P. Stanley's 1986 text, Enumerative Combinatorics, the
following problem is posed: Fix a natural number k. Consider the posets P
of cardinality n such that, for 0 < i < n , P has exactly k order ideals
(down-sets) of cardinality i . Let f_k(n) be the number of such posets.
What is the generating function \sum f_k(n) x^n? I will give a solution to
this problem (joint work with Ryan Klippenstine.) I will also relate this
to a problem of Ivo Rosenberg (University of Montreal) from the 1981 Banff
Conference on Ordered Sets.
Biography of Speaker:
-
|
| 2007/04/03 |
General |
Place: MC437
Time: 15:30 - 16:30
Speaker:
Caitlin
Kelleher
Affiliation: School of Computer Science, Carnegie Mellon University
Area:
Human-Computer Interaction
Title: Storytelling Alice: presenting programming as a means to the end of storytelling
Abstract: The Higher Education Research Institute (HERI) estimates that the number of
incoming college students intending to major in computer science has dropped
by 70% since 2000, despite the fact that the projected need for computer
scientists continues to grow. Increasing the numbers of female students who
pursue computer science has the potential both to help fill projected
computing jobs and improve the technology we create by diversifying the
viewpoints that influence technology design. Numerous studies have found
that girls begin to turn away from math and science related disciplines,
including computer science, during middle school. By the end of eighth
grade, twice as many boys as girls are interested in pursuing science,
engineering, or technology based careers.
In this talk, I will describe the development of Storytelling Alice, a
programming environment that gives middle school girls a positive first
experience with computer programming. Rather than presenting programming as
an end in itself, Storytelling Alice presents programming as a means to the
end of storytelling, a motivating activity for a broad spectrum of middle
school girls. More than 250 girls participated in the formative user testing
of Storytelling Alice. To determine girls' storytelling needs, I observed
girls interacting with successive versions of Storytelling Alice and
analyzed their storyboards and the programs they developed. To enable and
encourage middle school girls to create the kinds of stories they envision,
Storytelling Alice includes high-level animations that enable users to
program social interaction between characters, a gallery of 3D objects
designed to spark story ideas, and a story-based tutorial presented using
Stencils, a novel tutorial interaction technique.
To determine the impact of the storytelling focus on girls' interest in and
success at learning to program, I conducted a study comparing the
experiences of girls introduced to programming using Storytelling Alice with
those of girls introduced to programming using a version of Alice without
storytelling features (Generic Alice). Participants who used Storytelling
Alice and Generic Alice were equally successful at learning basic
programming concepts. However, I found that users of Storytelling Alice show
more evidence of engagement with programming. Storytelling Alice users spent
42% more time programming and were more than three times as likely to sneak
extra time to continue working on their programs (51% of Storytelling Alice
users vs. 16% of Generic Alice users snuck extra time).
Biography of Speaker:
-
Caitlin Kelleher is currently a post-doctoral researcher in Computer
Science and Human-Computer Interaction at Carnegie Mellon University. She
received her bachelor's degree in Computer Science from Virginia Tech and
her Ph.D. in Computer Science from Carnegie Mellon University with Professor
Randy Pausch. Caitlin was a National Science Foundation Graduate Fellow
|
| 2007/04/02 |
Bioinformatics |
Place: MC437
Time: 9:30 - 10:30
Speaker:
Chen-Hsiang
Yeang
Affiliation: University of California, Santa Cruz
Title: Computational methods for reconstructing Computational methods for reconstructing
Abstract: Recent progress in high throughput technologies has generated an enormous
amount of data which allow us to study biology at systems level. Two
research directions arise from computational systems biology: to
reconstruct a biomolecular system by integrating multiple sources of data
and to study the evolution of the components within the system. In this
talk I will present research works in each of those directions.
In the first part, I will describe a computational model capturing the
co-evolution between the components in a molecular system. It extends the
continuous-time Markov process of sequence substitution by rewarding
co-variation and penalizing single-site transitions in the rate matrix. The
model accurately predicts the secondary interactions of tRNA and 16S rRNA
molecules, and identifies the tertiary interactions which do not follow
typical Watson-Crick or GU base pairing rules. We then apply the model to
screen co-evolving amino acid pairs among all the protein domain families in
the Pfam database. The inferred pairs are highly enriched with the domains
which are physically or functionally coupled. Among the top 100 inferred
family pairs, 82 occur in the same proteins or share the same functional
annotations. By inspecting the 3D protein structures, we find many
co-evolving positions are either close and exhibit compensatory substitution
across species, or located at functionally important sites of the proteins.
In the second part, I will describe a constraint-based modeling framework of
inferring gene regulatory networks by data integration. The model treats
each piece of evidence as a constraint over attributes in the system and
builds a joint probabilistic graphical model from all constraints, and
applies statistical inference algorithms to calculate attribute values. To
demonstrate its use we apply this framework to four problems. First we
infer the causal/functional order of genes in the regulatory network of
colon cancer invasiveness using RNAi knock-down expression data. Second we
infer the causal order and combinatorial functions of the regulatory network
for the surface roughness phenotype of Vibrio cholerae, using multiple
knock-out expression data. Third we identify the active pathways and edge
directions/signs in the physical interaction network of yeast that explain
the knock-out expression data. Finally, we propose information theoretic
criteria for suggesting new knock-out experiments that
best disambiguate the existing models.
Biography of Speaker:
-
|
| 2007/04/02 |
Math |
Place: Burnside 1205
Time: 16:30 - 17:30
Speaker:
Janos
Pach
Affiliation: City College and Courant Institute, New York
Title: Turan-type results on intersection graphs
Abstract: We establish several geometric extensions of the Lipton-Tarjan separator
theorem for planar graphs. For instance, we show that any collection $C$
of Jordan curves in the plane with a total of $m$ crossings has a
partition into three parts $C=S\cup C_1\cup C_2$ such that
$|S|=O(\sqrt{m}),$ $\max\{|C_1|,|C_2|\}\leq\frac{2}{3}|C|,$ and no element
of $C_1$ has a point in common with any element of $C_2$. These results
are used to obtain various properties of intersection patterns of
geometric objects in the plane. In particular, we prove that if a graph
$G$ can be obtained as the intersection graph of $n$ convex sets in the
plane and it contains no complete bipartite graph $K_{k,k}$ as a subgraph,
then the number of edges of $G$ cannot exceed $c_kn$, for a suitable
constant $c_k$. Joint work with Jacob Fox.
Biography of Speaker:
-
|
| 2007/04/02 |
Bioinformatics |
Place: Duff Medical Building, 3775 University Street, Main Amphitheatre 1
Time: 12:00 - 13:00
Speaker:
Dr. Edwin
Wang
Affiliation: National Reseach Council, Biotechnology Research Institute
Title: Cellular signaling networks: from regulation to information superhighways
Abstract: During the last 50 years, cellular signaling data and information have been generated worldwide and accumulated in literature. In recent years, high-throughput technologies allow to generating a large mount of DNA sequence, microarray and protein data. We manually curated signaling events from literature and combined with other high-throughput data to construct a human signaling network. By integrative analysis of the network with other types of high-throughput biological data, we explored the regulation and signal propagation on the human signaling network. I will summarize the principles of gene regulation, protein phosphorylation and signal information superhighways of the network.
Biography of Speaker:
-
|
| 2007/03/30 |
Software Engineering |
Place: McConnell 103
Time: 11:00 - 12:30
Speaker:
Eric Bodden and Laur
Affiliation: McGill University
Area:
Conference Report (RV 2007 and AOSD 2007)
Title:
Abstract: We will present an overview of the RV (Runtime Verification)
workshop and AOSD 2007 conference.
We will start with an overview of what is hot and then give a
more detailed presentation of three of the papers.
|
| 2007/03/23 |
Software Engineering |
Place: McConnell 103
Time: 11:45 - 12:30
Speaker:
Jan
Rupar
Affiliation: McGill University
Title: Software Security Analysis on Stripped Binaries - An Exploration
Abstract: This talk will give an overview of the major problems faced by security
auditors today: how does one conduct the security evaluation of
executable files without the support of documentation or source code?
Current tools and methodologies will be presented, along with their
challenges and limitations. This talk has the goal of stimulating a
discussion on how we may apply current software engineering and dataflow
analysis research in order to address these challenges.
|
| 2007/03/23 |
Software Engineering |
Place: McConnell 103
Time: 11:00 - 11:45
Speaker:
Barthelemy
Dagenais
Affiliation: McGill University
Title: Recommending adaptive changes in the face of evolving software.
Abstract: Software constantly evolves and sometime, backward compatibility breaks and
developers must then adapt dependent programs. Unfortunately, adapting a
client program to support a new version of a used software can be a tedious
task and is a low value activity since no new features are introduced and
the quality of the program does not necessarily increase. We present the
ideas behind a recommendation system that 1) mine version histories of a
software to analyze evolution changes and 2) recommend ways (e.g. method
calls) to reconcile the client program with the new version of the used
software. During this talk, I will cover in more details how we can mine
software version histories, infer simple changes like refactoring and infer
complex changes using partial program analysis. Since this is still a work
in progress, comments on the approach will be more than welcome!
|
| 2007/03/21 |
General |
Place: MC437
Time: 9:30 - 10:30
Speaker:
Steve
Oudot
Affiliation: Stanford University
Area:
Computational Geometry
Title: Manifold Reconstruction Using Witness Complexes
Abstract: The problem of reconstructing manifolds from scattered data finds
applications in a number of areas of computer science, including image or
signal processing, computer graphics, robotics, machine learning, or
scientific visualization. It has been extensively studied by both the
computational geometry community and others. Ideas from computational
geometry have led to elegant solutions for the cases of sampled curves and
surfaces, based on the use of the Delaunay triangulation D(S) of the input
point set S. In these methods, an output simplicial complex is extracted
from D(S) that is guaranteed to be topologically equivalent to the
underlying manifold M, and geometrically close to M, provided that S
satisfies some mild sampling conditions.
After introducing these Delaunay-based methods, I will show that their nice
properties do not carry over to manifolds of arbitrary dimensions, since in
this more general setting the input point set may well-sample several
manifolds with different topological types. I will then introduce a novel
approach to manifold reconstruction, which builds a one-parameter family of
complexes approximating the input data at multiple scales. This approach is
motivated by recent advances in the theory of topological persistence,
showing that the "topology of a point cloud" depends on the scale at which
this point cloud is considered. To generate the family of complexes, our
algorithm uses a special construction, called the witness complex, which
mimics the Delaunay construction using only elementary geometric tests.
This makes the algorithm applicable in virtually any metric space. In
Euclidean spaces, its efficiency can be guaranteed theoretically, both in
terms of complexity and in terms of approximation. To support these claims,
I will show some experimental results obtained on real-life data sets from
computer graphics and medical imaging.
To conclude the talk, I will give some prospects for future research,
emphasizing on the obstacles and challenges of the sampling and analysis of
manifolds in arbitrary dimensions.
Biography of Speaker:
-
|
| 2007/03/14 |
Algorithms |
Place: MC437
Time: 9:30 - 10:30
Speaker:
Yufeng
Wu
Affiliation: Department of Computer Science, University of California, Davis
Area:
Recombination and Applications
Title: Algorithms for Inferring Recombination and AssociationMapping in Populations
Abstract: With increasingly available population-scale genetic variation data, a
current high priority research goal is to understand how genetic variations
influence complex diseases (or more generally genetic traits). Recombination
is an important genetic process that plays a major role in the logic behind
association mapping, a currently intensely studied method widely hoped to
efficiently find genes (alleles) associated with complex genetic diseases.
In this talk, I will present algorithmic and computational results on
inferring historical recombination and constructing genealogical networks
with recombination and applications to two biologically important problems:
association mapping of complex diseases and detecting recombination
hotspots.
On association mapping, I will present a method that generates the most
parsimonious genealogical networks uniformly and show how it can be applied
in association mapping. I will introduce results on evaluating how well the
inferred genealogy fits the given phenotypes (i.e. cases and controls) and
locates genes associated with the disease. Our recent work on detecting
recombination hotspots by inferring minimum recombination will also be
briefly described. For both biological problems, I will demonstrate the
effectiveness of these methods with experimental results on simulated or
real data.
Biography of Speaker:
-
Yufeng Wu received his Master degree in Computer Science from the
University of Illinois at Urbana-Champaign in 1998 and his Bachelor degree
from Tsinghua University, China in 1994. From 1998 to 2003 he was a
software engineer at a startup company in Illinois, USA. He currently works with Professor Dan Gusfield on algorithms in computational biology and
bioinformatics. His current research is focused on computational problems
arising in population genomics.
|
| 2007/03/12 |
General |
Place: MC437
Time: 9:30 - 10:30
Speaker:
Sylvain
Paris
Affiliation: Massachusetts Institute of Technology (MIT)
Area:
Graphics - HCI
Title: Exploiting the Richness of Digital Photographs and Videos
Abstract: Cameras have become popular with the recent development of digital
equipment. Today, it is extremely simple to take high-resolution pictures and movies, thereby giving easy access to an abundance of information. However, the size and number of acquired images challenge most existing algorithms. In this context, I will present my work on low-level image processing and computational photography.
I will first describe the bilateral filter, a nonlinear edge-preserving process which is becoming ubiquitous in computational photography. I will reformulate this filter in a higher-dimensional homogenous space and show that, using signal-processing arguments, it is possible to compute very quickly an approximation visually similar to the exact result. In a second part, I will use this technique to analyze digital photographs and enhance them by automatically transferring the visual qualities of an artist masterpiece. Finally, I will briefly present my on-going projects that extend these ideas to image segmentation and video processing.
Biography of Speaker:
-
Computer Science and Artificial Intelligence Laboratory (CSAIL)
|
| 2007/03/02 |
Software Engineering |
Place: McConnell 103
Time: 11:00 - 12:00
Speaker:
Richard
Halpert
Affiliation: McGill University - Sable Group
Title: Improving Lock Allocation with Supporting Analyses
Abstract: Multi-threaded programming using locks and monitors in Java (or any
other language) is difficult and error prone, even for experts. For
this reason, we have developed an automatic lock allocator. This lock
allocator could be used to enable a transactional version of the Java
language without the need for new code in the JVM, or simply to warn
programmers if their own lock allocation results in a program with
data races.
If a pair of lock regions have side effects to the same object, then
they are said to "interfere". A rudimentary solution to determining
which lock regions potentially interfere with each other is to inspect
the results of a Points-To analysis. Using more accurate Points-To
analyses can remove some spurious edges on the interference graph.
However, in order to remove the majority of spurious edges, more
thread-specialized analyses were needed.
|
| 2007/02/16 |
Software Engineering |
Place: McConnell 103
Time: 11:00 - 12:00
Speaker:
Dayong
GU
Affiliation: McGill University
Title: Hardware-related optimizations in a JVM
Abstract: Detecting repetitive "phases" in program execution is helpful for
program understanding, runtime optimization, and for reducing
simulation/profiling workload. The nature of the phases that may be
found, however, depend on the kinds of programs, as well how programs
interact with the underlying hardware.
We present a technique to detect program phases by monitoring
microarchitecture-level hardware events. We also
show a set of sample adaptive optimizations of our phase
detection technique.
|
| 2007/02/02 |
Software Engineering |
Place: McConnell 103
Time: 11:00 - 12:30
Speaker:
Ekwa
Duala-Ekoko
Affiliation: McGill University
Title: Tracking Code Clones in Evolving Software
Abstract: Code clones - source code regions with identical syntax and semantics -
are generally considered harmful in software development, and the
predominant approach is to try to eliminate them through refactoring.
However, recent research has provided evidence that it may not always
be practical, feasible, or cost-effective to eliminate certain clone
groups. In this talk, I will propose a technique that enables
developers to maintain clone groups of interest, and to monitor and
support modification tasks that intersect with the documented clone
model as a system evolves. Our technique relies on the concept of
abstract clone region descriptors (CRD), which describe clone regions
within methods in a robust way that is independent from the exact text
of the clone region or its location in a file. Next, I will present our
definition of CRDs, and describe a complete clone tracking system
capable of producing CRDs from the output of a clone detection tool,
notify developers of modifications to clone regions, and support the
simultaneous editing of clone regions. Finally, I will discuss the
results of two experiments and a case study conducted to assess the
performance and usefulness of our approach. (This is a joint work with Martin Robillard - Assistant Professor, McGill)
|
| 2007/01/15 |
Bioinformatics |
Place: Duff Medical Building, 3775 University Street, Main Amphitheatre 1
Time: 15:30 - 16:30
Speaker:
Corey
Yanofsky
Affiliation: Biomedical Engineering, McGill University
Area:
proteomics experiments
Title: Peptide retention time prediction from high-throughput proteomic data
Abstract: In high-throughput proteomics experiments, complex mixtures of peptides usually undergo chromatographic separation prior to identification by mass spectrometry. Chromatographic separation was originally incorporated into the experimental design principally to limit the complexity of the peptide mixture entering the mass spectrometer at any one time. Over the course of years, large databases have been built to store information, including chromatographic retention time information, about peptides identified experimentally. This creates the possibility of using retention time data as an extra dimension to improve the identification of peptides based on mass spectra. However, there is considerable variability in the elution time for a particular peptide which must be corrected before the information can be use for peptide identification. At McGill University, many thousands of proteomic experimental runs have been run using a variety of chromatographic protocols. These runs have different chromatographic dead times, elution gradients, and experimental precisions, all of which interfere with a straightforward prediction of future retention times from previous data. To overcome these complications, we have developed a Bayesian model to estimate the physical property of peptides underlying experimental retention times, explicitly modelling varying chromatographic dead time, elution gradient, and run-specific variance as separate effects. The model was fit using a data set of 113160 peptide identifications, comprising 6681 unique peptides in 3163 runs, thus providing estimates of the “true” retention time (relative to a chosen reference run) of the 6681 peptides. A cross-validation study was performed, showing that the predictive error of the model was typically wthin +/- 2 minutes. (This covers about 8% of the total elution time for a run with a 60 minute gradient and 10 minutes of dead time.) A second stage was necessary to predict the retention times of peptides which are not present in the data set, so the results of the first stage were used to fit a peptide-sequence-based retention time model. In a leave-one-out cross-validation study, the sequence-based model was able to predict retention times with an overall accuracy of +/- 5 minutes, which covers about 20% of the total elution time for a run with a 60 minute gradient and 10 minutes of dead time.
|
| 2007/01/04 |
General |
Place: MC13
Time: 15:30 - 16:30
Speaker:
last
test
Affiliation: test
Area:
tests
Title: testing
Abstract:
Biography of Speaker:
-
tet
|
|
