| |
The first Wednesday of every month, the Adaptive
Systems Laboratory has a "brown bag" seminar at
noon in the Columbia Conference Room (BCB).
This is the archive of past brown bag presentations.
| |
| 04/26/2006 | Finite Element Modeling of Hemodynamic Forces during Heart Developmen presented by Sandra Rugonyi from OGI/OHSU.
Abstract: Mechanical forces play a key role in heart development. Deviations from normal flow conditions in the developing heart alter the shear stresses and pressures exerted by the flow of blood on the walls of the heart, leading to congenital heart disease. The actual mechanisms by which flow-induced mechanical forces alter the development of the heart, however, are still not well understood, in part because measuring flows and flow-induced forces in vivo is challenging. Our goal is to quantify flow-induced forces in the heart during normal and abnormal conditions and develop a model to predict the effect of these alterations on development. To this end, we study the heart of chick embryos that have been incubated for approximately 3.5 days of a 21-day incubation period (stage HH21). At this stage, the heart consists of a looped tube, which has a primitive atrium followed by a ventricle. We focus on models of the heart outflow tract (OFT), which connects the primitive ventricle with the chick arterial sac. To better characterize flow-induced forces in vivo, we use a combination of experimental data and numerical modeling of the heart. The movement of the heart wall during the cardiac cycle is captured via high-frequency ultrasound (and OCT), and pressure in the ventricle is measured using a servo-null system. These data are incorporated into a finite element model of the OFT, from which flow-induced mechanical forces can be quantified. A future step is to correlate flow-induced mechanical forces with gene expression patterns and the morphogenesis of the OFT wall under normal and abnormal flow conditions, to enhance our understanding of congenital heart disease.
Speaker Bio: Sandra Rugonyi joined the Department of Biomedical Engineering at OGI as Assistant Professor in 2005. She earned a Nuclear Engineering degree from Balseiro Institute, Argentina in 1995, and MS and PhD degrees in Mechnical Engineering at MIT in 1999 and 2001 respectively. >From 2001-2005 she was a Post-Doctoral Fellow in the Department of Medicine at OHSU working with Dr. Steve Hall.
|
| 03/15/2006 | Photoacoustic imaging of biological tissue presented by Ricky Wang from OGI/OHSU. [abstract & bio] |
| 01/25/2006 | Physics Based Signal Processing for Radar Applications presented by Lisa M. Zurk, Shari Matzner and Reto Toengi from PSU. [abstract & bio] |
| 11/2/2005 | Registration of Microscopic Iris Images Using Probabilistic Elastic Mesh presented by Xubo Song from OGI/OHSU. [abstract & bio] |
| 9/7/2005 | A Self-healing Sensor Network Architecture presented by Nirupama Bulusu from PSU. [abstract & bio] |
| 8/3/2005 | Tremor Signal Analysis for Movement Disorders presented by Sungham Kim from PSU. [abstract & bio] |
| 7/6/2005 | Identifying Informative Spatiotemporal Features for Brain Computer Interfaces presented by Deniz Erdogmus from OGI/OHSU. [abstract & bio] |
| 4/20/2005 | The \"Autonomous\" helicopter project; from boats to rotorcrafts and lessons learned from dealing with DARPA presented by Eric Wan from OGI/OHSU. [abstract & bio] |
| 1/19/2005 | Proximity Graphs for Clustering and Manifold Learning presented by Miguel Á. Carreira-Perpiñán from OGI/OHSU. [abstract & bio] |
| 1/7/2005 | An all-too-quick romp through the lovely landscape of Predictive Coding presented by Jonathan Shaw from University of Rochester. [abstract & bio] |
| 12/1/2004 | Intelligent Coordination in Heterogeneous Sensor Networks presented by Nirupama Bulusu from PSU. [abstract & bio] |
| 11/3/2004 | Lyapunov exponents: A practical bound on model likelihood presented by Andy Fraser from PSU. [abstract & bio] |
| 10/6/2004 | Towards Rational Stochastic Modeling -- Beyond Simplistic Truncation and Overzealous Refutations presented by Todd Leen from OGI/OHSU. [abstract & bio] |
| 4/29/2004 | Information Theoretic Signal Processing presented by Deniz Erdogmus from OGI/OHSU. [abstract & bio] |
| 4/12/2004 | Quasi-Static Object Discovery presented by Brandon C.S. Sanders from University of Rochester. [abstract & bio] |
| 4/9/2004 | Towards a Better Representational Scheme for Cognitive Vision presented by Harry Foundalis from Indiana University. [abstract & bio] |
| 1/16/2004 | Synchrony, Oscillations and Temporal Coding in the Vertebrate Retina presented by Garrett Kenyon from NSI. [abstract & bio] |
| 12/15/2003 | Learning Distance Functions for Image Retrieval presented by Tomer Hertz from . [abstract & bio] |
| 11/21/2003 | Learning to use IP Puzzles presented by Wu-chang Feng from OGI/OHSU. [abstract & bio] |
| 10/24/2003 | Computer-Based Classification Results of an Extensive Set of Dysmorphic Faces presented by Stefan Boehringer from University of Essen, Germany. [abstract & bio] |
| 8/7/2003 | Dynamic Updates of Medical Models: Examples from Research in Medical Informatics presented by Holly Jimison from . [abstract & bio] |
| 6/18/2003 | Biomedical Signal Processing at Portland State presented by James McNames from PSU. [abstract & bio] |
| 5/7/2003 | Training Conditional Random Fields via Gradient Boosting presented by Tom Dietterich from OSU. [abstract & bio] |
| 3/12/2003 | Computational Models of Prefrontal Behavior: What can they teach us about the brain? presented by Tamara Hayes from OGI/OHSU. [abstract & bio] |
| 3/7/2003 | Mining, Mapping, Modeling and Crawling the Web presented by Filippo Menczer from University of Iowa. [abstract & bio] |
| 2/26/2003 | The Attentive Brain presented by Terrence Sejnowski from OGI/OHSU. [abstract & bio] |
| 2/5/2003 | Analogy-Making as Perception presented by Melanie Mitchell from OGI/OHSU. [abstract & bio] |
| 1/21/2003 | Particle filters for state estimation in mobile robotics presented by Dieter Fox from University of Washington. [abstract & bio] |
| 1/8/2003 | Semiconductor Engineering Meets Neuroscience presented by Dan Hammerstrom from OGI/OHSU. [abstract & bio] |
| 12/4/2002 | Sensory Adaptation in Mormyrid Electric Fish presented by Patrick Roberts from NSI. [abstract & bio] |
04/26/2006 -- Finite Element Modeling of Hemodynamic Forces during Heart Developmen presented by Sandra Rugonyi from OGI/OHSU.
Abstract: Mechanical forces play a key role in heart development. Deviations from normal flow conditions in the developing heart alter the shear stresses and pressures exerted by the flow of blood on the walls of the heart, leading to congenital heart disease. The actual mechanisms by which flow-induced mechanical forces alter the development of the heart, however, are still not well understood, in part because measuring flows and flow-induced forces in vivo is challenging. Our goal is to quantify flow-induced forces in the heart during normal and abnormal conditions and develop a model to predict the effect of these alterations on development. To this end, we study the heart of chick embryos that have been incubated for approximately 3.5 days of a 21-day incubation period (stage HH21). At this stage, the heart consists of a looped tube, which has a primitive atrium followed by a ventricle. We focus on models of the heart outflow tract (OFT), which connects the primitive ventricle with the chick arterial sac. To better characterize flow-induced forces in vivo, we use a combination of experimental data and numerical modeling of the heart. The movement of the heart wall during the cardiac cycle is captured via high-frequency ultrasound (and OCT), and pressure in the ventricle is measured using a servo-null system. These data are incorporated into a finite element model of the OFT, from which flow-induced mechanical forces can be quantified. A future step is to correlate flow-induced mechanical forces with gene expression patterns and the morphogenesis of the OFT wall under normal and abnormal flow conditions, to enhance our understanding of congenital heart disease.
Speaker Bio: Sandra Rugonyi joined the Department of Biomedical Engineering at OGI as Assistant Professor in 2005. She earned a Nuclear Engineering degree from Balseiro Institute, Argentina in 1995, and MS and PhD degrees in Mechnical Engineering at MIT in 1999 and 2001 respectively. >From 2001-2005 she was a Post-Doctoral Fellow in the Department of Medicine at OHSU working with Dr. Steve Hall.
03/15/2006 -- Photoacoustic imaging of biological tissue presented by Ricky Wang from OGI/OHSU.
Abstract: When laser light shining into the biotissue, there are numerous phenomena that can be observed with which bio-physicists have successfully explored to sense different physically meaningful parameters for diagnostic and therapeutic purposes. I will be going to talk about acoustic wave (or pressure wave) emission from tissue when shining a pulsed laser light. By use of ultrasonic transducer (either single or array detector) to pick up the acoustic wave in vicinity of tissue, functional imaging is possible. The image so obtained has advantages over pure optical or ultrasound imaging alone because in photoacoustic image, its imaging contrast is relied on the optical property of tissue while the imaging resolution is determined by the frequency of sound wave detected. Examples of biological imaging using this technique will be given.
Speaker Bio: Ricky obtained his PhD degree in optical engineering in 1995 from Glasgow University, Scotland. After 2 years postdoctoral training at Glasgow, he took up a lecturer position in bio-imaging science at Keel University Medical School in 1997, and then promoted to a senior lecturer in 2000. In 2002, he moved to Cranfield University Institute of Bioscience and Technology where he took up his chair position in Biomedical Optics and directed biophotonics laboratories there. He then quite bravely moved across Atlantic Ocean to OGI here where he is currently employed as a professor in biomedical optics. He was an author for more than 80 pure refereed journal papers in biomedical optics area, and frequently being invited to serve as a special issue editor in Biomedical Optics for a number of international journals. He also sat on a number of research funding committees for UK research councils. His current research interests are focused on optical coherence tomography for biological applications, tissue optics, optical applications for tissue engineering, optical tissue clearing and photoacoustic imaging.
01/25/2006 -- Physics Based Signal Processing for Radar Applications presented by Lisa M. Zurk, Shari Matzner and Reto Toengi from PSU.
Abstract: The processing formulation for radar systems has been based upon an electromagnetic model that makes the simplifying assumption of uncorrelated, discrete, stationary point scatterers that are randomly distributed on a spherical earth surface. While this description has been adequate to drive the development of basic Synthetic Aperture Radar (SAR) image formulations, it does not fully exploit the highly structured signatures produced by scattering environments and embedded targets. The increase in sophistication of electromagnetic models - combined with the increase in computational capability and the advent of novel sensor design - suggests that an investigation into the synergies between electromagnetic scattering signatures and the radar signal processing framework could prove useful in developing physics-based exploitation algorithms that more effectively capture scene and target characteristics. In this work we develop and apply appropriately sophisticated electromagnetic models to capture the frequency dependent, spatially and temporally distributed electromagnetic backscatter from terrain and targets. We consider two specific radar applications: 1) correlation processing for the detection of sub-surface objects (i.e., land mines) and 2) automated scene discrimination for airborne radar imaging. In the first application, the primary goal is to develop techniques to discriminate the radar return from the buried object from the much stronger scattering from the rough terrain surface. To accomplish this, we developed a model describing the scattering from a semi-transparent rough boundary with homogeneous or inhomogeneous medium below it and targets embedded in this medium. We conducted the numerical simulation for a variable number of the targets and for typical situations describing the different soil conditions. A correlation-based processing scheme with a physically motivated weighting function was then developed to exploit multiple measurement geometries. By choosing the partial exponential terms to compensate for the electromagnetic wave decay in the medium we can isolate the signal from the desired depth where the object is located. In the second application, an end-to-end electromagnetic and radar simulation model has been developed to understand and evaluate the appearance of complicated scene features in spotlight SAR images as a function of scene parameters, sensor characteristics, and radar processing approaches. The scattering model has been developed to emulate the performance of the Lincoln Multi-Mission ISR Testbed (LiMIT), which is an airborne SAR sensor that was recently developed and deployed by MIT Lincoln Laboratory. Output of the model is compared and validated with LiMIT data from a recent collection campaign over San Clemente Island. Furthermore, the model is being employed to investigate the potential of advanced physics-based processing techniques that can be used to produce alternatives to or extensions of standard SAR image output for automated feature detection in urban environments.
Speaker Bio: Dr. Lisa M. Zurk received her Bachelor\'s degree in Computer Science at University of Massachusetts, Amherst in 1985, the Master\'s in Electrical and Computer Engineering at Northeastern University in 1991, and her PhD in Electrical Engineering at the University of Washington in 1995. She spent four years in industry in biomedical instrumentation and another nine years at MIT Lincoln Laboratory where she conducted research in understanding the physics of electromagnetic and acoustic wave propagation through modeling and measurement in order to devise advanced, physics-based signal processing techniques. For the 2000-2001 academic year, she was on sabbatical from MIT to teach and conduct research as a visiting Fulbright Professor in the University Helsinki Math Department. In January 2005 she joined the Electrical and Computer Engineering Department at Portland State University and subsequently founded the Northwest Electromagnetics and Acoustics Research (NEAR) Laboratory, of which she is the Director.
11/2/2005 -- Registration of Microscopic Iris Images Using Probabilistic Elastic Mesh presented by Xubo Song from OGI/OHSU.
Abstract: The recent development of video-microcropy technology for imaging immune responses within the eye is revolutionizing the way researchers are studying and understanding the immune mechanism. Such technology enables the visualization of T cell behavior in disease models without resorting to surgical trauma. The motion patterns of T cells are directly related to the cellular and chemical environment at the site of inflammation and thus can reveal the underlying disease mechanism. However, such technology needs the capability for computer-aided image processing and analysis due to images that are compromised by motion artifacts that obfuscate the true T cell motion. In addition, the current practice of manual tracking of the cell locations is a prohibitive task for massive amount of images. The goal of this project is to develop image-processing techniques for tracking and characterizing cell motion in microscopic video of the ocular uveal tract. This talk will focus on the development of techniques for image registration in order to stabilize images plagued with motion artifacts. We will discuss a method based on probabilistic elastic mesh that is able to handle local elastic geometric deformations, and demonstrate its effectiveness on a set of microscopic videos.
Speaker Bio: Xubo Song is currently an assistant professor in the Department of Computer Science and Electrical Engineering at OGI, OHSU. She also has a joint appointment with the Dept. of Biomedical Engineering. She has Masters and Phd degrees in Electrical Engineering, both from Caltech. Her research interests include machine learning, image processing and analysis, with a flavor for medical applications.
9/7/2005 -- A Self-healing Sensor Network Architecture presented by Nirupama Bulusu from PSU.
Abstract: For widespread adoption of sensor technology, robustness in the event of abnormal behavior such as a network intrusion, or failures of components or nodes is critical. Current research on robust and resilient sensor networking is focused on specific tasks - secure broadcast, secure aggregation, secure localization or fault-tolerant feature extraction. While these primitives provide useful functionality, what has been lacking is a comprehensive, holistic approach to sensor network robustness across various failure modalities. In this talk, I will describe a self-healing hybrid sensor network architecture called SASHA, that is inspired by and co-opts several mechanisms from the Acquired Natural Immune System to attain its autonomy, robustness, diversity and adaptability to unknown pathogens, and compactness. SASHA encompasses automatic fault recognition and response over a wide range of possible faults. Moreover, it is an adaptive architecture that can learn and evolve its monitoring and inference capabilities over time to deal with unknown faults. I will illustrate the workings of SASHA using the example of fault-tolerant distributed sensor data collection and outline an agenda for future research.
Speaker Bio: Nirupama Bulusu is an Assistant Professor of Computer Science at Portland State University. She received her Ph.D from UCLA in 2002.
8/3/2005 -- Tremor Signal Analysis for Movement Disorders presented by Sungham Kim from PSU.
Abstract: Tremor is one the most common disabling symptoms for the movement disorders such as Parkinson\'s disease (PD) and essential tremor (ET). Tremor activity can be measured with many types of instrumentation and sensors including EEG, MEG, EMG, and MER. We have developed processing methods to detect the presence of tremor, automatically detect action potentials, and track the instantaneous frequency of tremor in spike trains. We have also developed an index of phase coupling between two tremor signals. These methods help us understand the mechanism and pathophysiology of tremors.
Speaker Bio: Sunghan Kim is a member of the Biomedical Signal Processing Laboratory at Portland State University. He received his BS in Electrical and Computer Engineering and his MS degree in Electrical and Computer Engineering from Portland State University in 2003 and 2005 respectively. He is currently enrolled in the Ph.D. program in Electrical and Computer Engineering at PSU. His research interests are biomedical signal processing, digital signal processing, and applied regression analysis.
7/6/2005 -- Identifying Informative Spatiotemporal Features for Brain Computer Interfaces presented by Deniz Erdogmus from OGI/OHSU.
Abstract: Brain computer interface (BCI) refers to a family of engineered systems that modulate the interactions between human brains and computers. These systems typically involve active or passive participation of the human subject, were in the active mode, the human controls the computer or the mechanical system directly by thought, and in the passive mode the computer makes inferences regarding the mental state of the user to manipulate his environment. In either case, accurate estimation of the mental state from brain activity signals is the most crucial component to real-world BCI applications.Modern noninvasive BCI applications use information obtained from the user\'s electroencephalogram (EEG) to estimate the mental states. However, due to the noisy and nonstationary character of EEG signals, it is an extremely difficult problem to classify mental state directly from raw EEG signals. An accurate and robust real-time classification system requires a set of compact and stable features to be extracted from the raw EEG measurements.Using mutual information, we can select the features that contain the most relevant information regarding the classification solution, as well as eliminating the irrelevant, redundant, and noisy features, which will impair the performance. We have applied this technique to cognitive state estimation in the context of the Augmented Cognition Project. Experimental results demonstrate that this technique can successfully identify EEG sites and frequency bands of activity that best identify the cognitive state.
Speaker Bio: Deniz Erdogmus received his BS degrees in Electrical Engineering and Mathematics in 1997, and his MS in Electrical Engineering from the Middle East Technical University in Turkey. He received his PhD in Electrical Engineering from the University of Florida in 2002 after which he spent another two years as a post-doc working on invasive brain-machine interfaces. Deniz joined the OHSU in August 2004 as an Assistant Professor with a joint appointment at the CSEE and BME departments. His research interests include adaptive signal processing, machine learning, and information theory.
4/20/2005 -- The \"Autonomous\" helicopter project; from boats to rotorcrafts and lessons learned from dealing with DARPA presented by Eric Wan from OGI/OHSU.
Abstract: This talk looks back at the conclusion of the 5 year Software Enabled Control (SEC) DARPA project. In this candid presentation, I will discuss what we set out to do and what was really accomplished in the course of both technical hurdles and shifting project goals from above. Additional details are provided for some of the core algorithms developed, including Model Predictive Neural Control, State-Dependent Riccati Equation Control, and Sigma-Point Kalman filtering for Integrated Navigation. Future directions are discussed. We conclude with a video from the final DARPA demonstration, which took place last August at Ft. Benning, GA.
Speaker Bio: Eric Wan is an associate professor at the CSEE department of OGI/OHSU.
1/19/2005 -- Proximity Graphs for Clustering and Manifold Learning presented by Miguel Á. Carreira-Perpiñán from OGI/OHSU.
Abstract: Consider a cloud of points in Euclidean space. Machine learning problems such as clustering or dimensionality reduction, when formulated on the pairwise distances between points, assume a graph having as vertices the data points. The graph is then partitioned (clustering) or used to redefine metric information (dimensionality reduction). There has been much recent work on different algorithms for graph-based clustering and dimensionality reduction, but not much on learning the graph itself. Graphs typically used include the fully-connected graph, a fixed-grid graph (for image segmentation) or a nearest-neighbour graph.We suggest that the graph should adapt locally to the structure of the data and take into account its noisy nature. This can be attained by a graph ensemble obtained by combining multiple minimum spanning trees, each fit to a perturbed version of the data set. We show that such a graph ensemble usually produces a better representation of the data manifold than standard methods; and that it provides robustness to a subsequent clustering or dimensionality reduction algorithm based on the graph.Joint work with Rich Zemel (U. of Toronto).
Speaker Bio: Miguel Á. Carreira-Perpiñán is an assistant professor at the CSEE department of OGI/OHSU. He received a PhD in Computer Science from the University of Sheffield, UK in 2001, and did postdoctoral work at Georgetown University and the University of Toronto. His research interests lie in machine learning and computational neuroscience.
1/7/2005 -- An all-too-quick romp through the lovely landscape of Predictive Coding presented by Jonathan Shaw from University of Rochester.
Abstract: There has been much effort to discover guiding principles behind the brain\'s amazing perceptual abilities. One of the strongest contenders, and perhaps the only one that can fully address learning in high dimensional data sets without making strong prior assumptions (no offense to my more Bayesian colleagues), is predictive coding (but we\'ll still touch on temporal invariance). Predictive coding states that the chief job of perception is to filter out and discard everything about past input that is not predictive about future input. The talk will attempt to motivate this principle from very basic observations about perception. Next, it will move on to previous approaches that approximate predictive coding, and discuss why the standard approximation is hopelessly poor. Finally, it will explore two techniques that actually can perform predictive coding (one has not previously been recognized as such, and the other is novel). The talk will end with a pretty result that has little to do with anything. So if you are at all interested in perception, this talk is for you!
Speaker Bio: onathan Shaw is a doctoral student at the University of Rochester. He is interested in biologically plausible computer models of perception, with an emphasis on predictive coding in vision.
12/1/2004 -- Intelligent Coordination in Heterogeneous Sensor Networks presented by Nirupama Bulusu from PSU.
Abstract: The past few years have witnessed deployment of networks of low-powered homogeneous sensing devices in a variety of applications, ranging from environmental monitoring to health care. Heterogeneous sensor technologies can enormously enrich the range of potential sensing applications. Heterogeneity spans many dimensions such as sensing modalities, computation, communication, energy and mobility resources.In this talk, I will describe our experiences in building intelligent coordination mechanisms for heterogeneous sensor networks that systematically exploit and adapt to sensor diversity, for(i) mobility-assisted localization of sensor devices,(ii) adaptive resource management, and(iii) cane-toad monitoring using a hybrid sensor network.I will conclude my talk with specific examples and a discussion of the role machine learning can play in making these systems truly self-configuring.This talk represents joint work with Wen Hu, Pubudu Pathirana, Sanjay Jha, Andrew Taylor and Wu-chi Feng.
Speaker Bio: Nirupama Bulusu is an Assistant Professor of Computer Science at Portland State University. Her research interests lie in wireless sensor networks. She received her Ph.D in Computer Science from UCLA in 2002, working with Prof. Deborah Estrin and John Heidemann. For more information, see: http://www.cs.pdx.edu/~nbulusu
11/3/2004 -- Lyapunov exponents: A practical bound on model likelihood presented by Andy Fraser from PSU.
Abstract: For decades (since 1988, so 1.6 decades), I\'ve been saying that you can test model fitting techniques on chaotic data and compare the log likelihood that you get against chi, the sum of the positive Lyapunov exponents. If your log likelihood is less than -chi, then your technique is sub-optimal. When I\'ve said this in public, real mathematicians ask questions like, \"How do you know how good your Lyapunov exponent estimates are? Do you really believe that Lyapunov exponents exist for this system? What about the brittleness of the system: Does shadowing work here? But Pesin\'s relation only holds with equality if the measure is SRB. Is your measure SRB?\". In reply, I have mumbled, feeling like W faced with Kerry. But now, having spent a couple of weeks thinking about it, I\'m pretty confident that those questions are irrelevant. Running the Benettin procedure for estimating Lyapunov exponents on a simulation of a stochastic process does (within certain bounds on the noise scale, exponents, and measurement resolution) give a lower bound on the achievable log likelihood for models.
Speaker Bio:
10/6/2004 -- Towards Rational Stochastic Modeling -- Beyond Simplistic Truncation and Overzealous Refutations presented by Todd Leen from OGI/OHSU.
Abstract: The dynamics of Markov processes is a recurring subject in chemistry and physics, and pops up in machine learning in the theory of stochastic approximation algorithms, and in state estimation (Kalman filters and their nonlinear extensions). More recently the subject has captured the imagination of theoretical neuroscientists interested in the statistical properties of synaptic plasticity, particularly in systems exhibiting spike-timing-dependent plasticity.Unfortunately, the theoretical approach almost universally adopted to \"approximate\" these learning systems -- the nonlinear Fokker-Planck equation (FPE) -- fails miserably (in these applications) to satisfy a very simple requirement of any approximation scheme. The technique is provably incorrect. However it appears to give quite good results for some problems!The landscape is made even more confusing by theoretical expositions in both the chemistry and machine learning literature that grossly overstate the case against the FPE, and totally disregard the empirical demonstrations of its utility.We\'ll briefly discuss a valid approximation scheme, it\'s application to machine and biological learning problems, and propose its use to (hopefully) clear up a very confused literature.
Speaker Bio:
4/29/2004 -- Information Theoretic Signal Processing presented by Deniz Erdogmus from OGI/OHSU.
Abstract: Signal processing aims at extracting useful information from the available data. Traditionally, under the assumptions of linear filtering and Gaussian signal distributions, second order statistics have been heavily exploited for signal processing in an adaptive filtering framework. However, in the real world, signals are not Gaussian and systems are nonlinear. Consequently, linear correlation based filtering is not optimal. We propose information theoretic learning (ITL) as a unified framework for optimal signal processing. To this end, a nonparametric estimator for information theoretic quantities suitable for adaptation is developed. Stochastic gradient adaptation algorithms for on-line adaptation are derived and a global optimization methodology is proposed. Applications of the proposed ITL framework to various problems in system identification, digital communications, pattern analysis, and speech and biomedical signal processing will also be discussed to clearly demonstrate the advantages of the proposed information theoretic signal processing framework.
Speaker Bio: Deniz Erdogmus received his B.S. degrees in electrical engineering and mathematics in 1997, and his M.S. degree in electrical engineering, with emphasis on systems and control, in 1999, all from the Middle East Technical University, Ankara, Turkey. He received his Ph.D. degree in electrical and computer engineering from the University of Florida, Gainesville, in 2002. Deniz\'s research interests lie in the broad areas of adaptive systems, machine learning, and signal processing, including applications of techniques stemming from these theories to solve problems in communications, biomedical engineering, and control. In particular, he investigates the information theoretical aspects of adaptation and statistical learning.
4/12/2004 -- Quasi-Static Object Discovery presented by Brandon C.S. Sanders from University of Rochester.
Abstract: Humans can scarcely think a thought or speak a sentence that is not fundamentally about objects. Our mental reliance upon objects contrasts sharply with the nearly complete inability of our machines to discern them. \"Object Discovery\" (OD) addresses this imbalance by seeking to group all observations springing from a single object without including any observations generated by other objects. In the field of computer vision, observations of objects are provided by cameras. Each observation is composed of a measured value (red,green,blue), a spatial index (c,x,y) that tells the camera, column, and row of the sensing element (sensel) that made the measurement, and a time-stamp (t) indicating when the measurement was acquired. Largely due to historic limits on sensing and computational resources, most Object Discovery approaches rely primarily upon the spatial index.In this talk I will show how under the \"Quasi-Static\" world model, using time rather than space to relate observations allows reliable, unsupervised discovery of objects across multiple completely uncalibrated cameras. Our Quasi-Static Labeling Theorem proves that, under realistic observability conditions, we recover the maximally informative mapping from observations to the objects they stem from. We have embodied this theory in an offline, deterministic implementation that (1) ignores distracting motion, (2) correctly deals with complicated occlusions, (3) discovers objects it has never before encountered, and (4) correctly associates all the views of a single object across multiple completely uncalibrated cameras. At the end of the talk I will discuss our current work extending the original monolithic, offline, deterministic implementation to a distributed, online, probabilistic system. In the probabilistic formulation we employ sampling to obtain robustness, the Minimum Description Length (MDL) principle to evaluate priors, and a hierarchical pairwise Markov Random Field to combine local and global evidence and hypotheses.
Speaker Bio: Brandon Sanders is interested in machine learning, with special emphasis on perceptually grounding representations of object characteristics, relationships, and interactions. He received a BSE in EE from Walla Walla College in 1998, a Masters in CS from the University of Rochester in 1999, and is expecting to defend his Ph.D. dissertation \"Quasi-Static Object Discovery\" in June of 2004, also at the University of Rochester. Miscellaneous topics he has spent time studying include biometric identity validation, 3D person tracking to monitor physical therapy compliance, speech reading to augment aural recognition, microphone arrays for speaker localization, and single spike models of neural computation.
4/9/2004 -- Towards a Better Representational Scheme for Cognitive Vision presented by Harry Foundalis from Indiana University.
Abstract: I will describe Phaeaco, a novel cognitive architecture that specializes in the domain of cognitive vision. The sub-domain of Bongard problems, a set of 100 puzzles of visual pattern understanding, constitutes Phaeaco\'s current focus of expertise. The proposed architecture cannot be characterized as belonging squarely to either the connectionist or the symbolic paradigm; nor is it a hybrid system. Rather, it attempts to make use of the best ideas from both worlds.Phaeaco can be seen as a representational scheme for cognitive vision, which includes a way of forming visual patterns in short-term memory out of raw visual input; storing such patterns in long-term memory, retrieving, and enriching them by experience; a pattern-matching mechanism; a chunking mechanism for the formation of groups of visual objects; and a categorization mechanism that ultimately leads to the solution of Bongard problems. Other features of Phaeaco include using a fault-tolerant approach that, through multi-perception, allows the improvement of possibly initially incorrect representations; combining the prototype and exemplar models for conceptual representation; employing the so-called Generalized Context Model to perform pattern matching and compute a similarity measure; and adhering to a version of the minimum description length principle for the formation of sensible (psychologically plausible) visual patterns.
Speaker Bio: Harry Foundalis is a Ph.D. student in Douglas Hofstadter\'s group at Indiana University, studying visual pattern recognition.
1/16/2004 -- Synchrony, Oscillations and Temporal Coding in the Vertebrate Retina presented by Garrett Kenyon from NSI.
Abstract: The issue of how visual neurons encode information remains a subject of considerable debate. The vertebrate retina is an ideal preparation for investigating such questions, as the anatomy and physiology of its principle cell types are comparatively well understood and its inputs and outputs can be clearly identified. Moreover, the vertebrate retina exhibits a form of temporal coding characterized by high-frequency oscillations that are phase locked between neurons responding to the same visual feature. We have conducted computer modeling studies and have reanalyzed electrophysiological data in an effort to understand the cellular and synaptic mechanisms underlying stimulus-specific high-frequency oscillations in the vertebrate retina. Our results suggest a novel hypothesis as to what additional types of visual information are encoded by coherent high-frequency oscillations between retinal neurons.For directions to NSI, see http://www.ohsu.edu/nsi/contact/directions/IMPORTANT: There is a guard gate at the entrance to West Campus, where you will be asked to show your OHSU badge, so remember to take it with you. If you don\'t have one, please e-mail rosatoj@ohsu.edu to be put on a list for the guard. You will still be asked to show some form of ID. If you have a problem, call Jane Rosato at NSI: (503) 418-2500.
Speaker Bio:
12/15/2003 -- Learning Distance Functions for Image Retrieval presented by Tomer Hertz from .
Abstract: Image retrieval critically relies on the distance function used to compare a query image to images in the database. We suggest to learn such distance functions by training binary classifiers with margins, where the classifiers are de- fined over the product space of pairs of images. The classifiers are trained to distinguish between pairs in which the images are from the same class and pairs which contain images from different classes. The signed margin is used as a distance function. We explore several variants of this idea, based on using SVM and Boosting algorithms as product space classifiers. Our main contribution is a distance learning method which combines boosting hypotheses over the product space with a weak learner based on partitioning the original feature space. The weak learner used is a Gaussian mixture model computed using a constrained EM algorithm, where the constraints are equivalence constraints on pairs of data points. This approach allows us to incorporate unlabeled data into the training process. Using some benchmark databases from the UCI repository, we show that our margin based methods significantly outperform existing metric learning methods, which are based on learning a Mahalanobis distance. We then show comparative results of image retrieval in a distributed learning paradigm, using two databases: a large database of facial images (YaleB), and a database of natural images taken from a commercial CD. In both cases our GMMbased boosting method outperforms all other methods, and its generalization to unseen classes is superior.Joint work with Aharon Bar-Hillel and Prof Daphna Weinshall.
Speaker Bio:
11/21/2003 -- Learning to use IP Puzzles presented by Wu-chang Feng from OGI/OHSU.
Abstract: Over the years, distributed denial-of-service (DoS) attacks via worms and viruses have periodically disrupted the Internet. Client puzzles have been proposed as one mechanism for protecting protocols against such attacks. In the first part of this talk, we argue that puzzles must be placed within the IP layer in order to be effective. We then describe the design, implementation, and evaluation of an initial prototype system. While IP puzzles provide us with an effective weapon to combat DoS attacks, they clearly require a dynamic, intelligent, and automatic system that controls their use. In the second part of this talk, we will open the floor for a discussion on the appropriateness of using current machine learning techniques for this problem in the hopes of stimulating a collaborative research effort in this area.
Speaker Bio:
10/24/2003 -- Computer-Based Classification Results of an Extensive Set of Dysmorphic Faces presented by Stefan Boehringer from University of Essen, Germany.
Abstract: Our project explores the feasibility of computer aided diagnostics of certain genetic defects based on statistical classification of face features. An extensive set of digital images of faces from patients with certain syndromes was collected. These syndromes include Mucopolysaccharidosis type III, Cornelia de Lange, Fragile X, Prader-Willi, Williams-Beuren, 5p-, 22q-, Noonan, Sotos, and Smith-Lemli-Opitz syndrome. The analysis is performed in two steps: first obtaining a numeric decomposition of a facial image using Gabor-Wavelet transformations and second application of statistical classification methods. We show results for kth-nearest neighbour (kNN), support vector machines (SVM) and linear discriminant analysis (LDA). Prior to performing SVM and LDA we employed a principle component analysis to reduce the complexity of the data set. This step resulted in 30 components out of 3840 covariates in the original data set. The error rates as measured by cross validation are 37%, 49% and 31% for kNN, SVM and LDA, respectively. These results indicate that the methods presented can be helpful in guiding the diagnostic process. Reconstructions of average or typical images of faces for a given syndrome allow us to understand the decision process of a classification algorithm.
Speaker Bio: Stefan Boehringer received his Ph.D. in Molecular Human Genetics and is currently a research associate in the department of Human Genetics at the University of Essen in Germany.
8/7/2003 -- Dynamic Updates of Medical Models: Examples from Research in Medical Informatics presented by Holly Jimison from .
Abstract: Although most medical research and clinical studies provide us with population statistics leading to broad-based clinical guidelines, it is often important to incorporate patient specific information when making treatment decisions for individuals. In this talk I will present a decision analytic framework for tailoring population treatment decision models to individual patients using Bayesian networks. This framework provides a mechanism for prioritizing the assessment of information from patients, for displaying the relative importance and sensitivity of variables in the model, and for generating patient-specific educational materials. I will also provide an overview of examples of OHSU research projects in medical informatics where dynamic user models and adaptive algorithms are used in optimizing user interfaces and in detecting trends in patient state.
Speaker Bio:
6/18/2003 -- Biomedical Signal Processing at Portland State presented by James McNames from PSU.
Abstract: This talk will give an overview of the activities of the Biomedical Signal Processing Laboratory and some of our recent research projects in detail including microelectrode recording visualization and adaptive comb filters for electrocardiograms. The mission of the BSP lab is to advance the art and science of extracting clinically significant information from physiologic signals. The objectives of this program are to develop new methods of signal processing that extract useful information from physiologic signals, to provide students with a solid foundation in statistical data analysis and signal processing, to teach undergraduate and graduate students about the process of knowledge discovery and research, and to serve the needs of the Portland metropolitan area. We primarily focus on clinical projects in which the extracted information can help physicians make better critical decisions and improve patient outcome.
Speaker Bio: James McNames is an assistant professor in the Department of Electrical and Computer Engineering and Director of the Biomedical Signal Processing (BSP) Laboratory at Portland State University. He graduated with his Ph.D. from Stanford in 1999. His research interests include statistical signal processing, statistical learning theory, system identification, nonlinear modeling, pattern recognition, time series analysis, and statistical process control.
5/7/2003 -- Training Conditional Random Fields via Gradient Boosting presented by Tom Dietterich from OSU.
Abstract: Recently, Lafferty, McCallum and Pereira introduced the Conditional Random Field as a new model for solving sequential supervised learning problems. Many applications of machine learning can be formalized as Sequential Supervised Learning (SSL). Each training example in SSL has the form (X,Y), where X is a sequence (x1, ..., xT) of items (each usually described by a vector in R^n), and Y is a sequence (y1, ..., yT) of class labels from {1, ..., K}. Typical applications include part-of-speech tagging, information extraction from web pages, and text-to-speech mapping. There are no robust, off-the-shelf methods for solving SSL problems in any commercial or academic statistical or data mining software systems. This talk will provide an introduction to the problem, discuss why the CRF is a good candidate for an off-the-shelf method, and describe our work on applying Friedman\'s Gradient Tree Boosting algorithm to efficiently and flexibly fit CRF models to the large NETtalk text-to-speech data set. Preliminary performance results will be presented.
Speaker Bio: Dr. Dietterich (AB Oberlin College 1977; M.S. University of Illinois 1979; Ph.D. Stanford University 1984) joined the OSU faculty in January 1985. In 1987, he was named a Presidential Young Investigator for the NSF. In 1990, he published, with Dr. Jude Shavlik, the book entitled Readings in Machine Learning, and he also served as the Technical Program Co-Chair of the National Conference on Artificial Intelligence (AAAI-90). From 1992-1998 he held the position of Executive Editor of the journal, Machine Learning. The American Association for Artificial Intelligence named him a Fellow in 1994, and the Association for Computing Machinery did the same in 2003. In 2000, he co-founded a new, free electronic journal: The Journal of Machine Learning Research. He served as Technical Program Chair of the Neural Information Processing Systems (NIPS) conference in 2000 and General Chair in 2001. He is currently President of the International Machine Learning Society and he also serves on the Board of Trustees of the NIPS Foundation.
3/12/2003 -- Computational Models of Prefrontal Behavior: What can they teach us about the brain? presented by Tamara Hayes from OGI/OHSU.
Abstract: In recent years, an number of reseachers have begun to use computational approaches such as Neural Networks to model the complex behaviors mediated by prefrontal cortex. These models are then \"lesioned\" in various ways in an attempt to replicate the changes in behavior that occur due to particular disease states. While it is clear that computational models can be created to perform most any task, what\'s not obvious is that this approach contributes to our understanding of cortical function or of the disease states. In particular, because our understanding of the underlying neuroanatomy and behavior is so poorly understood, creating a model that is sufficiently detailed to generate testable hypothesis is a challenging problem. In this seminar I will review recent work in this area and engage the debate about the value of modelling ill-defined neuronal systems.
Speaker Bio: Tamara Hayes is Assistant Professor in the Department of Biomedical Engineering and in the Department of Computer Science and Engineering at OGI. Before joining the OGI School of Science and Engineering, Hayes was a senior database engineer and director of the distributed systems management group at Informix Software and IBM. She also worked as a senior research associate on a telemedicine research project at OHSU from 1994 to 1996. Hayes received a bachelor\'s degree in engineering science and a master\'s in electrical engineering from the University of Toronto. Her doctoral degree in neuroscience is from the University of Pittsburgh. Hayes is interested in the interface between technology and human language function. She will explore ways to use technology to assist, improve or rehabilitate function in people with cognitive impairments, and will research computer systems and devices that emulate cognitive behavior.
3/7/2003 -- Mining, Mapping, Modeling and Crawling the Web presented by Filippo Menczer from University of Iowa.
Abstract: Can we model the scale-free distribution of Web links under realistic assumptions about the behavior of page authors? Can a Web crawler efficiently locate an unknown relevant page? These questions are receiving much attention due to their potential impact for understanding the structure of the Web and for building better search engines. This talk will discuss the semantic maps obtained by analyzing the connection between similarity functions based on text, link and semantic cues across a massive number of page pairs. These maps uncover some striking relationships. For example link probability displays a phase transition between a region where it is not determined by content and one where it decays with textual distance according to a power law. This relationship suggests a novel Web growth model that is shown to accurately predict the distribution of page degree, based on textual content and assuming only local knowledge of degree for existing pages. A similar phase transition is found between link probability and semantic distance, and both results indicate that efficient paths can be discovered by Web crawling algorithms based on textual and/or categorical cues. I will conclude by surveying a number of applications of these findings to the evaluation and design of more efficient, effective, and scalable search engines and crawlers.
Speaker Bio: Filippo Menczer is an Assistant Professor in the Department of Management Sciences at the University of Iowa, and a faculty of the graduate program in Applied Math and Computational Sciences. He received a Laurea in Physics from the University of Rome in 1991 and a dual Ph.D. in Computer Science and Cognitive Science from the University of California at San Diego in 1998. Dr. Menczer has been the recipient of Fulbright, Rotary Foundation, and NATO fellowships, and is a fellow-at-large of the Santa Fe Institute. He is also the recipient of a CAREER Award from the National Science Foundation.Dr. Melanie Mitchell, Associate Professor, is the local event host.
2/26/2003 -- The Attentive Brain presented by Terrence Sejnowski from OGI/OHSU.
Abstract: Brains are never at rest. Even in the absence of sensory stimuli, neurons are spontaneously active and form highly coherent patterns of activity. Experiments and models suggest that these dynamic patterns reflect top-down influences that allow us to expect, attend and flexibly respond.
Speaker Bio:
2/5/2003 -- Analogy-Making as Perception presented by Melanie Mitchell from OGI/OHSU.
Abstract: Analogy-making is generally acknowledged to be central and ubiquitous in thought. However, in the cognitive science community, analogy-making has most often been modeled as a high-level, goal-driven reasoning process -- a ``big gun\'\' for problem solving, to be invoked when a problem cannot be solved in a conventional way.In this talk I will argue that the ``analogy-making as reasoning\'\' approach does not address the most important and central aspects of analogy-making. I will argue that instead, we must model analogy-making as a form of high-level perception -- a result of the same mental mechanisms underlying categorization, recognition, reminding, and other unconscious perceptual activities. This view requires models of analogy-making to focus on the ``fluidity\'\' of concepts -- their overlapping and associative nature, their blurry boundaries, their dynamic and graded relevance in a given situation, their flexibility as a function of context---in short, their fluid rather than rigid adaptability to different situations. Such fluidity is a hallmark of human thought and its source is not well understood.To illustrate these arguments, I will describe ``Copycat\'\', a computer program developed by myself and Douglas Hofstadter, which models the mental mechanisms underlying fluid concepts and analogy-making. Copycat creates analogies between idealized situations involving strings of letters. Analogy-making in this stripped-down, seemingly simple domain requires many of the same abilities humans use to understand and to make analogies between more complex, real-world situations.Copycat\'s architecture is neither symbolic nor connectionist, but occupies an unique intermediate level in the spectrum of cognitive models. The claim of our model is that this level is at present the most appropriate one for understanding human concepts and analogy-making.I will conclude the talk with a discussion of my current research on applying these ideas to the domain of image understanding.
Speaker Bio: Melanie Mitchell received a Ph.D. in Computer Science from the University of Michigan in 1990. Her dissertation work with Douglas Hofstadter was on cognitive modeling of high-level perception and analogy-making. In 1990 she was awarded a Junior Fellowship in the University of Michigan Society of Fellows. From 1992 to 1999 she was Research Professor at the Santa Fe Institute, and directed the Institute\'s program in Adaptive Computation. From 1999-2000 she was a staff member of the Biophysics Group at the Los Alamos National Laboratory. From 2000-2002 she was a researcher at the Santa Fe Institute and directed the Institute\'s Educational Outreach program. In 2002 she joined the faculty of the Computer Science and Engineering Department at the OGI School of Science and Engineering, Oregon Health and Science University. She is also a member of the external faculty of the Santa Fe Institute.
1/21/2003 -- Particle filters for state estimation in mobile robotics presented by Dieter Fox from University of Washington.
Abstract: Over the last years particle filters have been applied with great practical success to various types of state estimation problems. In this talk I will present some applications of particle filters to mobile robotics. In particular, I will describe how we can use particle filters to mobile robot localization, which is the problem of estimating the position of a robot from sensor data. A variant of this method allows to track the positions of people in a robot\'s vicinity. Furthermore, I will present recent work on adaptive and real-time particle filters. The techniques will be illustrated and discussed in the context of museum tour-guide robots and soccer playing AIBO robots.
Speaker Bio: Dieter Fox is currently an Assistant Professor of Computer Science at the University of Washington, Seattle. Fox\'s research interests lie in artificial intelligence, probabilistic state estimation for mobile robotics, and multi-robot collaboration. Dr. Fox obtained his Ph.D. from the University of Bonn, Germany. Before joining UW, he spent two years as a postdoctoral researcher at the robot learning lab at Carnegie Mellon University. Together with colleagues, he successfully deployed the mobile robots Rhino and Minerva as museum tour-guides in two populated museums, one of them the Smithsonian\'s National Museum of American History in Washington, DC.
1/8/2003 -- Semiconductor Engineering Meets Neuroscience presented by Dan Hammerstrom from OGI/OHSU.
Abstract: As we approach molecular scale semiconductor electronics, there is an increasing discrepancy between our computational models and the functional substrate provided by the hardware. In this talk I propose that we should look to computational neurobiology as one source for inspiration in creating new models that solve important problems and are a better match to molecular scale computing. Consequently, the goal of The Center for Biologically Inspired Information Engineering (CBIIE) at OGI is to develop applications using models derived from computational neurobiology. One important model is associative memory. In this talk, I describe the development of a basic association model, which is a useful function with broad applicability.Creating a robust, scalable association model has been an interesting engineering process. During this development, which is still on-going, we have encountered a number of problems. We have developed solutions for some of these problems, but are struggling with others. This talk will describe the design process and current state for the development. An important issue is the ability to execute large association models in real-time, so a number of implementation issues will also be discussed.
Speaker Bio: Dan Hammerstrom received the PhD degree from the University of Illinois in 1977 in Electrical Engineering. He was an Assistant Professor in the Electrical Engineering Department at Cornell University from 1977 to 1980. In 1980 he joined Intel in Oregon, where he participated in the development and implementation of the iAPX-432, the i960, and iWarp. In 1988 he founded Adaptive Solutions, Inc., which specialized in high performance silicon technology (the CNAPS chip set) for image processing and pattern recognition. He served as President from 1988-1990, and Chief Technical Officer from 1990 to 1997. In 1997 he moved to the ECE department at the Oregon Graduate Institute. Dr. Hammerstrom is also a Professor in the IDE (Information, Computation, and Electronics) Department at Halmstad University, Halmstad, Sweden.
12/4/2002 -- Sensory Adaptation in Mormyrid Electric Fish presented by Patrick Roberts from NSI.
Abstract: In the electrosensory lateral line lobe (ELL) of mormyrid electric fish, synaptic plasticity allows Purkinje-like medium ganglion cells to store and retrieve temporal information. The stored image in the ELL is a prediction about sensory input built from associations between centrally originating predictive signals and peripherally originating sensory input.Recent experiments carried out in vitro have confirmed that the pairing of an excitatory postsynaptic potential (EPSP) with a postsynaptic spike in medium ganglion cells depresses the strength of synapses that normally carries temporally correlated signals. The depression has a tight dependence on the temporal order of pre- and postsynaptic events. The postsynaptic spike must follow the onset of the EPSP within a window of up to about 60 msec. Otherwise, there is a non-associative enhancement of the EPSP.To test whether the mechanisms of plasticity studied in vitro are responsible for the in vivo temporal responses of neurons in the ELL, the system is studied using mathematical analyses and computer simulations. The experimental timing relations measured in vitro are found to lead to activity patterns observed in vivo showing the importance of precise timing in pre- and postsynaptic activity for accurate storage of information in the ELL.
Speaker Bio: Patrick Roberts is Assistant Scientist at the Neurological Sciences Institute at OHSU. Dr. Roberts received his Ph.D. in theoretical physics from the University of Gothenburg in 1993. He is a theoretical neuroscientist, specializing in the development of mathematical methods, both analytical and computational, to study dynamics of neural activity patterns, and to help understand the relationship between those dynamics and behavior. His specific areas of research are:1. The storage of temporal patterns in cerebellum-like structures: the dynamics of synaptic plasticity at the site of initial electrosensory information processing in mormyrid electric fish.2. Dynamics of neural activity in the cerebellum: test various hypothetical mechanisms underlying neuronal activity patterns in the cerebellum. The modeling effort will help to bridge the gap between cellular- and systems-level experimental findings.3. Biological learning rules: to analyze the neural dynamics that result from different biological learning rules. Since the timing relations of biological learning rules result from molecular events at the synapse, this line of research helps to link the implications of dynamics at the molecular level, through dynamics at the network level, to the behavior of whole organisms.
|
|
|
![[Navigation]](images/nav-bar.jpg)