November 13
Petros Boufounos
Mitsubishi Electric Research Laboratories

Recent Advances in Signal Acquisition, Sensing, and Quantization

Abstract:

The increasing availability of computing power, thanks to the advances of Moore's law, has put significant pressure on sensing technology to follow suit. Although sensor hardware cannot always keep pace, recent theoretical developments such as Compressive Sensing and Computational Imaging have demonstrated how smart sensor design can exploit cheap computation to improve sensing and signal acquisition technology. The hallmarks in these theoretical developments are randomization, non-linear reconstruction, and emphasis on signal models.

In this talk we emphasize how a model of the acquisition system can be taken into account in the design of the reconstruction algorithms. Specifically, we examine how the randomization of the measurements interacts with measurement quantization in analog-to-digital conversion. We first consider finite-range quantizers and demonstrate that, counter to common intuition, we can often decrease the error due to quantization by increasing the saturation rate of the quantizer. Then we consider the extreme case of 1-bit quantization and we demonstrate that we can significantly improve performance by explicitly incorporating the appropriate quantization model in the reconstruction. Finally, we consider signals measured though non-linear distortions, and we demonstrate that we can still reconstruct the signal from the measurements, even if the distortion itself is not known.

Bio:

Petros Boufounos completed his undergraduate and graduate studies at MIT. He received the S.B. degree in Economics in 2000, the S.B. and M.Eng. degrees in Electrical Engineering and Computer Science (EECS) in 2002, and the Sc.D. degree in EECS in 2006. Since January 2009 he is with Mitsubishi Electric Research Laboratories (MERL) in Cambridge, MA.

Between September 2006 and December 2008, Dr. Boufounos was with the Digital Signal Processing Group at Rice University doing research in the area of Compressive Sensing. Before that he was a postdoctoral associate in the MIT Digital Signal Processing Group. In addition to Compressive Sensing, his immediate research interests include signal processing, data representations, frame theory, and machine learning applied to signal processing. He is also looking into how compressed sensing interacts with other fields that use sensing extensively, such as robotics and mechatronics. Dr. Boufounos has received the Ernst A. Guillemin Master Thesis Award for his work on DNA sequencing and the Harold E. Hazen Award for Teaching Excellence, both from the MIT EECS department. He has also been an MIT Presidential Fellow. Dr. Boufounos is a member of the IEEE, Sigma Xi, Eta Kappa Nu, and Phi Beta Kappa.

October 30
C. Lee Giles
Pennsylvania State University

SeerSuite: Enterprise Search and Cyberinfrastructure for Science and Academia

Abstract:

Cyberinfrastructure or e-science has become crucial in many areas of science as data access often defines scientific progress. Open source systems have greatly facilitated design and implementation and supporting cyberinfrastructure. However, there exists no open source integrated system for building an integrated search engine and digital library that focuses on all phases of information and knowledge extraction, such as citation extraction, automated indexing and ranking, chemical formulae search, table indexing, etc. We propose the open source SeerSuite architecture which is a modular, extensible system built on successful OS projects such as Lucene/Solr and discuss its uses in building enterprise search and cyberinfrastructure for the sciences and academia. We highlight application domains with examples from computer science, CiteSeerX, chemistry, ChemXSeer, and archaeology, ArchSeer.

CiteSeerX, the successor to CiteSeer, currently offers or intends to offer some unique aspects of search not yet present in other scientific search services or engines, such as table, figure, algorithm and author search. In addition, CiteSeerX continuously crawls the web and author submissions and now has nearly 1.5 million documents, close to 30 million citations, a million authors and comparable database tables. It has nearly 1 million unique users with several million hits a day.

In chemistry, the growth of data has been explosive and timely, and effective information and data access is critical. The ChemXSeer (funded by NSF Chemistry) system is a portal and search engine for academic researchers in environmental chemistry, which integrates the scientific literature with experimental, analytical and simulation datasets. ChemXSeer consists of information crawled from the web, manual submission of scientific documents and user submitted datasets, as well as scientific documents and metadata provided by major publishers. Information gathered from the web is publicly accessible whereas access to restricted resources such as user submitted data will be determined by those users. Thus, instead of being a fully open search engine and repository, ChemXSeer will be a hybrid one, limiting access to some resources.

Because such enterprise systems require unique information extraction approaches, several different machine learning methods, such as conditional random fields, support vector machines, mutual information based feature selection, sequence mining, etc. are critical for performance. We draw lessons for other e-science and cyberinfrastructure systems in terms of design, implementation and research and discuss future directions and systems.

References:

[ChemXSeer] http://chemxseer.ist.psu.edu

[CiteSeerx] http://citeseerx.ist.psu.edu

[SeerSuite] http://en.wikipedia.org/wiki/SeerSuite

[Solr] http://lucene.apache.org/solr/

Bio:

Dr. C. Lee Giles is the David Reese Professor at the College of Information Sciences and Technology at the Pennsylvania State University, University Park, PA. He is also Professor of Computer Science and Engineering, Professor of Supply Chain and Information Systems, and Director of the Intelligent Systems Research Laboratory. He directs the CiteSeer^X <http://citeseerx.ist.psu.edu> project and codirects the Chem_X Seer <http://chemxseer.ist.psu.edu> project at Penn State. He has been associated with Columbia University, the University of Maryland, University of Pennsylvania, Princeton University, the University of Pisa and the University of Trento.

October 23
Alex Waibel
LTI

From Research Lab to Jungle Ops: Computer Speech Translation for Humanitarian Relief

Abstract:

Our world continues to be rocked by natural disasters and political calamities at unpredictable times and places, generating humanitarian needs that can often not be met locally alone. As the world responds, however, relief organizations and volunteers face communication challenges imposed by language barriers and cultural differences.

To satisfy the communication needs, we have worked for many years on building and deploying speech communication systems that would provide human-to-human language interpretation and have explored them in actual field use. In addition to the scientific problems associated with speech and language technology, such a goal harbors numerous practical, logistical and financial pitfalls that one stumbles upon when one leaves the comfort of our research labs.

In this talk I will tell the story of this (ongoing) adventure in three parts:

1.) Technology: How does a speech-to-speech translation system work and what does it do? What levels of performance can we expect, and what language assistance can it offer? Can it handle many languages and how flexible is it to be used in different field situations? What interfaces and what platforms are needed for field use? How can it be built, maintained and ported inexpensively?

2.) Money: How much does it cost to develop a speech translator? Who pays for the development of such systems? I will discuss our Robin Hood approach: we sign up people in the developing world to provide language expertise over the internet; we build, improve and market the technology in the developed world; with the proceeds we improve and adapt the technology for health care missions and redeploy it in the developing world. We have formed several companies around the world that provide system development, product distribution, marketing, and data collection.

3.) Deployment: We examine, how it all fits together and what lessons we have learned from three different healthcare missions that we are attempting to support: in the mountains of Honduras, remote villages of Thailand, and in the jungles of Papua, Indonesia.

Warning: No equations, but plenty of system demos, pictures, movies and 'war' stories from the field.

Bio:

Alex Waibel is a Professor of Computer Science at the Language Technology Institute at Carnegie Mellon University, and at the Institute of Anthropomatics at the University of Karlsruhe. He directs the international Center for Advanced Communication Technologies (InterACT) with research interests in multimodal and multilingual human communication systems. Dr. Waibel's team pioneered many of the first domain-limited and the domain­unlimited speech translators. He was one of the founders and chairmen (1998-2000) of C-STAR, the consortium for speech translation research. He has published extensively in the field, received several patents and awards, and built several successful companies. He received his BS, MS and PhD degrees at MIT and CMU, respectively.

October 22
Tuomas Virtanen
Tampere University of Technology

New Machine Learning Algorithms Enable Recognizing Sounds in Mixture

Abstract:

Automatic recognition of sounds in the presence of interfering sounds has been an unsolved problem for a long time, which has caused severe problems in applications such as automatic speech recognition (ASR). Additive noise degrades significantly the performance of conventional pattern recognition algorithms, for example the hidden Markov models used in ASR. New machine learning algorithms based on factorizing the sound spectrogram have proven to be good tools in separating sounds from mixtures. Specifically, non-negative spectrogram decompositions are able to learn the structure of the data automatically without the need to tune sophisticated separation algorithms. We present source separation algorithms and models based on non-negative matrix factorization and maxrix deconvolution, and extensions such as the source-filter model. We discuss both supervised and unsupervised use of the methods. The ability of the models to deal with mixtures of sounds has provided superior results in the recognition of noisy signals. We present state-of-the art results in robust speech recognition, speech enhancement, and music information retrieval. Audio demonstrations will be given.

Bio:

Tuomas Virtanen is a senior researcher at Tampere University of Technology, Finland. He finished his PhD about monaural sound source separation in 2006. His research interests include sound source separation, noise-robust automatic speech recognition, audio content analysis, and music information retrieval.