CMU talk: Preference Elicitation in Constraint-based Decision Problems

Craig Boutilier, Department of Computer Science, University of Toronto

WHEN: 10:30am Wed., Nov. 30
ABSTRACT: Preference elicitation is generally required when making or recommending decisions on behalf of users whose utility function is not known with certainty. Although one can engage in elicitation until a utility function is perfectly known, in practice, this is infeasible. This talk tackles this problem in constraint-based optimization.

I will first describe a graphical model for utility representation and issues associated with elicitation in this model. I then discuss two methods for optimization with imprecise utility information: a Bayesian approach in which utility function uncertainty is quantified probabilistically; and a distribution-free minimax regret model. Finally, I will describe several heuristic strategies for elicitation.

This work describes several joint projects with: Darius Braziunas, Relu Patrascu, Pascal Poupart and Dale Schuurmans.

SPEAKER BIO: Craig Boutilier received his Ph.D. in Computer Science (1992) from the University of Toronto, Canada. He is Professor and Chair of the Department of Computer Science at the University of Toronto. He was previously an Associate Professor at the University of British Columbia, a consulting professor at Stanford University, and has served on the Technical Advisory Board of CombineNet, Inc. since 2001.
Dr. Boutilier's research interests span a wide range of topics, with a focus on decision making under uncertainty. He has been awarded the Isaac Walton Killam Research Fellowship, and an IBM Faculty Award. He also received the Killam Teaching Award.

CMU talk: Learning Image Manifolds != Manifold Learning

Robert Pless
Washington University in St. Louis

Monday, December 5, 2005

Abstract
This talk will detail my explorations in applying Manifold Learning techniques to real problems in image processing. Initial experiments with natural image sets (What is the intrinsic dimension of a Charlie Chaplin video clip?... Do cardio-pulmonary MR-images have a natural 2D parameterization?) illuminate several limitations of existing algorithms. First, using Euclidean (sum-of-squared pixel intensity difference) distance is usually a poor choice of image distance functions for natural images. Second, many natural image manifolds have a cyclic topology (and thus cannot be cleanly embedding into a Euclidean space). Third, natural data sets often include unlabeled examples from multiple, intersecting low-dimensional manifolds.

I will talk about several heuristic (and occasionally well founded) algorithms for choosing effective local image distance measures, finding minimal parameterizations for cyclic manifolds, and simultaneously clustering and parameterizing data from multiple intersecting manifolds. These have been brought together in an end-to-end application which automatically learns the 2D manifold structure of (ungated, free-breathing) cardiac MRI images of a patient, and uses the manifold structure of the images to regularize the segmentation of the left ventricle simultaneously in all frames.


Short Bio
Robert Pless is an Assistant Professor of Computer Science, and Assistant Director of the Center for Security Technologies at Washington University. His research interests focus on video processing; motion estimation for video surveillance and manifold learning for applications in biomedical imaging. He received a BS from Cornell University in 1994 and a PhD from the University of Maryland in 2000, and was chairman of the IEEE OMNIVIS workshop in 2003.

CMU talk: Human-Robot Systems for Planetary Exploration

Title: Human-Robot Systems for Planetary Exploration

Speaker: Salvatore Domenick Desiano, Research Scientist, Intelligent Systems Division (QSS Group, Inc), NASA Ames Research Center

Date: Thursday, December 1

Abstract:
Planetary robots will be used in many contexts -- Martian and Lunar, alone and with humans, for construction and for scientific exploration, to name a few. The Intelligent Robotics Group at the NASA Ames Research Center develops cross-cutting capabilities that enable robots to perform autonomously in all of these situations.

In this talk, I will focus on the results of the Collaborative Design Systems FY05 demonstration, performed in September. This was the largest demonstration of integrated robotic systems ever carried out at NASA Ames. The demonstration included visual target tracking, autonomous multi-SCIP (Single Cycle Instrument Placement), constraint-based temporal planning, human-robot collaboration, spoken dialog interfaces, multi-agent systems, and 3D visualization tools.

In addition to the results of this specific demonstration, I will briefly present some of the open research problems on which our group is interested in working or collaborating. I will also provide some inside perspective on the current state of NASA's robotics programs and funding sources.

Speaker Bio:
Salvatore Domenick Desiano is a robotics research scientist at the NASA Ames Research Center. As a member of the Intelligent Robotics Group, he leads the K-9 Rover Team of the CDS project, the most elaborate combination of human and robot planetary exploration ever demonstrated at NASA Ames. His research focuses on developing fundamental navigation capabilities for mobile robots, and he works extensively with the NASA Office of Education. He has also served as the Integration Lead for the Personal Satellite Assistant project. Salvatore is currently on leave from being a doctoral candidate at the Robotics Institute and will be returning to the program early 2006.

My talk this Wednesday

Paper:

Multi-Plannar Projection by Fixed-Center Pan-Tile Projectors
{Ikuhisa Mitsugami , Norimichi Ukita , Masatsugu Kidode}

Abstract:

We describe a new steerable projector, whose projection center precisely corresponds with its rotation center, which we call a “fixed-center pan-tilt (FC-PT) projector.” This mechanism allows it be set up more easily to display graph- ics precisely on the planes in the environment than for other steerable projectors; wherever we would like to display graphics, all we have to do are locating the FC-PT projec- tor in the environment, and directing it to the corners of the planes whose 2D sizes have been measured. Moreover, as the FC-PT projector can recognize automatically whether each plane is connected to others, it can display visual in- formation that lies across the boundary line of two planes in a similar way to a paper poster folded along the planes.

Link:
Multi-Plannar Projection by Fixed-Center Pan-Tile Projectors

Paper: The Smart Wheelchair Component System

Richard Simpson, PhD, ATP; Edmund LoPresti, PhD; Steve Hayashi, PhD; Illah Nourbakhsh, PhD; David Miller, PhD

Journal of Rehabilitation Research & Development (JRRD) Volume 41, Number 3B, Pages 429–442 May/June 2004

Abstract—While the needs of many individuals with disabilities can be satisfied with power wheelchairs, some members of the disabled community find it difficult or impossible to operate a standard power wheelchair. To accommodate this population, several researchers have used technologies originally developed for mobile robots to create“smart wheelchairs”that reduce the physical, perceptual, and cognitive skills necessary to operate a power wheelchair. We are developing a Smart Wheelchair Component System (SWCS) that can be added to a variety of commercial power wheelchairs with minimal modification. This paper describes the design of a prototype of the SWCS, which has been evaluated on wheelchairs from four different manufacturers. [PDF]

What's New @ IEEE in Signal Processing, November 2005

-- PROTOTYPE COMBINES THERAPEUTIC AND 3-D IMAGING CAPABILITY
High frequency ultrasound waves may allow physicians to both visualize the heart's interior in three dimensions and selectively destroy heart tissue with heat to correct arrhythmias, according to engineers at Duke University who are developing the technology. Building on previous work, the Duke team has created dual-function ultrasound probes that use tiny cables, as many as two hundred of them in a three-millimeter catheter. To destroy aberrant tissue in the heart, physicians currently use electrodes that must touch the target tissue, and are guided by x-rays, which do not provide sharp images of soft tissue. The Duke engineers say their prototype can destroy target tissue without touching it, and is guided by much cleared three-dimensional imaging. The work is described in two research papers published in last month in the journals "IEEE Transactions on Ultrasonics, Ferroelectronics and Frequency Control" and "Ultrasonic Imaging." Read more

-- USING ULTRASOUND TECHNOLOGY TO LOOK INSIDE CONCRETE
Researchers at Cambridge Ultrasonics, in conjunction with UK firm Sonatest, have developed an ultrasound sensor that can "see" inside concrete. The sensor works by firing sound waves from up to six different transducers and then registering the returning echoes. A visual map of the inside of the concrete is then displayed as a three-dimensional image. The system, still in the testing stage, was designed to monitor concrete structures for interior corrosion, such as cracks and fissures, particularly in a building's tendons, which act as the skeletons of structures. But the sensor is also of particular interest to police organizations, which could use the device to locate corpses buried in concrete. Bodies buried in concrete break down and leave voids which the sensor would record. Cambridge Ultrasonics is also working on a monitoring system which could be attached to structures to provide regular feedback on corrosion. Read more

What's New @ IEEE in Wireless, November 2005

-- PLANNING SMART BUILDINGS TO AID FIRST RESPONDERS
The National Institute of Standards and Technology (NIST) says it is working with the building industry, public safety officials and information technologists to study how "intelligent" building systems can be used by firefighters, police and other first responders to assess emergency conditions in real-time. NIST is developing standards for various types of communication networks (including wireless networks) to transmit real-time building sensor information on mechanical systems, elevators, lighting, security and fire systems, occupant locations, and temperature and smoke conditions to first responders. According to NIST, the network information would include floor plans and live data from motion, heat, biochemical and other sensors and video cameras. Read more

In related first-responder news, the article "Service-Based Computing on
Manets: Enabling Dynamic Interoperability of First Responders," can be
found in the current issue of IEEE Intelligent Systems magazine

CMU talk: Consistent Segmentation for Optical Flow Estimation

Larry Zitnick

When: Monday, November 21, 3:30 p.m.- 4:45 p.m.
Abstract:
The computation of optical flow in the presence of large displacements and occlusion boundaries is a difficult problem, both in terms of accuracy and computational efficiency. Recently, work in stereo vision has shown promising results using image segmentation to constrain the matching process. Unfortunately, these same segmentation approaches are highly inefficient for optical flow estimation, due to the increased search space (2D vs. 1D) needed for optical flow. We propose a new approach that simultaneously computes a consistent segmentation across images while estimating optical flow. This approach leads to a computationally efficient algorithm while producing accurate results. In addition, we'll present results in video interpolation and exaggerated motion blur using the computed flow fields.

MIT talk: Vision-based robotics: Representation, Mapping and Exploration

Speaker: Robert Sim, University of British Columbia
Date: Tuesday, November 22 2005

Autonomous mobile robot systems have an important role to play in a wide variety of application domains. A key component for autonomy is the capability to explore an unknown environment and construct a representation that a robotic agent can use to localize, navigate, and reason about the world. In this talk I will present results on the automatic construction of visual representations. First, the Visual Map representation will be introduced as a method for modelling the visual structure of the world. Second, I will present a flexible architecture for robust real-time vision-based mapping of an unknown environment. Finally, I will conclude with a discussion of recent progress on the problem of autonomous robotic exploration, and illustrate issues in the problem of developing robotic explorers that are naturally curious about their environment.

The Visual Map framework is an approach to representing the visual world that enables a robot to learn models of the salient visual features of an environment. A key component of this representation is the ability to learn mappings between camera pose and image-domain features without imposing a priori assumptions about the structure of the environment, or the optical characteristics of the visual sensor. These mappings can be employed as generative models in a Bayesian framework for solving the robot localization problem, as well as for visual servoing and path planning.

The second part of this talk demonstrates an architecture for performing simultaneous localization and mapping with vision. The main goals of our work are to facilitate robust large-scale mapping in real time using vision. We employ a Rao-Blackwellised particle filter for managing uncertainty and examine a variety of robust proposal distributions, as well as the run-time and scaling characteristics of our architecture.

The latter part of this builds on representation and mapping to address robotic exploration. In order to acquire a representation of the world, a robot must first acquire data. From an information-theoretic point of view, this problem involves moving through the world so as to maximize the information that can be gained
from what is observed along the robot's trajectory. However, computing the optimal trajectory is complicated by several factors, including the presence of noise, the time horizon over which the robot plans, the specific objective function that is optimized, and the robot's choice of sensor. I will present several results in this area that lead to the development of robust robotic systems that can plan over the long term and successfully demonstrate an emergent sense of curiosity.

MIT PhD Defense: A Unified Information Theoretic Framework for Pair- and Group-wise Registration of Medical Images

Speaker: Lilla Zollei , MIT CSAIL Vision Research Group
Date: Tuesday, November 22 2005
Time: 10:30AM to 12:00PM
Location: Star Seminar Room, 32-D463, Stata Center
Host: Prof. Eric Grimson, Head, EECS Dept.; MIT CSAIL Vision Research Group

Abstract:
The field of medical image analysis has been rapidly growing for the past two decades. Besides a significant growth in computational power, scanner performance, and storage facilities, this acceleration is partially due to an unprecedented increase in the amount of data sets accessible for researchers. Medical experts traditionally rely on manual comparisons of images, but the abundance of information now available makes this task increasingly difficult. Such a challenge prompts for more automation in processing the images.

In order to carry out any sort of comparison between multiple medical images, one frequently needs to identify the proper correspondence between them. This step allows us to follow the changes that happen to anatomy throughout a time interval, to identify differences between individuals, or to acquire complementary information from different data modalities. Registration achieves such correspondences. In this dissertation we focus on the unified analysis and characterization of statistical registration approaches.

First we formulate and interpret a select group of pair-wise registration methods in the context of a unified statistical and information theoretic framework. This clarifies the implicit assumptions of each method and yields a better understanding of their relative strengths and weaknesses. This guides us to a new registration algorithm that incorporates the advantages of the previously described methods. Next we extend the unified formulation with analysis of the group-wise registration algorithms that align a population as opposed to pairs of data sets. Finally, we present our group-wise registration framework, stochastic congealing. That algorithm runs in a simultaneous fashion, with every member of the population approaching the central tendency of the collection at the same time. It eliminates the need for selecting a particular reference frame a priori, resulting in a non-biased estimate of a digital template. Our algorithm adopts an information theoretic objective function which is optimized via a gradient-based stochastic approximation process embedded in a multi-resolution setting. We demonstrate the accuracy and performance characteristics of stochastic congealing via experiments on both synthetic and real images.

11/23 lab meeting

Dear all,

This is the paper I'll present.

Modeling the Static and the Dynamic Parts of the Environment
to Improve Sensor-based Navigation

ICRA 2005

11/16 lab meeting

Dear all,

This is the paper I'll present.

Simultaneous Localization and Mapping with
Detection and Tracking of Moving Objects

-C.-C. Wang and C. Thorpe.In IEEE International Conference on Robotics and Automation (ICRA'02), May, 2002.

CMU talk: Automatic Filters for the Detection of Coherent Structure in Spatiotemporal Systems

Cosma Shalizi
November 29, 2005
Abstract: Current methods for identifying coherent structures in spatially-extended systems rely on prior information about the form which those structures take. This talk describes two new approaches to automatically filter the changing configurations of spatial dynamical systems and extract coherent structures. One, local sensitivity filtering, gauges the ability of locally-applied perturbations to produce large-scale changes in the system configuration. The other, local statistical complexity filtering, calculates the amount of information needed for optimal prediction of the system's behavior in the vicinity of a given point. By examining the changing spatiotemporal distributions of these quantities, we can find the coherent structures in a variety of pattern-forming systems, without needing to guess or postulate the form of that structure. The results are at least comparable to those obtained with older techniques based on formal language theory or the statistical- mechanical theory of order parameters. Paper URL: http://arxiv.org/abs/nlin.CG/0508001

CMU talk: Scalable Inference in Hierarchical Models of the Neocortex

Tom Dean

November 21, 2005
Title: Scalable Inference in Hierarchical Models of the Neocortex
Abstract:
Borrowing insights from computational neuroscience, we present a class of generative models well suited to modeling perceptual processes and an algorithm for learning their parameters that promises to scale to learning very large models. The models are hierarchical, composed of multiple levels, and allow input only at the lowest level, the base of the hierarchy. Connections within a level are generally local and may or may not be directed. Connections between levels are directed and generally do not span multiple levels. The learning algorithm falls within the general family of expectation maximization algorithms. Parameter estimation proceeds level-by-level starting with components in the lowest level and moving up the hierarchy. Having learned the parameters for the components in a given level, those parameters are fixed and needn't be revisited for the purposes of learning. These parameters do, however, play an important role in learning the parameters for higher-level components by helping to generate the samples used in subsequent parameter estimation. Within levels, learning is decomposed into many local subproblems suggesting a straightforward parallel implementation. The inference required for learning is carried out by local message passing and the arrangement of connections within the underlying networks is designed to facilitate this method of inference. Learning is unsupervised but can be easily adapted to accommodate labeled data. In addition to describing several variants of the basic algorithm, we present preliminary experimental results demonstrating the pattern-recognition capabilities of our approach and some of the characteristics of the approximations that the algorithms produce.

Stanford Talk: Rethinking State, Action, and Reward in Reinforcement Learning

Satinder Singh
November 7, 2005, 4:15PM

Abstract
Over the last decade and more, there has been rapid theoretical and empirical progress in reinforcement learning (RL) using the well- established formalisms of Markov decision processes (MDPs) and partially observable MDPs or POMDPs. At the core of these formalisms are particular formulations of the elemental notions of state, action, and reward that have served the field of RL so well. In this talk, I will describe recent progress in rethinking these basic elements to take the field beyond (PO)MDPs. In particular, I will briefly describe older work on flexible notions of actions called options, briefly describe some recent work on intrinsic rather than extrinsic rewards, and then spend the bulk of my time on recent work on predictive representations of state. I will conclude by arguing that taken together these advances point the way for RL to address the many challenges of building an artificial intelligence.

About the Speaker
Satinder Singh is an Associate Professor of Electrical Engineering and Computer Science in the University of Michigan, Ann Arbor. His main research interest is in the old-fashioned goal of Artificial Intelligence, that of building autonomous agents that can learn to be broadly competent in complex, dynamic, and uncertain environments. The field of reinforcement learning (RL) has focused on this goal, and accordingly his deepest contributions are in RL.

MIT talk: Visual Recognition: From Generative to Discriminative Models

Speaker: Pietro Perona , Caltech
Date: Monday, November 14 2005

We can easily recognize objects and properties of the world by looking. If machines had this ability they could be much more intelligent and useful. I will present a taxonomy of visual recognition, review the state of the art and discuss a number of fascinating open problems.

Pietro Perona studies the computational aspects of vision; his current focus is visual recognition. He has published on applications of PDEs to image segmentation, human texture perception and segmentation, dynamic vision, grouping, perception of human motion, learning and recognition of object categories, categorization of scenes in human vision, human perception of 3D shape, interaction of attention and recognition. Perona is Professor of Electrical Engineering and of Computation and Neural Systems at the California Institute of Technology (Caltech). He is the Director of the National Science Foundation Engineering Research Center in Neuromorphic Systems Engineering at Caltech.

MIT talk: A robust layered control system for a mobile robot

Speaker: Rodney Brooks , MIT
Date: Tuesday, November 15 2005

Rod will present a historical perspective on robot control, planning, and intelligence and discuss his influential trend-changing paper A robust layered control system for a mobile robot published in IEEE Transactions on Robotics and Automation, 2(1), pages 14-23, April 1986.

The paper is available to download here:
http://people.csail.mit.edu/brooks/papers/AIM-864.pdf

CMU talk: Fast Inference and Learning in Large-State-Space HMMs

Speaker: Sajid Siddiqi, CMU
http://www.ri.cmu.edu/people/siddiqi_sajid.html

Date: November 14

Abstract:
For Hidden Markov Models (HMMs) with fully connected transition models, the three fundamental problems of evaluating the likelihood of an observation sequence, estimating an optimal state sequence for the observations, and learning the model parameters, all have quadratic time complexity in the number of states. We introduce a novel class of non-sparse Markov transition matrices called Dense-Mostly-Constant (DMC) transition matrices that allow us to derive new algorithms for solving the basic HMM problems in sub-quadratic time. We describe the DMC HMM model and algorithms and attempt to convey some intuition for their usage. Empirical results for these algorithms show dramatic speedups for all three problems. In terms of accuracy, the DMC model yields strong results and outperforms the baseline algorithms even in domains known to violate the DMC assumption.

Fast Inference and Learning in Large-State-Space HMMs
S. Siddiqi and A. Moore
Proceedings of the 22nd International Conference on Machine Learning, August, 2005 Paper.

New @ IEEE in Communications, November 2005

1. SOCIAL NETWORKING LAYS FOUNDATION FOR INNOVATIVE DESIGNS
The current special issue of IEEE Internet Computing (v. 9, no. 5) explores how ideas from social networking can propel creative designs in the communications technology field. To help people use communication technologies to understand and manage their social networks more effectively, the issue' guest editors have selected three articles that position social networks and social networking in terms of relationships among individuals. One author focuses on the networked ego of the e-mail user to present sociograms as end-user visualizations of connections between individuals who have been co-addressed on e-mail messages. The second article examines social isolation and depression in elderly individuals and uses social networking and computing technologies to help reduce these feelings and provide health feedback displays. The third article connects physical place, mobile technologies, and social networks into the P3 framework, a system which helps designers determine appropriate geographic context clues for specific social interactions. The guest editors' introduction, along with a sidebar entitled "Resources on Social Networks, Social Networking, and Social Analysis," are available to all readers online: the link


5. FUTURE SPACE MISSIONS MAY LINK MULTIPLE PLATFORMS IN SENSOR NETWORKS
Missions to other planets and moons may one day use combined space, aerial and ground vehicles to deploy sensors and a communications network more robust and adaptable than current one-vehicle missions, researchers say. The new concept would ensure that the failure of one instrument or vehicle would not doom a mission, say scientists from the California Institute of Technology, the University of Arizona and the U.S. Geological Survey. The researchers propose multi-tiered robotic space missions that link orbiting spacecrafts, blimps and balloons with ground robots, all of which will carry instruments which can communicate and interact with instruments on the other platforms to exploit local weather and geographic conditions.
Read more: the link.


14. PAPER SUBMISSIONS TO INFORMATION FUSION CONFERENCE DUE MID-JANUARY
Papers to the 9th International Conference on Information Fusion should be submitted by 15 January 2006. The conference, sponsored by the IEEE Aerospace & Electronic Systems Society, seeks papers on advancements and applications in information fusion, particularly those with special emphasis on non-traditional topics. Some areas of interest include foundational tools, algorithmic developments, technological advancements and applications. The conference will take place in Florence, Italy, next June. For more details, visit: the link.

New @ IEEE for Students, November 2005

7. IEEE MAGAZINE EXAMINES HUMAN-MACHINE COMMUNICATION
Speech technology gets the special-focus treatment in the current issue of IEEE Signal Processing Magazine (v. 22, no. 5). The issue contains nine articles around the theme of "Speech Technology in Human-Machine Communication," along with an introduction by the guest editors, who write that "the full potential of speech technology still remains to be uncovered." The table of contents and abstracts for all papers in the current issue can be found in the IEEE Xplore digital library, where subscribers may also access the full text of the articles: the link.

10. NEPTUNE RISING: IEEE SPECTRUM REPORTS
Scientists studying the world's oceans are limited to short trips during times of the year when weather conditions are most favorable, and the underwater instruments they leave behind lack the power and bandwidth to deliver much useful information. But this month, construction begins on an Internet-connected undersea observatory covering hundreds of thousands of square kilometers of sea floor. When the project, called the North-East Pacific Time Series Undersea Networked Experiments (NEPTUNE), is completed in 2007, instruments such as hydrophones, current sensors, high-definition video cameras and even robotic crawlers will deliver data around the clock.
IEEE Spectrum has more: the link.