Seminar Home
Fall 2012 Schedule
Winter 2013 Schedule
Summer 2013 Schedule
Archives
|
Date
(
Winter 2012
) |
Category |
Seminar Info |
| 2012/04/27 |
Graduate Seminar Series |
Place: MC103
Time: 11:30 - 12
Speaker:
Andrew
Cave
Affiliation: McGill, PhD student of Prof. Brigitte Pientka
Area:
Computation and Logic
Title: What programming language research can do for you
Abstract: Research in types and programming languages often advertises safety as
its key feature. But let's be honest: for many programmers, safety
isn't something to get excited about. In this talk we will explore
what types are for, not just what they are against. We will see some
examples of ways in which types guide programming: how they can enable
the machine to find some of our programs for us, how they can cut down
the need for unit tests and documentation (in surprisingly
sophisticated ways!), and how they can provide immediate answers and
reassurance for some of the complex questions we ask ourselves when
programming or reading code (is this case really impossible? is this
function evidently terminating? does this function maintain the
datastructure invariant?). The general theme is that types enable us
to offload to the machine much of the routine thought (and
increasingly not-so-routine thought!) that goes into programming. This
informs tool and IDE design, turning programming into an interactive
refinement process, with help and feedback from the machine at every
step.
|
| 2012/03/29 |
Vision, Graphics, and Robotics |
Place: ENGMC 437
Time: 11:00 - 12:00
Speaker:
Scott
McCloskey
Affiliation: Honeywell Labs (Minneapolis MN)
Area:
Computational Photography
Title: Capture of Sharp Images from Moving Objects , with Application to Iris Biometrics
Abstract: This talk addresses the capture of sharp images of moving objects using a computational photographic technique known as coded exposure, and summarizes recent papers from the European Conference on Computer Vision (ECCV) and International Conference on Computer Vision (ICCV). We extend the flutter shutter capture work of Raskar, et al. by first demonstrating that a given shutter sequence has a limited range of velocities over which it produces invertible motion blur, and then by proposing an algorithm to find velocity-specific sequences. In order to enable a fully-automated flutter shutter system, we further present the first automatic blur estimation technique which is applicable to coded exposure images, and demonstrate its effectiveness over a range of motion types: constant velocity, constant acceleration, and
periodic motion. Coded exposure using a temporal sequence of flash illumination is also analyzed, and we demonstrate its improved robustness to velocity variation. Finally, the talk will demonstrate the utility of flutter shutter image capture to a specific image exploitation task: iris capture for biometric identification.
Biography of Speaker:
-
Scott McCloskey joined Honeywell’s Integrated Security Technology section in 2007 as a Computer Vision Research Scientist, and has since served as the principle investigator for several government-funded projects from DARPA, IARPA, and the Biometrics Task Force (BTF). His research interests include computational photography, computer vision, and biometrics. Dr. McCloskey received his PhD in Computer Science from McGill University in 2008, a MS in Computer Science from the Rochester Institute of Technology in 2002, and a BS in Computer Science and Math from the University of Wisconsin-Madison in 1998. He is the author of several peer-reviewed conference papers including recent work at CVPR, ECCV and ICCV. As an active part of the computer vision research community, Dr. McCloskey has served as a panelist for the National Science Foundation (NSF), on the program committee for the IEEE Workshop on Applications of Computer Vision, and as reviewer for numerous conferences and journals.
|
| 2012/03/29 |
CQIL - Cryptography and Quantum Information |
Place: McConnell 103
Time: 14:30 - 15:30
Speaker:
Saikat
Guha
Affiliation: BBN Technologies
Title: Capacity of quantum reading
Abstract: This talk will focus on a quantum analysis of optical reading of classical data encoded in a reflective memory that encodes information via passive amplitude and phase modulation of an optical probe beam, and detecting the modulated light via a suitable receiver. We will address fundamental quantum-information-theoretic capacity bounds and the gap in performance between conventional systems and the theoretical bounds, as well as explicit codes and structured optical receivers aimed at achieving those performance bounds. It turns out that there is no fundamental upper limit to the number of information bits that can---in principle---be read reliably per expended probe photon when reading classical data encoded in a memory made of reflective pixels that passively modulate the probe light. We show that using a noiseless coherent-state (classical laser light) probe, an on-off amplitude-modulation pixel encoding, and direct (intensity) detection at the receiver (an idealized model of commercial CD/DVD disc drives), the highest photon information efficiency achievable in principle, is about half a bit per transmitted photon. We will discuss how a coherent-state probe can in principle read unlimited number of bits per photon when the receiver is allowed to make collective (joint) measurements on the reflected light from a large block of memory cells, and when phase modulation is used at the memory cells. We will then show how a spatially-entangled non-classical light probe can read any number of error-free bits of information as we please. Finally, we will discuss how an entangled-state probe, along with a sequential-decoding receiver that uses structured optical elements plus a binary quantum non-demolition measurement, can attain the Holevo bound. We will close the talk with open problems, and a discussion of ongoing work.
|
| 2012/03/23 |
Graduate Seminar Series |
Place: MC103
Time: 11:30 - 12
Speaker:
Xi
Chen
Affiliation: McGill, PhD candidate of Profs Xue Liu and Xiao-Wen Chang
Area:
Computer Networks
Title: QoS control in networked real-time systems
Abstract: This week I will give a talk about the quality of service (QoS) control in networked real-time systems. Networked real-time systems are systems with multiple processors connected to execute the real-time applications. These systems are widely applied in our real world, i.e. environment monitoring systems, drive-by-wire automotive systems, modern server clusters etc. In the networked real-time systems, the QoS metrics mainly include schedulability of the real-time applications , efficiency of the system resources usage and predictability of the system workload. To provide good QoS in the networked real-time systems, we need to design efficient resource allocation strategies to ensure most or all of the real-time tasks can meet their deadlines and at the same time the system resources can be effectively utilized even under the workload uncertainties or variations during the runtime. My research focuses on providing good QoS in networked real-time systems using feedback-based resource allocation strategies, where the system performance is leveraged as a feedback to estimate the system workload online and adjust the system resources allocated to the real-time applications. In this talk, I will present the problems I work on in QoS control area, including the system model formulation, the resource allocation strategy design and some preliminary results.
|
| 2012/03/16 |
Graduate Seminar Series |
Place: MC103
Time: 11:30 - 12
Speaker:
Annie
Ying
Affiliation: McGill, PhD student of Prof. Martin Robillard
Area:
Software Engineering
Title:
Abstract: Why do we care about negative results? After all, the halting problem is a negative result! I will share with you three experiences I had with negative results and how these results have positively shaped my current research. I would also like to hear your experiences with negative results.
|
| 2012/03/09 |
Graduate Seminar Series |
Place: MC103
Time: 11:30 - 12
Speaker:
Karyn
Moffatt
Affiliation: McGill, Assistant Professor in the School of Information Studies
Area:
Human Computer Interaction
Title: How <strike>Users</strike> Interfaces Fail: A Guide to the Human-Computer Interface
Abstract:
This talk will appeal to developers and users alike. For users, it will
examine why computer interfaces are so notoriously frustrating. And why
tools meant to make life easier, instead frustrate and annoy. For
developers, it will explain why users are so notoriously frustrating. And
why tools meant to make their lives easier, instead confuse and bewilder
them. For everyone, it will present an overview of the challenges of user
interface design and evaluation, and an introduction to the tools and
methods of the domain.
Biography of Speaker:
-
Dr. Karyn Moffatt is an assistant professor in the McGill School of
Information Studies, where she studies human-computer interaction,
particularly with respect to interface design for older adults. Prior to
joining McGill she held a postdoctoral fellowship with the Technologies for
Aging Gracefully Lab (TAGLab) at the University of Toronto. She earned her
PhD in Computer Science from the University of British Columbia in 2010.
|
| 2012/03/02 |
Graduate Seminar Series |
Place: MC103
Time: 11:30 - 12
Speaker:
Pablo
Duboue
Affiliation: Les Laboratoires Foulab
Area:
Natural Language Processing
Title: Computer Science in the times of high error rate: survival tips from the NLP trenches
Abstract: Natural language processing is receiving a lot of interest of late due
to a number of reasons: Maturity of some techniques in the field, showcased by some high profile successes; available computational power (both in terms of number of cores and accessibility to huge amount of storage); and increased availability of data.
In this talk, I will give a few insights regarding how to approach NLP
work for complete newbies, which can be summarized into three major
points: (1) Learn a toolkit (or more) and develop intuitions about the type of
errors it makes. (2) Look at the data. (3) Don't be afraid of annotating data yourself.
I have been working in NLP for more than 15 years and I have always
been on the hands-on, practical side of things.
Look forward meeting you all, you can expect something more
interactive than your usual seminar.
Feel free to bring plenty of questions and be ready to answer my very own.
Biography of Speaker:
-
Dr. Duboue is an independent language technologist. His work focuses
on applied language technology and natural language generation. He
received a Licenciatura en Computacion degree from Cordoba University
(Argentina) in 1998 and M.S., M.Phil and Ph.D. degrees in Computer
Science from Columbia University in the City of New York in 2001, 2003
and 2005. He is passionate about improving society through language
technology and splits his time between teaching, doing research and
contributing to free software projects. He had taught at Cordoba
University, Columbia University, Siglo21 University and had worked for
IBM TJ Watson Research as a Research Staff Member. Past projects he had actively contributed include the Watson QA system
showcased in the IBM Jeopardy Challenge and the Debian Developers
Conference in New York City. He is currently an independent consultant
living in Montreal, Quebec.
|
| 2012/02/03 |
Graduate Seminar Series |
Place: MC103
Time: 11:30 - 12
Speaker:
Peter
Rigby
Affiliation: Mcgill, PostDoc with Prof. Martin Robillard
Area:
Software Engineering
Title: Peer Review on Open Source Software Projects
Abstract:
Peer review is seen as an important quality assurance mechanism in both
industrial development and the open source software (OSS) community. In
contrast to industry, where reviews are typically assigned to specific
individuals, in OSS, changes are broadcast to hundreds of potentially
interested stakeholders. What is surprising is that this approach works
very well, despite concerns that reviews may be ignored, or that
discussions will deadlock because too many uninformed stakeholders are
involved.
In this talk, I will describe the multi-case study that I used to develop a
theory of OSS peer review. The theory is developed in three research
stages: review processes, parameters and statistical models, and underlying
behaviours and mechanisms of OSS peer review. I find that OSS reviews can
be described as (1) early, frequent reviews (2) of small, independent,
complete contributions (3) that, despite being asynchronously broadcast to
a large group of stakeholders, are reviewed by a small group of
self-selected experts (4) resulting in an efficient and effective peer
review technique. I am currently working with the DND, AMD, and Microsoft
to understand how lightweight peer review can be adopted into industry.
Biography of Speaker:
-
Peter C. Rigby is a postdoctural researcher working with Dr. Robillard at
McGill University in Montreal. He received his PhD from the University of
Victoria for his examination of the peer review practices used by OSS
projects. His PhD advisers were Dr. Storey and Dr. German. He received a
Bachelor degree at the University of Ottawa in Software Engineering and has
taught two third year Software Engineering courses (Software Maintenance
and HCI). His primary research interest is in Mining Empirical Software
Engineering data to understand the how people collaborate to design and
develop large, successful software systems. His three current research
areas are: informal API documentation quality, lightweight industrial
review techniques, and the affect distributed version control is having on
developer collaboration. Please see helium.cs.uivc.ca for more details.
|
| 2012/01/25 |
CQIL - Cryptography and Quantum Information |
Place: Rutherford 103
Time: 13:00 - 14:00
Speaker:
Patrick
Hayden
Affiliation: McGill University
Title: Entanglement, entropy and quantum error correction
Abstract: A search of the cond-mat arXiv reveals hundreds of papers containing
"entanglement entropy" in their titles, but what's so special about this
recently fashionable function? I'll give a quick introduction to the
information theoretic interpretation of entanglement entropy, through its
connection to compression and quantum error correction. Time-permitting,
I'll describe the special behaviour of entanglement entropy in holographic
field theories and what can be learned from it.
|
| 2012/01/20 |
Graduate Seminar Series |
Place: MC103
Time: 11:30 - 12
Speaker:
Mathieu
Boespflug
Affiliation: Mcgill, PostDoc with Prof. Brigitte Pientka
Area:
Logic and programming languages
Title: First class reflection for shorter proofs
Abstract: An increasingly common and relied upon feature of programming
languages is the view according to which functions - all functions,
defined locally or globally, named or unnamed - are values just like any
other. It is becoming well understood that making functions first class
citizens of your programming language in this way greatly facilitates
modularity and code reuse. But what is less well known is that first
class functions are a boon to everyone who wishes to represent and
manipulate programs as data, such as in compiler and program generators.
In this talk, I'll discuss encoding programs using a datatype with
embedded functions, and show that this same encoding can be used to
represent theorem statements in an interactive proof assistant equally
well. The ability to represent theorem statements and manipulate them as
data - or /reflect/ on them - will allow us to throw away long winded
hand written proofs of some theorem statements and conveniently replace
these proofs with short programs that analyze a theorem statement to
answer with "true" or "false", depending on whether the theorem
statement is provable, but in such a way that the short programs are
just as trustworthy as the long winded hand written proofs they replace.
|
| 2012/01/18 |
CQIL - Cryptography and Quantum Information |
Place: Rutherford 103
Time: 13:00 - 14:00
Speaker:
Mark
Wilde
Affiliation: McGill University
Title: Potential Consequences of Closed Timelike Curves for Quantum Information Processing
Abstract: Einstein's field equations for general relativity predict the existence of
closed timelike curves (CTCs) in certain exotic spacetime geometries, but the
bizarre consequences lead many physicists to doubt that such "time machines" could
exist. CTCs, if they existed, would allow particles to interact with their former
selves, suggesting the possibility of grandfather-like paradoxes in both classical
and quantum theories. Quantum information theorists have devised self-consistency
conditions to avoid such grandfather-like paradoxes, by either enforcing
self-consistency on the density matrix of the chronology-violating qubits or by
considering a "postselected quantum teleportation protocol" which eliminates the
possibility of such paradoxes.
In this talk, I will describe a simple protocol that allows for a "CTC-assisted"
party to violate the uncertainty principle in both of the above CTC models (though,
I should clarify that this result depends upon your viewpoints for how to describe
quantum states resulting from preparations or measurements). In both CTC models, it
is even possible to violate the no-signaling principle (again depending upon your
viewpoints). If you consider these models to be legitimate ways of imposing causal
self-consistency, then any future theory of quantum gravity should reconcile this
intersection of quantum information theory and general relativity.
This talk is intended for non-specialists and should be comprehensible to anyone
familiar with quantum mechanics.
|
| 2012/01/13 |
Graduate Seminar Series |
Place: MC103
Time: 11:30 - 12
Speaker:
Jordan
Frank
Affiliation: Mcgill
Title: Machine learning using time delay embeddings
Abstract: I will discuss my research on machine learning with time
series data. Our approach is based on a technique called time delay
embedding, which has been used for nonlinear time series analysis in
various fields (e.g., statistical physics, computational physiology).
We extend this approach to machine learning problems such as
classification and clustering of time series data. I will present the
results of experiments showing that our approach is competitive, and
often outperforms, existing state-of-the-art approaches, while being
more computationally efficient. Time-permitting, I will demonstrate
how our classifier can detect people based on their gait, via data
collected from accelerometer sensors in smartphones (audience
participation required).
|
|
