Seminars

201006

Spoken Language Processing: Where Do We Go From Here?

Professor Roger Moore, Department of Computer Science, University of Sheffield
June 21, 2010
11:00 am
McGill University, McConnell Engineering, room 603

* McGill Seminar in Speech and Language Technology

Abstract

Recent years have seen steady improvements in the quality and performance of speech-based human-machine interaction driven by a significant convergence in the methods and techniques employed. However, the quantity of training data required to improve state-of-the-art systems seems to be growing exponentially, yet performance appears to be reaching an asymptote that is not only well short of human performance, but which may also be inadequate for many real-world applications. This suggests that there may be a fundamental flaw in the underlying architecture of contemporary systems, and the future direction for research into spoken language processing is currently uncertain.

This talk addresses these issues by stepping outside the usual domains of speech science and technology, and instead draws inspiration from recent findings in the neurobiology of living systems. In particular, four areas will be discussed: the growing evidence for an intimate relationship between sensor and motor behaviour in living organisms, the power of negative feedback control to accommodate unpredictable disturbances in real-world environments, mechanisms for imitation and mental imagery for learning and modeling, and hierarchical models of temporal memory for predicting future behaviour and anticipating the outcome of events.

The talk will conclude by showing how these results point towards a novel architecture for speech-based human-machine interaction that blurs the distinction between the core components of a traditional spoken language dialogue system; an architecture in which cooperative and communicative behaviour emerges as a by-product of a model of interaction where the system has in mind the needs and intentions of a user, and a user has in mind the needs and intentions of the system.

Roger Moore has been professor of spoken language processing in the Speech and Hearing Research Group at the University of Sheffield since 2004 and is a visiting professor in the Department of Linguistics and Phonetics at University College London. Previously, he was chief scientific officer at 20/20 Speech and before that Head of the UK government Speech Research Unit. He is also past president of both the International Speech Communication Association and the Permanent Council of the International Conferences on Spoken Language Processing. He is currently the editor of the journal of Computer Speech and Language.

Top of page

201004

Canonical Representations and their Application to Formal Verification and Synthesis

Prof. Maciej Ciesielski, University of Massachusetts
April 22, 2010
11:00 am
McGill University, McConnell Engineering, room 603

* MITACS/IML seminar

Abstract

This talk will provide a review of several canonical, graph-based representations used in formal verification and synthesis of digital systems. These representations, commonly known as "decision diagrams", include Binary Decision Diagrams (BDD), Binary Moment Diagrams (BMD), and Taylor Expansion Diagrams (TED).

The emphasis of this talk is on TED, a recently invented canonical representation for dataflow and computation-intensive designs, used in digital signal processing and computer graphics applications. TED representation is based on a word-level rather than binary decomposition principle, and as such can represent designs on higher levels of abstraction. Several applications of TEDs are discussed, including: equivalence checking, high-level transformations, and high-level synthesis of designs specified on algorithmic and behavioral levels.

Maciej Ciesielski is Professor in the Department of Electrical & Computer Engineering (ECE) at the University of Massachusetts, Amherst. He received M.S. in Electrical Engineering from Warsaw Technical University, Poland, in 1974 and Ph.D. in Electrical Engineering from the University of Rochester, N.Y. in 1983. From 1983 to 1986 he worked at GTE Laboratories on a silicon compilation project. He joined the University of Massachusetts, Amherst in 1987, where he teaches and conducts research in the area of electronic design automation (EDA), and specifically in synthesis, optimization and verification of digital systems. He is recipient of Doctorate Honoris Causa from the Université de Bretagne Sud, Lorient, France.

Top of page

201003

Network Science for Communications

Ali Tizghadam, University of Toronto
March 5, 2010
11:00 am
McGill University, McConnell Engineering, room 603

* MITACS seminar

Abstract

Network Science is a newly emerging scientific discipline that investigates common attributes and differences (static and dynamic) among diverse physical and engineered networks, such as social networks, biological networks, Internet, and communication networks. Generally speaking, Network Science is the combination of Graph Theory, Control Theory, and cross discipline applications.

In this talk we focus on the application of Network Science in Communication Networks. One important aspect of data communication networks is the flow of traffic between different nodes. The notion of traffic is not explicitly entered into the basic concepts of Network Science, since the focus of Network Science has been more on social networks. We show one approach to explicitly account for the effect of traffic. We extend the definition of betweenness centrality and introduce traffic aware betweenness (TAB).

We investigate the properties of TAB and show that it is directly related to some important measures in data communication networks. More specifically, we show that the average network utilization is directly proportional to the average normalized traffic aware betweenness, which is referred to as traffic aware network criticality (TANC). We investigate the behavior of TANC and verify that TANC is a linear function of point to point effective resistances of the graph. As a result, normalized TAB is a convex function of link weights and can be minimized using convex optimization techniques. We study the mathematical properties of the optimization problem and derive useful results to be used in network planning and traffic engineering purposes in wired and wireless domain.

Ali Tizghadam is currently a Post Doctoral research fellow at the University of Toronto. He is also managing the Network Architecture Lab (NAL) in the Communications group. He received his M.A. Sc in 1994 for University of Tehran in Electrical Engineering. After graduation he went to the industry for about 10 years where he gained experience in telecommunications especially local exchange switches and access networks. He then came back to the university to pursue his PhD studies in the Electrical and Computer Engineering department, at the University of Toronto. His research interests span the areas of Network Control, Virtualization, Green Communications, Network Resource Management, Optimization Theory, Autonomic Networking, and their applications in different networking areas.

Top of page

201003

Stability and Distributed Power Control in MANETs with Outages and Retransmissions

Anastasios Giovanidis, Fraunhofer Heinrich Hertz Institute for Telecommunications
March 4, 2010
2:00 - 3:30
McGill University, McConnell Engineering building, room 603

* MITACS seminar

Abstract

The current talk investigates the effects of hop by hop packet loss and retransmissions via ARQ protocols within a Mobile Ad hoc NET work (MANET). Errors occur due to outages and a success probability function is related to each link, which can be controlled by power and rate allocation. We first derive the expression for the network's capacity region, where the success function plays a critical role. Properties of the latter as well as the related maximum goodput function are presented and proved.

A Network Utility Maximization problem (NUM) with stability constraints is further formulated which decomposes into (a) the input rate control problem and (b) the scheduling problem. Under certain assumptions problem (b) is relaxed to a weighted sum maximization problem with number of summants equal to the number of nodes. This further allows the formulation of a non cooperative game where each node decides independently over its transmitting power through a chosen link.

Use of supermodular game theory suggests a price based algorithm that converges to a power allocation satisfying the necessary optimality conditions of (b). Implementation issues are considered so that minimum information exchange between interfering nodes is required. Simulations illustrate that the suggested algorithm brings near optimal results.

Anastasios Giovanidis received the Diploma in Electrical and Computer Engineering from the National Technical University of Athens, Greece in 2005. From Sept. 2005 to Feb. 2010 he has been with the Fraunhofer Heinrich Hertz Institute for Telecommunications in Berlin, Germany, working as a Research Associate, while pursuing his Dr. Ing. degree from the Technical University of Berlin (defence to take place in April 2010). His research interests include stochastic control, optimization and probability theory applied to telecommunication systems with emphasis on queuing networks.

Top of page

201002

Temporal Debugging using URDB

Gene Cooperman, Northeastern University
February 11, 2010
10:30 am
McGill University, MacDonald Engineering, room 357

* MITACS seminar

Abstract

This talk presents URDB, a universal reversible debugger. URDB supports reversible debugging for: gdb, MATLAB, python, and perl. The standard reversible commands (for example, reverse next, reverse step, reverse continue, reverse finish) are all provided. The novelty is two fold: the approach of debugging entire multi process debugging sessions; and temporal search to automate certain types of debugging.

In order to checkpoint entire debugging sessions, DMTCP (Distributed MultiThreaded CheckPointing) has been extended to checkpoint ptrace based debuggers such as gdb. We believe this is the first time that anyone has been able to checkpoint a gdb session (both gdb and target process).

The current showcase for temporal search is efficient expression watchpoints: the ability to execute binary search through a process lifetime to see when an expression becomes bad. For example, a debugging session has discovered that a certain expression has become bad. The expression originally had a good value. One wishes to exhibit the bug. To do so, one must stop in the debugger just before the expression is about to change from good to bad. (There may be many such times when this happens since the bug may express itself many times, but the user requires only to observe one occurrence of this bug.) As an example, perhaps a user defined testing function has tested an internally represented lattice graph and discovered that a cycle exists. URDB then brings the user of the debugging session to a point in time when the user test function reports that there is no cycle, and that the execution of the next statement will cause the test expression become bad (reports a cycle).

In a computation of N statements, URDB needs to execute the test function only O(log N) times. A traditional expression watchpoint implementation by a debugger such as gdb requires the test function to be executed O(N) times. Current work in progress is developing the ability to reversibly execute and hence reversibly debug both MPI (Message Passing Interface for parallel computing) and X Windows. The work reported here is joint with Kapil Arya, Tyler Denniston, Xin Dong, Artem Polyakov, Praveen S. Solanki, and Ana Maria Visan.

Gene Cooperman received his Ph.D. from Brown University in 1978. He spent two years as a post doc, followed by six years at GTE Laboratories. He has been a professor at Northeastern University since 1986, and a full professor since 1992. His interests lie in high performance computation and symbolic algebra. The combination of these two subjects has also led to his joint work with his students on a second theme: disk based parallel computation. He leads the High Performance Computing Laboratory at Northeastern University, where he currently advises five PhD students. He has over 80 refereed publications.

Top of page