Engineering Seminar
Abstract: Motivated by the popularity of multimedia content and applications, researchers have been very active in developing new techniques and standards for audio-visual content description. Such descriptions can be used to facilitate innovative ways of accessing, filtering, and delivering multimedia information, potentially from distributed on-line sources.
We will present our research and prototype results in this area.
First, we will present our research in searching images/videos from large archives. We will describe our works in using multimedia features in content categorization, knowledge representation networks for integrating perceptual-level and conceptual-level knowledges, and iconic templates for searching visual-oriented concepts.
Second, we will present our works in digital video structuring and summarization. We will discuss approaches to scene segmentation and skimming for generic domains using a unique computational scene model and joint audio-visual features. We will present techniques for exploring syntactic structures and production rules in videos in specific domains, such as sports and medical.
Lastly, we will briefly discuss how the content description and structuring results from above can be used to enhance streaming and transmission of digital video. One application called ``content-aware resource allocation'' uses content analysis techniques in classifying and predicting the rate-distortion utility functions of video streams, which in turn facilitate optimization of resource allocation among multiple streams.
Speaker Bio: Shih-Fu Chang is an Associate Professor of Electrical Engineering at Columbia University. He currently leads Columbia's ADVENT university- industry consortium. The ADVENT group conducts research in representation, searching, transmission, and security of multimedia information. Several innovative systems have been developed and prototyped by his group, including VideoQ, WebSEEk, MetaSEEk for image/video searching, WebClip for networked video editing, and Sari for online image authentication.
He has been a PI or co-PI in several large-scale cross-disciplinary projects, including Columbia's Health Care Digital Library Project supported by the multi-agency DLI-2 initiative, the Digital News System, and a K-12 Art Image Education Project. His group has played an active role in developing MPEG-7. He has been a general co-chair of ACM 8th Multimedia Conference 2000, and associate editor for several journals, and a consultant in several new media companies.
Professor Chang has been awarded a Navy ONR Young Investigator Award, a Faculty Development Award from IBM, a CAREER Award from NSF, and authored or co-authored three best paper awards in the areas of multimedia manipultation and accessing. He is currently a Distinguished Lecturer in IEEE Society of Circuits and Systems.
Brown Applied Mathematics Pattern Theory and Vision Seminar
New York, NY | |
Abstract: I will describe an approach to the analysis of single-and multi- neuron spike trains based on a metric-space formalism, and contrast this with traditional vector-space approaches.
The approach will be applied to recordings from the primate visual cortex. The analysis provides strong evidence that spike trains should be considered as more than mere estimators of average firing rates or population activity, and also suggests how the temporal structure of spike trains might be used to represent and decode sensory information.
Brown Analysis Seminar
Scientific Computing Seminar
University of Texas at Austin | |
Abstract: I will start by reviewing shortly the idea and importance of the de Rham diagram in finite element approximations of time-harmonic Maxwell equations. First of all, convergence of Maxwell eigenvalues is sufficient and necessary for the (asymptotic) stability of Galerkin discretizations. Secondly, it has been shown that Kikuchi's discrete compactness property is not only sufficient but (in a certain sense) necessary for the convergence of the eigenvalues. And finally, commutativity of the diagram is sufficient and, again, necessary in some sense for proving the discrete compactness property.
In order to prove the discrete compactness property though, we need error estimates for the interpolation operators entering the diagram, with minimum regularity assumptions on interpolated functions. Whereas such estimates for the h-method have been widely known for a long time, the only known estimates for the p-method, obtained by Peter Monk, are valid for hexahedral (and square) edge elements, and they are suboptimal.
I will present first such optimal p-interpolation error estimates for the triangular edge elements of variable order. The proofs are based on two crucial technical results: a polynomial extension theorem for H^{1/2}_{00} spaces by Babuska et al, and a discrete Friedrichs inequality for polynomial spaces.
The interpolation result paves a way for the theory of optimal hp- discretizations for Maxwell's equations in 2D.
Joint work with Ivo Babuska.
PDE Seminar
Abstract: The classical theory of domain coarsening during phase transitions in materials science involves a conservative law for the size distribution of a family of particles. I will describe recent results concerning:
(a) The initial value problem: We give a theory that yields existence, uniqueness and continuous dependence on initial data for arbitrary measure-valued size distributions of compact support. We use a physically natural topology given by a Wasserstein distance between size distributions.
(b) Long-time behavior: Universal self-similar behavior was predicted classically based on physical stability arguments. Our rigorous analysis shows, however, that the classical model does not yield the predicted behavior. Instead, long-time behavior depends sensitively on the initial distribution of the largest particles. E.g., for a dense set of initial data, convergence to any self-similar solution is impossible.
This is a joint work with Barbara Niethammer, University of Bonn.
<--- 2001 Index