TALKS

Name: Michelle Spina

Email: spina@sls.lcs.mit.edu

Group: Spoken Language Systems Group, M.I.T.

Title: Automatic Transcription of General Audio Data: Effect of Environmental Segmentation on Phonetic Recognition

Statement:

In addition to the vast majority of text-based information available on the World Wide Web, an increasing amount of video and audio based information is becoming available to users as a result of emerging multimedia computing technologies. The addition of these multimedia sources of information have presented us with new research challenges. Mature information retrieval methods have been developed for the problem of finding relevant items from a large collection of text-based materials given a query from a user. Only recently has there been any work on similarly indexing the content of multimedia sources of information. If we concentrate on general audio data from sources such as radio, television, movies, meeting recordings, etc., one natural solution to this problem would be to generate a full transcription of the audio, in effect converting the audio data to text data, such as is done with closed captioning. Traditional text-based information retrieval techniques could then be used directly on the transcriptions. In this talk, I will concentrate on the automatic transcription of general audio data. The task of transcribing general audio data is very different from those usually confronted by current automatic speech recognition systems. Here, we determine the optimal training strategies for recognizing such data. Specifically, we have studied the effects of different speaking environments on a phonetic recognition task using data collected from a radio news program.

Name: Dimitris Hristu Email: hristu@arcadia.harvard.edu Group: Division of Applied Sciences, Harvard Title: Control with Limited Communication" Statement: This work is an attempt to bring together in an elegant manner, issues in control and communication. We consider a class of continuous-time systems where inputs and outputs must be communicated from and to a controller, over a bus of finite capacity. The controller must typically spend some of its time performing computations and the dimension of the bus may be less than that of the input/output signals. Consequently, communication with the control system occurs at discrete times and the controller must choose which inputs/outputs to update/read at a particular time. This leads to a theory of sampled-input continuous systems where communication and control are naturally coupled. The theory allows us to solve tracking and stabilization problems as well as to define and represent ``attention'' in the context of control systems with limited communication. The theory has relevance in areas where system complexity is an issue, as in the coordinated control of aerial vehicles (UAV's), MEMS devices or any other setting where many systems must share the attention of a decision--maker. In addition, it can be a promising framework for posing questions related to optimal communication and polling versus ``greedy'' communication policies.

Name: Christoforos Hadjicostis Email: chadjic@allegro.mit.edu Group: LEES/DSPG, M.I.T. Title: Fault-Tolerant Monitoring of Discrete Event Systems using Petri Net Embeddings (see Statement at the end of email) Statement: Petri nets are a powerful graphical and mathematical model for a variety of information and processing systems. In this talk, I will present a methodology for monitoring failures and other activity in discrete event systems that are described by Petri nets. The method is based on embedding the given Petri net model into a larger Petri net which retains the functionality and properties of the given one, perhaps in a {\em non-separate} (that is, not immediately identifiable) way. This {\em redundant Petri net embedding} introduces ``structured redundancy'' which can be used to facilitate fault detection, identification and correction, or to offer increased capabilities for monitoring and control. I will focus primarily on {\em separate embeddings} in which the functionality of the original Petri net is retained in its exact form. Using these embeddings, I will construct monitors that detect and identify different types of failures by performing consistency checks between the state of the original Petri net and that of the monitor. The methods that I propose are attractive because the resulting monitors are robust to failures, they may not require explicit acknowledgments from each activity, and their construction is systematic and easily adaptable to restrictions in the available information.

Name: Stark Draper Email: scd@sllegro.mit.edu Group: DSP, M.I.T. Title: Multiple Source Descriptions via Lossy Lempel-Ziv Statement: The objective of the multiple source descriptions problem is to formulate a number of encodings of a source such that each encoding meets a minimal fidelity description on its own, but the more descriptions received the higher fidelity can be achieved. We think about approaching this problem through a modification of the Lempel-Ziv algorithm.

---------------------------------------------------------------------- Name: Ilya Pollak Email: ipollak@mit.edu Group: SSG, MIT Title: A Nonlinear Diffusion Equation as a Fast and Optimal Solver of Edge Detection Problems. Statement: A nonlinear diffusion process known to be effective for image segmentation is analyzed in 1-D. It is shown that it provides maximum likelihood solutions for certain edge detection problems. This fact means that classical change detection results can be used for performance analysis in some special cases. The robustness of the algorithm is also analyzed using an H-infinity-like criterion. The algorithm's connections with linear programming are explored. A fast implementation of the algorithm is introduced.

--------------------------------------------------------------------------- Name: Chalee Asavathiratham Email: chalee@mit.edu Group: LEES, MIT Title: Multidimensional Time-Frequency Uncertainty Principle Statement: The time-frequency uncertainty principle (UP) establishes the limit to which a signal can be simultaneously concentrated both in the time and frequency domain. The main result of this paper is the extension of the UP to $m$-D signals by expressing the bound in the form of matrix inequalities. By associating the covariance matrix of a given signal to an ellipsoid of effective support, geometrically insightful interpretations of the $m$-D UP can be made. The proof of the main result is based on operator-theory, which allows us to easily derive a similar set of inequalities for $m$-D discrete-time signals.

----------------------------------------------------------------------------- Name: Nick Laneman Email: jnl@allegro.mit.edu Group: DSPG, MIT Title: Temporal Diversity for Fading Channels Statement: Frequently, system designers employ an error-correcting code followed by an interleaver to achieve temporal diversity over time-selective fading channels. This talk will discuss efficient, rate-preserving alternatives to coding plus interleaving that take the form of linear transformations (blockwise or convolutional) of the input symbol sequence. These schemes lend themselves to low-complexity decoding methods, such as linear equalization followed by symbol-by-symbol detection. We are examining the performance vs. complexity tradeoffs associated with several suboptimal decoding techniques for these schemes, and exploring combinations of these particular temporal diversity techniques with linear spectral and spatial diversity techniques. We are also searching for other, possibly nonlinear, transformations which offer significant diversity benefit for similar computational cost.

----------------------------------------------------------------------- Name: Michael Neely Email: mjneely@mit.edu Group: LIDS, MIT Title: Queue occupancy in single server, deterministic service time tree networks. Statement: Tree networks of single server, deterministic service time queues are often used as models for packet flow in systems with ATM traffic. In this tTitle: Universal Decoding alk, we present a simple method of analyzing packet occupancy in these systems without making any assumptions about the nature of the underlying input processes. We demonstrate how analysis of multi-stage tree systems can be reduced to the analysis of a much simpler 2-stage equivalent model. We also develop an expression for first moments of queue occupancy in terms of first moments of a simple 1-stage equivalent model. From this, we observe an interesting phenomenon for general types of "distributable inputs": Expected occupancy at any queue within a multi-stage tree network will be a concave function of the loadings produced by the multiple exogenous inputs.

---------------------------------------------------------------------- Name: Sridevi Sarma Email: sree@mit.edu Group: CSG, MIT Title: Analysis of Cerebellar-Mediated Wave-Variable Servo Control System Statement: The peformance and stability of a cerebellar-mediated wave-variable servo control system proposed by Steve Massaquoi will be addressed. In addition, validity of model assumptions will be assessed.

---------------------------------------------------------------------- Name: Thierry Klein Email: teklein@mit.edu Group: LIDS, MIT Title: Universal Power Control for Finite-State Statement:

----------------------------------------------------------------------- Name: Aaron Cohen Email: Group: Title: Universal Decoding Statement:

--------------------------------------------------------------------- Name: Babak Ayazifar Email: babak@mit.edu Group: LEES, MIT Title: Statement:

--------------------------------------------------------------------- Name: Anand Ganti Email: aganti@mit.edu Group: LIDS, MIT Title: Mismatch Channel Capacity per Unit Cost Statement: The Mismatch Channel Capacity per Unit Cost represents the maximum number of bits per unit cost that can be transmitted reliably across a channel under conditions of mismatch. It's reciprocal gives the cheapest cost of transmitting a bit reliably. We derive lower bounds for the Mismatch Channel Capacity per Unit Cost as the cost tends to zero and discuss some of its properties.

--------------------------------------------------------------------- Name: Jorge Goncalves Email: jmg@mit.edu Group: CSG, MIT Title: Analysis of Relay Feedback Systems Statement: This paper presents sufficient stability conditions of limit cycles for relay feedback systems. Local stability conditions exist. These are based on analyzing the linear part of the Poincare map. We know that when a certain limit cycle satisfies those local conditions, a neighborhood around the limit cycle exists such that any trajectory starting in this neighborhood converges to the limit cycle as time goes to infinity. However, tools to characterize this neighborhood do not exist. In this work, we present conditions, in the form of Linear Matrix Inequalities (LMIs), that guarantee the stability of a limit cycle in a reasonably large set around it. These results differ from previous local results as they take into account the high order terms of the Poincare map.

-------------------------------------------------------------------- Name: Hisham Kassab Email: Group: Title: Routing in Packet Radio Networks Statement:

----------------------------------------------------------------------- Name: Andrew Kim Email: Group: Title: Non-parametric Estimation of Anisotropic Statement:

------------------------------------------------------------------------- ????????????????????????????? Name: Asuman Koksal Email: Group: Title: Statistical Object Recognition Statement:

------------------------------------------------------------------------ Name: Can Emre Koksal Email: Group: Title: Communication Networks Theory Statement:

------------------------------------------------------------------------ Name: Navid Sabbaghi Email: Group: Title: Verification and Recognition Statement:

------------------------------------------------------------------------- Name: Paul Sotiriadis Email: Group: Title: Asynchronous (Chaotic) Computations Statement:

----------------------------------------------------------------------- Name: Elif Uysal Email: elif@mit.edu Group: LIDS, MIT TITLE: Maximizing Diversity in Slow Frequency Hopping Statement: This talk will address situations where channel memory is kept over a finite block length. A motivating example is slow frequency hopping in wireless communication, in particular, SFH/TDMA which is related with the GSM standard. For a TDMA based scheme, a way of achieving additional diversity by exploiting discontinuous transmission (such as voice activity) will be shown. The operation of diversity in general will be observed, and limits on the achievable diversity in general block-fading (or block-interference) channels will be discussed.

------------------------------------------------------------------------ Name: Sean Warnick Email: warnick@mit.edu Group: CSG, MIT Title: Piloting Epitaxy Through Ellipsometric Feedback Statement: Epitaxial processes are methods of systematically depositing specified materials on a substrate. These processes are critical in the production of various semiconductor devices, especially compound semiconductor devices such as vertical cavity surface emitting lasers (VCSELs) and other optical or optoelectronic devices. The primary challenge for epitaxy is to realize a desired structure. That is, we want to grow films such that the index of refraction, or material composition, of the resulting device satisfies a desired function of thickness. Recent advances in in-situ sensor technology have made feedback available as a tool to address this problem. This study investigates the issues of designing controllers for gas-source molecular beam epitaxy using spectroscopic ellipsometry as a sensor technology.

------------------------------------------------------------------------ Name: Richard Barron Email: Group: DSPG, MIT Title: Abstract:

POSTERS Name: Soosan Beheshti Title: System identification Email: soosan@mit.edu Group: CSG, MIT Statement:

----------------------------------------------------------------------- Name: Maurice Chu Title: Learning and Vision Email: Group: Statement:

-------------------- Name: Sae-Young Chung Title: Coding Theory Email: Group: Statement:

---------------------- Name: Angelia Geary Title: Nonlinear Optimization Email: Group: Statement:

------------------------- Name: John Richards Email: Group: Title: Target Model Generation for Multiple SAR Images Statement:

---------------------------------- Name: Tengo Saengudomlert Email: Group: Title: Networks Statement:

-------------------------------- Name: Anant Sahai Email: Group: Title: Information Theory and Control Statement:

--------------------------