CMU talk: Learning Language from its Perceptual Context

Joint Intelligence/LTI Seminar
November 21, 2008

Learning Language from its Perceptual Context
Raymond J. Mooney, University of Texas at Austin

Current systems that learn to process natural language require laboriously constructed human-annotated training data. Ideally, a computer would be able to acquire language like a child by being exposed to linguistic input in the context of a relevant but ambiguous perceptual environment. As a step in this direction, we present a system that learns to sportscast simulated robot soccer games by example. The training data consists of textual human commentaries on Robocup simulation games. A set of possible alternative meanings for each comment is automatically constructed from game event traces. Our previously developed systems for learning to parse and generate natural language (KRISP and WASP) were augmented to learn from this data and then commentate novel games. The system is evaluated based on its ability to parse sentences into correct meanings and generate accurate descriptions of game events. Human evaluation was also conducted on the overall quality of the generated sportscasts and compared to human-generated commentaries.

Bio:
Raymond J. Mooney is a Professor in the Department of Computer Sciences at the University of Texas at Austin. He received his Ph.D. in 1988 from the University of Illinois at Urbana/Champaign. He is an author of over 150 published research papers, primarily in the areas of machine learning and natural language processing. He is the current President of the International Machine Learning Society, was program co-chair for the 2006 AAAI Conference on Artificial Intelligence, general chair of the 2005 Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing, and co-chair of the 1990 International Conference on Machine Learning. He is a Fellow of the American Association for Artificial Intelligence and recipient of best paper awards from the National Conference on Artificial Intelligence, the SIGKDD International Conference on Knowledge Discovery and Data Mining, the International Conference on Machine Learning, and the Annual Meeting of the Association for Computational Linguistics. His recent research has focused on learning for natural-language processing, text mining for bioinformatics, statistical relational learning, and transfer learning.

CMU talk: Rain in Vision and Graphics

Special VASC Seminar
Tuesday, November 18, 2008

Rain in Vision and Graphics
Kshitiz Garg
intuVision

Abstract:
Rain produces sharp intensity fluctuations in images and videos which severely degrade the performance of outdoor vision systems. Considering that bad weather is common, a city like New York has bad weather 23% of time, it is important to remove the visual effects of rain to make outdoor vision robust. In contrast, in graphics, rain effects are desirable. They are often used in movies to convey scene emotions and in other graphics applications, such as games, to enhance realism. In this talk, I will present rain from the perspective of vision and graphics. I will show how physics based modeling of the visual appearance of rain leads to efficient algorithms both for handling its effects in vision and for its realistic rendering in graphics. I will also briefly discuss some of the recent projects I have done on recognition and tracking at intuVision.


Bio:
Kshitiz Garg is a research scientist and software developer at intuVision. His research interests are in the areas of computer vision, pattern recognition and computer graphics. He has a Masters in Physics and a PhD. in Computer Science from Columbia University, NY. He specializes in physics-based modeling and algorithm development. During his graduate work he developed physics based models for the intensity fluctuations produced by rain in images. He is also interested in Computer Graphics and has developed efficient algorithms for realistic rendering of rain. Since joining the intuVision team, he has worked on algorithms to improve object tracking and recognition especially in the presence of background motion, illumination changes and shadows. He is the research lead for development of intuVision's object classification, face detection, and soft biometry algorithms.

CMU talk: Techniques for Learning 3D Maps

Title: Techniques for Learning 3D Maps

Dr. Wolfram Burgard
Dept. of Computer Science
University of Freiburg

Monday, November 17th

Abstract: Learning maps is a fundamental aspect in mobile robotics, as maps support various tasks including path planning and localization. Whereas the problem of learning maps has been extensively studied for indoor settings, novel field robotics projects have substantially increased the interest in effective representations of outdoor environments. In this talk, we will present our recent results in learning highly accurate multi-level surface maps, which are an extension of elevation maps towards multiple levels. We will describe how multi-level surface maps can be utilized for motion planning and localization. We present an application, in which Junior, the DARPA Grand Challenge entry robot of Stanford University, autonomously drives through a large parking garage and carries out an autonomous parking maneuver. Finally, we will briefly describe our approaches to learning surface maps using variants of Gaussian Processes.

Speaker Bio: Wolfram Burgard is an associate professor for computer science at the University of Freiburg where he heads of the Laboratory for Autonomous Intelligent Systems. He received his Ph.D.~degree in Computer Science from the University of Bonn in 1991. His areas of interest lie in artificial intelligence and mobile robots. Over the past years his research mainly focused on the development of robust and adaptive techniques for state estimation and control of autonomous mobile robots. He and his group developed several innovative probabilistic techniques for robot navigation and control. They cover different aspects such as localization, map-building, path-planning, and exploration.

CMU RI Thesis Proposal: Geolocation from Range: Robustness, Efficiency and Scalability

Joseph Djugash

Robotics Institute

Carnegie Mellon University

Abstract:

In this thesis I explore the topic of geolocation from range. A robust method for localization and SLAM (Simultaneous Localization and Mapping) is proposed. This method uses a polar parameterization of the state to achieve accurate estimates of the nonlinear and multi-modal distributions in range-only systems. Several experimental evaluations on real robots reveal the reliability of this method. 

Scaling such a system to large network of nodes, increases the computational load on the system due to the increased state vector. To alleviate this problem, we propose the use of a distributed estimation algorithm based on the belief propagation framework. This method distributes the estimation task, such that each node only estimates its local network, greatly reducing the computation performed by any individual node. However, the method does not provide any guarantees on the convergence of its solution in general graphs. Convergence is only guaranteed for non-cyclic graphs (ie. trees). Thus, I propose to formulate an extension to this approach that provides guarantees on its convergence and an improved approximation of the true graph inference problem. 

Scaling in the traditional sense involves extensions to deal with growth in the size of the operating environment. In large, feature-less environments, maintaining a globally consistent estimate of a group of mobile agents is difficult. In this thesis, I propose the use of a multi-robot coordination strategy to achieve the tight coordination necessary to obtain an accurate global estimate. The proposed approach will be demonstrated using both simulation and experimental testing with real robots.

Full text

Lab Meeting November 10, 2008 (Yu-chun): “Try something else!” — When users change their discursive behavior in human-robot interaction

ICRA 2008

Authors

Manja Lohse, Katharina J. Rohlfing, Britta Wrede, and Gerhard Sagerer

Abstract

This paper investigates the influence of feedback provided by an autonomous robot (BIRON) on users' discursive behavior. A user study is described during which users show objects to the robot. The results of the experiment indicate, that the robot's verbal feedback utterances cause the humans to adapt their own way of speaking. The changes in users' verbal behavior are due to their beliefs about the robots knowledge and abilities. In this paper they are identified and grouped. Moreover, the data implies variations in user behavior regarding gestures. Unlike speech, the robot was not able to give feedback with gestures. Due to the lack of feedback, users did not seem to have a consistent mental representation of the robot's abilities to recognize gestures. As a result, changes between different gestures are interpreted to be unconscious variations accompanying speech.

pdf

Lab Meeting November 10, 2008 (Alan): An image-to-map loop closing method for monocular SLAM (IROS 2008)

Title: An image–to–map loop closing method for monocular SLAM
Authors: Brian Williams, Mark Cummins, Jos´e Neira, Paul Newman, Ian Reid and Juan Tard´os

Abstract: In this paper we present a loop closure method for a handheld single–camera SLAM system based on our previous work on relocalisation. By finding correspondences between the
current image and the map, our system is able to reliably detect loop closures. We compare our algorithm to existing techniques for loop closure in single–camera SLAM based on both image–
to–image and map–to–map correspondences and discuss both the reliability and suitability of each algorithm in the context of monocular SLAM.

Link

Lab Meeting November 10, 2008 (Any): Efficiently Learning High-dimensional Observation Models for Monte-Carlo Localization using Gaussian Mixtures

Title: Efficiently Learning High-dimensional Observation Models for Monte-Carlo Localization using Gaussian Mixtures

Authors: Patrick Pfaff, Cyrill Stachniss, Christian Plagemann, and Wolfram Burgard

Abstract: Whereas probabilistic approaches are a powerful tool for mobile robot localization, they heavily rely on the proper definition of the so-called observation model which defines the likelihood of an observation given the position and orientation of the robot and the map of the environment. Most of the sensor models for range sensors proposed in the past either consider the individual beam measurements independently or apply uni-modal models to represent the likelihood function. In this paper, we present an approach that learns place-dependent sensor models for entire range scans using Gaussian mixture models. To deal with the high dimensionality of the measurement space, we utilize principle component analysis for dimensionality reduction. In practical experiments carried out with data obtained from a real robot, we demonstrate that our model substantially outperforms existing and popular sensor models.

Local copy

A $1 Recognizer for User Interface Prototypes

It requires under 100 lines of easy code and achieves 97% recognition rates with only one template defined for each gesture below. With 3+ templates defined, accuracy exceeds 99%. Gestures should be regarded as fully rotation, scale, and position invariant.

UIST'07 Paper

Javascript Example

CMU VASC Seminar: What does the sky tell us about the camera?

What does the sky tell us about the camera?
Jean-Francois Lalonde
Robotics Institute, Carnegie Mellon

VASC Seminar
Monday, November 10


Abstract: As the main observed illuminant outdoors, the sky is a rich source of information about the scene. However, it is yet to be fully explored in computer vision because its appearance in an image depends on the sun position, weather conditions, photometric and geometric parameters of the camera, and the location of capture. In this talk, I will present an analysis of two sources of information available within the visible portion of the sky region: the sun position, and the sky appearance. By fitting a model of the predicted sun position to an image sequence, we show how to extract camera parameters such as the focal length, and the zenith and azimuth angles. Similarly, we show how we can extract the same parameters by fitting a physically-based sky model to the sky appearance. In short, the sun and the sky serve as geometric calibration targets, which can be used to annotate a large database of image sequences. We use our methods to calibrate 22 real, low-quality webcam sequences scattered throughout the continental US, and show deviations below 4% for focal length, and 3 degrees for the zenith and azimuth angles. Once the camera parameters are recovered, we use them to define a camera-invariant sky appearance model, which we exploit in two applications: 1) segmentation of the sky and cloud layers, and 2) data-driven sky matching across different image sequences based on a novel similarity measure defined on sky parameters. This measure, combined with a rich appearance database, allows us to model a wide range of sky conditions.

Bio: Jean-Francois Lalonde received his B.E. in Computer Engineering from Laval University, Canada in 2004. He received his M.S. in Robotics from Carnegie Mellon University in 2006 under Martial Hebert, and he has been a Robotics Ph.D. student advised by Alexei A. Efros in that institution since. His research interests are in computer vision and computer graphics, focusing on image understanding and synthesis
leveraging large amounts of data.

CMU talk: Computing with Language and Context over Time

Speaker: Gregory Aist, Arizona State University

Title: Computing with Language and Context over Time

What: Joint LTI/RI Seminar
When: Friday November 7, 2008, 2:00pm - 3:00pm
Where: 1305 NSH

Abstract:
How do language and context interact in learning and performance by humans and machines? To explore this broad area of inquiry, I have studied interactions between natural language and a wide range of different contexts: visual context, social and team context, written context and world knowledge, procedure and task context, dialogue and temporal context, and instructional context. Specific research questions have included how machines can process spoken language continuously and integrate speech and visual context during understanding; how computers can help pilots and astronauts learn and perform tasks; and how to automatically generate, present, and evaluate the effects of vocabulary help for children. One key challenge in addressing all of these questions is to model and compute representations of language and context that unfold over time as the interaction progresses. This talk will illustrate the need for such interactive time-sensitive processes, describe computational approaches to understanding language and context as dialogue and interactions unfold across time, and evaluate the effectiveness of such approaches.

Short bio:
Gregory Aist is currently at Arizona State University as an Assistant Research Professor in the School of Computing and Informatics and the Applied Linguistics Program. His research interests are in natural language processing and computer-assisted learning. His research addresses fundamental issues in language and learning, tackles computational challenges of automatic processing of human language and computer support for human learning, and is applied to provide users with learning experiences and new capabilities in authentic settings for educational domains such as traditional literacy (reading and writing) and new literacies (virtual worlds), and physical domains such as aerospace and human-robot interaction. During summers 2007 and 2008 he was an Air Force Summer Faculty Fellow. Previously he has held research and visiting positions at the University of Rochester, RIACS/NASA Ames Research Center, and the MIT Media Lab. He received a Ph.D. in Language and Information Technology from Carnegie Mellon University in 2000, where he was an NSF Graduate Fellow.

Lab Meeting November 3rd, 2008 (swem): Learning Patch Correspondences for Improved Viewpoint Invariant Face Recognition

Title: Learning Patch Correspondences for Improved Viewpoint Invariant Face Recognition

Author: Ahmed Bilal Ashraf, Simon Lucey, Tsuhan Chen

Abstract:
Variation due to viewpoint is one of the key challenges
that stand in the way of a complete solution to the face
recognition problem. It is easy to note that local regions of
the face change differently in appearance as the viewpoint
varies. Recently, patch-based approaches, such as those of
Kanade and Yamada, have taken advantage of this effect resulting
in improved viewpoint invariant face recognition. In
this paper we propose a data-driven extension to their approach,
in which we not only model how a face patch varies
in appearance, but also how it deforms spatially as the viewpoint
varies. We propose a novel alignment strategy which
we refer to as “stack flow” that discovers viewpoint induced
spatial deformities undergone by a face at the patch level.
One can then view the spatial deformation of a patch as
the correspondence of that patch between two viewpoints.
We present improved identification and verification results
to demonstrate the utility of our technique.

Link

Lab Meeting November 3rd, 2008 (Shao-Chen): Blind spatial subtraction array with independent component analysis for hands-free speech recognition

Title:
Blind spatial subtraction array with independent component analysis for hands-free speech recognition

Authors:
Yu Takahashi, Tomoya Takatani, Hiroshi Saruwatari and Kiyohiro Shikano

Abstract:
In this paper, we propose a new blind spatial subtraction array (BSSA) which contains an accurate noise estimator based on independent component analysis (ICA) to realize a noise-robust hands-free speech recognition. First, a preliminary experiment suggests that the conventional ICA is proficient in the noise estimation rather than the direct speech estimation in real environments, where the target speech can be approximated to a point source but real noises are often not point sources. Secondly, based on the above-mentioned findings, we propose a new noise reduction method which is implemented in subtracting the power spectrum of the estimated noise by ICA from the power spectrum of noise-contaminated observations. This architecture provides us with a noise-estimation-error robust speech enhancement which is well applicable to the speech recognition. Finally, the effectiveness of the proposed BSSA is shown in the speech recognition experiment.

Link:
IWAENC 2006
http://www.commsp.ee.ic.ac.uk/~sap/iwaenc/Proceedings/2006/pdf/A38.pdf

Lab Meeting October 27th, 2008 (slyfox): Bearings-Only Tracking of Manoeuvring Targets Using Particle Filters

Abstract:

We investigate the problem of bearings-only tracking of manoeuvring targets using particle filters (PFs). Three different (PFs) are proposed for this problem which is formulated as a multiple model tracking problem in a jumpMarkov system (JMS) framework. The proposed filters are (i) multiple model PF (MMPF), (ii) auxiliary MMPF (AUX-MMPF), and (iii) jump Markov system PF (JMS-PF). The performance of these filters is compared with that of standard interacting multiple model (IMM)-based trackers such as IMM-EKF and IMM-UKF for three separate cases: (i) single-sensor case, (ii) multisensor case, and (iii) tracking with hard constraints. A conservative CRLB applicable for this problem is also derived and compared with the RMS error performance of the filters. The results confirm the superiority of the PFs for this difficult nonlinear tracking problem.

Link:
EURASIP Journal on Applied Signal Processing, 2004
http://portal.acm.org/citation.cfm?id=1289537

Lab Meeting Octobor 27th, 2008 (fish60): Smooth Nearness-Diagram Navigation

Abstract:

This paper presents a new method for reactive collision avoidance for mobile robots in complex and cluttered environments. Our technique is to adapt the "divide and conquer" approach of the Nearness-Diagram+ Navigation (ND+) method to generate a single motion law which applies for all navigational situations.
The resulting local path planner considers all the visible obstacles surrounding the robot, not just the closest two. With these changes our new navigation method generates smoother motion while avoiding obstacles. Results from comparisons with ND+ are presented as are experiments using Erratic mobile robots.

2008 IROS Paper
Link

Lab Meeting Octobor 27th, 2008 (Jeff):Incremental vision-based topological SLAM

Title: Incremental vision-based topological SLAM

Authors: Adrien Angeli, Stephane Doncieux, Jean-Arcady Meyer, and David Filliat

Abstract:

In robotics, appearance-based topological map building consists in infering the topology of the environment explored by a robot from its sensor measurements. In this paper, we propose a vision-based framework that considers this data association problem from a loop-closure detection perspective in order to correctly assign each measurement to its location. Our approach relies on the visual bag of words paradigm to represent the images and on a discrete Bayes filter to compute the probability of loop-closure. We demonstrate the efficiency of our solution by incremental and real-time consistent map building in an indoor environment and under strong perceptual aliasing conditions using a single monocular wide-angle camera.

Link:
IROS2008 Paper
http://www.ensta.fr/~filliat/papers/Angeli_IROS08.pdf

Lab Meeting October 20, 2008 (Jimmy): Learning in Dynamic Environments with Ensemble Selection for Autonomous Outdoor Robot Navigation

Title: Learning in Dynamic Environments with Ensemble Selection for Autonomous Outdoor Robot Navigation (IROS2008)

Authors: Michael J. Procopio, Jane Mulligan, and Greg Grudic

Abstract:
Autonomous robot navigation in unstructured outdoor environments is a challenging area of active research. The navigation task requires identifying safe, traversable paths which allow the robot to progress toward a goal while avoiding obstacles. Machine learning techniques—in particular, classifier ensembles—are well adapted to this task, accomplishing near-to-far learning by augmenting near-field stereo readings in order to identify safe terrain and obstacles in the far field. Composition of the ensemble and subsequent combination of model outputs in this dynamic problem domain remain open questions. Recently, Ensemble Selection has been proposed as a mechanism for selecting and combining models from an existing model library and shown to perform well in static domains. We propose the adaptation of this technique to the time-evolving data associated with the outdoor robot navigation domain. Important research questions as to the composition of the model library, as well as how to combine selected models’ outputs, are addressed in a two-factor experimental evaluation. We evaluate the performance of our technique on six fully labeled datasets, and show that our technique outperforms memoryless baseline techniques that do not leverage past experience.

[Link]

Lab Meeting October 20, 2008(Casey): 3D Head tracking and pose-robust 2D Texture Map-Based Face Recognition using a Simple Ellipsoid Model

Title: 3D Head tracking and pose-robust 2D Texture Map-based Face Recognition using a Simple Ellipsoid Model (IROS2008)

Authors: Kwang Ho An and Myung Jin Chung

Abstract: 

A human face provides a variety of different communicative functions such as identification, the perception of emotional expression, and lip-reading. For these reasons, many applications in robotics require tracking and recognizing a human face. A novel face recognition system should be able to deal with various changes in face images, such as pose, illumination, and expression, among which pose variation is the most difficult one to deal with. Therefore, face registration (alignment) is the key of robust face recognition. If we can register face images into frontal views, the recognition task would be much easier. To align a face image into a canonical frontal view, we need to know the pose information of a human head. Therefore, in this paper, we propose a novel method for modeling a human head as a simple 3D ellipsoid. And also, we present 3D head tracking and pose estimation methods using the proposed ellipsoidal model. After recovering full motion of the head, we can register face images with pose variations into stabilized view images which are suitable for frontal face recognition. By doing so, simple and efficient frontal face recognition can be easily carried out in the stabilized texture map space instead of the original input image space. To evaluate the feasibility of the proposed approach using a simple ellipsoid model, 3D head tracking experiments are carried out on 45 image sequences with ground truth from Boston University, and several face recognition experiments are conducted on our laboratory database and the Yale Face Database B by using subspace-based face recognition methods such as PCA, PCA+LAD, and DCV.

[Link]

MIT CSAIL talk: Modeling Appearance via the Object Class Invariant

Modeling Appearance via the Object Class Invariant
Speaker: Matthew Toews, Harvard Medical School

Date: Friday, October 17 2008
Time: 2:00PM to 3:00PM
Host: Polina Golland, CSAIL

As humans, we are able to identify, localize, describe and classify a wide range of object classes, such as faces, cars or the human brain, by their appearance in images. Designing a general computational model of appearance with similar capabilities remains a long standing goal in the research community. A major challenge is effectively coping with the many sources of variability operative in determining image appearance: illumination, noise, unrelated clutter, occlusion, sensor geometry, natural intra-class variation and abnormal variation due to pathology to name a few. Explicitly modeling sources of variability can be computationally expensive, can lead to domain-specific solutions and may ultimately be unnecessary for the computational tasks at hand.

In this talk, I will show how appearance can instead be modeled in a manner invariant to nuisance variations, or sources of variability unrelated to the tasks at hand. This is done by relating spatially localized image features (e.g. SIFT) to an object class invariant (OCI), a reference frame which remains geometrically consistent with the underlying object class despite nuisance variations. The resulting OCI model is a probabilistic collage of local image patterns that can be automatically learned from sets of images and robustly fit to new images, with little or no manual supervision. Due to its general nature, the OCI model can be used to address a variety of difficult, open problems in the contexts of computer vision and medical image analysis. I will show how the model can be used both as a viewpoint-invariant model of 3D object classes in photographic imagery and as a robust anatomical atlas of the brain in magnetic resonance imagery.

IROS 2008 Keynote speech: Understanding Human Faces

At IROS 2008, Takeo Kanade delivered a great speech summarizing what he has being working on in terms of understanding human faces. Below is the abstract:

A human face conveys important information: identity, emotion, and intention of the person. Technologies to process and understand human faces have many applications, ranging from biometrics to medical diagnosis, and from surveillance to human-robot interaction. This talk will give an overview of the recent progress that the CMU Face Group has made, in particular, robust face alignment, facial Action Unit (AU) recognition for emotion analysis, and facial video cloning for understanding human dyadic communication.


The video I took is available at http://robotics.csie.ntu.edu.tw/~bobwang/iros2008/. As I selected a wrong/low resolution to record this one-hour talk, it is hard to see the slides. Fortunately, the audio is clear. Take a look (or listen to this excellent talk)!

-Bob

CMU RI Thesis Proposal: Pretty Observable Markov Decision Processes: Exploiting Approximate Structure for Efficient Planning under Uncertainty

Title: Pretty Observable Markov Decision Processes: Exploiting Approximate Structure for Efficient Planning under Uncertainty

Nicholas Armstrong-Crews
Robotics Institute
Carnegie Mellon University

NSH 1507
10:00 AM
20 Oct 2008

Abstract:
Planning under uncertainty is a challenging task. POMDP models have become a popular method for describing such domains. Unfortunately, solving a POMDP to find the optimal policy is computationally intractable, in general. Recent advances in solving POMDPs include finding near-optimal policies and exploiting structured representations of the problems. We believe that using these two tools together synergistically, we can tame the complexity of many POMDPs. In this thesis, we propose to further advance these approaches by analyzing new types of structure and new approximation techniques, as well as the methods combining the intersection of the two.

Some of the research we have done to lay the groundwork for this thesis falls into these categories, with promising results. We introduced the Oracular POMDP framework, which takes advantage of an MDP-like structure by allowing direct observation of the state as a (costly) action by the agent, but otherwise the agent receives no information from the environment and in between invocations of this “oracle'' action the agent is again afflicted by uncertainty. We have given an anytime algorithm for solving Oracular POMDPs which we've proven is efficient (poly-time) in all but the number of actions. At any iteration of the anytime algorithm, we have a (provably) near-optimal policy, which we have achieved efficiently by exploiting the structured observation function.

Another vein of our past work addressing solving general POMDPs by approximating them as finite state MDPs. It is a well-known result that POMDPs are equivalent to continuous MDPs whose state space is the belief simplex (the probability distribution over possible hidden states). We sample a finite number of these beliefs to create a finite-state MDP that approximates the original POMDP. We then solve this MDP for an optimal policy, improve our sample of belief states with this policy so that it better approximates the POMDP, and continue in this fashion.

These prior works exhibit an important common methodology: anytime algorithms that give near-optimal policies at every iteration, and in the limit converge to the optimal policy. This property is paramount for tackling problems with approximate structure. We can focus early iterations on the structured portion of the problem, which we can solve quickly; and later iterations can handle the complex, unstructured portion of the problem. In this way, we can quickly reach a near-optimal solution, while guaranteeing convergence to an optimal solution in the limit. Our method of evaluating an algorithm's performance on a given problem, then, is the entire curve of policy quality versus algorithm runtime.

Although the AI literature is rich with attempts to exploit different types of structure, in this thesis we focus on a small subset. Our prior work includes Oracular POMDPs, an extremely structured observation function; and the finite-state MDP approximation to POMDPs, which takes advantage of a structured belief-state space that is learned as the algorithm progresses.

For the remainder of the thesis work, we propose to generalize the concept of Oracular POMDPs to include nearly perfect information from oracles, with nearly no information provided from the environment otherwise; we will also extend the oracle concept to factored state problems, where an oracle can reveal one state variable reliably but not the others. We will investigate automated techniques for learning structure from a given unstructured representation. Finally, we wish to examine in greater detail what can be proven about the optimality-runtime tradeoff of these approximately structured POMDPs.

To evaluate our methods, we will apply them to several types of problems. First, we will introduce new synthetic domains that exhibit the structure we wish to exploit. Second, we will use our structure learning methods on existing domains in the literature. Finally, we will attempt to apply the methods to a real-world robot problem, in order to address doubts (in the minds of the community and of the author) about the usefulness of POMDP methods real robots.

Full text

Lab Meeting October 13, 2008(Tiffany): Graph Laplacian Based Transfer Learning in Reinforce-ment Learning

Authors:
Yi-Ting Tsao, Ke-Ting Xiao, Von-Wun Soo

Abstract:
The aim of transfer learning is to accelerate learning in related domains. In reinforcement learning, many different features such as a value function and a policy can be transferred from a source domain to a related target domain. Many researches focused on transfer using hand-coded translation functions that are designed by the experts a priori. However, it is not only very costly but also problem dependent. We propose to apply the Graph Laplacian that is based on the spectral graph theory to decompose the value functions of both a source domain and a target domain into a sum of basis functions respectively. The transfer learning can be carried out by transferring weights on the basis functions of a source do-main to a target domain. We investigate two types of domain transfer, scaling and topological. The results demonstrated that the transferred policy is a better prior policy to reduce the learning time.

[Link]

Lab Meeting October 13 (Andi) Extrinsic Laser Scanner / Camera calibration

I will summarize three papers with different approaches for Laser/Camera calibration.

[1] An Algorithm for Extrinsic Parameters Calibration of a Camera and a Laser Range Finder Using Line Features
[2] An efficient extrinsic calibration of a multiple laser scanners and cameras’ sensor system on a mobile platform
[3] Extrinsic calibration of a camera and laser range finder (improves camera calibration)

Abstracts:
[1]This paper presents an effective algorithm for calibrating the extrinsic parameters between a camera and a laser range finder whose trace is invisible. On the basis of an analysis of three possible features, we propose to design a right-angled triangular checkerboard and to employ the invisible intersection points of the laser range finder’s slice plane with the edges of the checkerboard to set up the constraints equations. The extrinsic parameters are then calibrated by minimizing the algebraic errors between the measured intersections points and their corresponding projections on the image plane of the camera....
[2] ...In this research, we present a practical method for extrinsic calibration of multiple laser scanners and video cameras that are mounted on a vehicle platform. Refering to a fiducial coordinate system on vehicle platform, a constraint between the data of a laser scanner and of a video camera is established. It is solved in an iterative way to find a best solution from the laser scanner and from the video camera to the fiducial coordinate system. On the other hand, all laser scanners and video cameras are calibrated for each laser scanner and video camera pair that has common in feature points in a sequential way....
[3] We describe theoretical and experimental results for the extrinsic calibration of sensor platform consisting of a camera and a 2D laser range finder. The calibration is based on observing a planar checkerboard pattern and solving for constraints between the “views” of a planar checkerboard calibration pattern from a camera and laser range finder. We give a direct solution that minimizes an algebraic error from this constraint, and subsequent nonlinear refinement minimizes a re-projection error....

IEEE News: Smart Phones May Detect Sleep Disorders

Technology for screening and diagnosing sleep disorders, and for waking users at the best times in their sleep cycles, has been developed by researchers at two Finnish universities, Tampere University of Technology and the University of Helsinki, who say the first application of the new technology, a smart alarm clock for mobile phones, HappyWakeUp, is now available. The researchers first noticed that a common microphone is very sensitive to any sounds and voices produced by movements in the bed during night-time, they say, and can adapt that technology for the detection of restless sleep, leg movements associated with restless leg syndrome and screening for snoring and sleep apnea. The technology makes it possible to perform several repeated all-night recordings and to diagnose sleep disorders in countries and areas with no previous sleep recording facilities, according to researchers, who say the new technology is extremely cost-efficient, compared to the use of existing s! pecial medical recording devices. Read more
Learn more about broadband wireless in the IEEE Xplore® digital library

Research Scientist Position in Robotics at MIT CSAIL

************************************************************
Research Scientist
Computer Science and Artificial Intelligence Laboratory
Massachusetts Institute of Technology


RESEARCH SCIENTIST, Computer Science and Artificial Intelligence Laboratory (CSAIL), to perform research in the development of perception, planning, control, and human interface software and algorithms for autonomous robots; manage research in autonomous vehicles, including development and testing of techniques for vision, lidar, and radar data processing for mapping, localization, and autonomous path control; and development and field deployment of novel robotic systems for land, air, and sea environments.

REQUIREMENTS: a Ph.D. in robotics or computer vision; and five or more years' experience in perception algorithm and human-computer interface and robotic system programming for autonomous vehicles. Seek motivated, enthusiastic roboticist who demonstrates exceptional programming skills and the ability to perform independent research and manage complex research projects. Must be able to help mentor graduate students and postdocs. Position requires expert knowledge of Bayesian state estimation and computer vision algorithms such as Kalman filters, particle filters, and SIFT feature detection; and general experience in robot system integration, C/C++ network programming in Linux and Windows, CVS, SVN, openGL, perl, and HTML. Must have experience with configuration and management of Linux computer systems using Ubuntu/Debian distributions; deployment and operation of mobile ad-hoc wireless networks; and code development for public-domain robot control software packages such as CARMEN and LCM. Should also have experience creating real-time interfaces to vision, laser, and radar sensors using serial, USB, CANbus, and tcp/ip connections; and in the configuration and operation of SICK laser range scanners.

Applicants may apply online at http://hrweb.mit.edu/staffing/
(Search for position mit-00005935)

John Leonard (jleonard@mit.edu)

HRI 2009 Evaluation Criteria

The criteria are good guidelines for doing human robot interaction research. -Bob

The Evaluation Criteria for papers are now available. All papers must:

a) Be relevant to the field of human-robot interaction. So, for example, a paper that describes a new face tracking algorithm needs to demonstrate how it is of direct use to human-robot interaction. A paper contributing a face recognition technology should use standard recognition metrics (e.g., ROC curve) as well as demonstrate or highlight a path to “feasibility” in human-robot domains with regard to interactive performance, sufficient accuracy, integration with closed loop control, etc. Similarly, a study of the elderly must show how insights from the study directly inform the design of robots for this population and a wizard-of-oz experiment should show how findings contribute to our understanding of how people might interact with robotic capabilities that are plausible (if not currently available).

b) Clearly articulate: 1) the contribution to HRI, 2) how the contribution advances the state-of-the art or knowledge in HRI, and 3) how the contribution relates to other work in HRI as well as the fields of study on which the paper draws (e.g. psychology, cognitive science, anthropology, computer vision, artificial intelligence, speech recognition, etc.).

c) Be technologically and methodologically sound based on the criteria generally used for that technology/method within a given field. For example, conventions used in psychology for conducting experiments with people and analyzing the data, and reporting the study (e.g. hypotheses, manipulation checks, the creation of scales, ANOVA analyses, correlation tables, etc.) should be applied. Authors should take care to use correct terminology for their methods to avoid being evaluated against the incorrect set of criteria. For example, a user study of 5 people should be referred to as a user study or evaluation and not an experiment.

d) For empirical papers, provide adequate detail for readers to understand what was done, how the data were collected, from how many people, what were the characteristics of these people, what questions were people asked, what type of robot was involved (if a robot was used), etc.

e) Be written to be accessible for a broad, interdisciplinary/multidisciplinary HRI audience.

We particularly encourage papers that bring together subfields and investigate problems that have not been explored and are novel to HRI.

CMU RI Thesis Proposal: Generalized Backpropagation

Title: Generalized Backpropagation

David Bradley

Robotics Institute

Carnegie Mellon University

Place: NSH 1507

Time: 2:00 PM

Date: 2 Oct 2008

Abstract:

Robotics problems are often ill-suited for the standard supervised learning paradigms. Consequently, complex tasks like autonomous driving through difficult off-road terrain are decomposed into a series of small subproblems which are each solved with independent modules through learning or hand-engineering. Unfortunately, this approach exacerbates the problem of data labeling, as training data sets must now be labeled for intermediate outputs where the "correct" label that will lead to optimal system performance may be hard to define. Additionally, even if each of the subproblems is solved reasonably well, the performance of the system as a whole may suffer because of the accumulation of small errors in each module. Figuring out which individual modules in a complex system are responsible for system-wide errors, and how to adjust them to improve performance, has consumed many man-hours. 

The gradient backpropagation algorithm has been integral for neural network training because it provides a way to translate errors in the final system performance into updates for each of the parameters in a complex network. Although it is commonly associated with neural networks, backpropagation can be used to train any system of differentiable modules. Backpropagation depends exclusively on labeled data, however, and fails on deep networks where the "blame" for errors becomes too diffuse. This thesis proposes a set of tools, collectively termed "generalized backpropagation", which are derived from recent research into learning in the online, semi-supervised, transductive, and multi-task settings, and address some key limitations of the backpropagation algorithm. These tools make it easier to learn networks of modules if labeled training data is limited and can be understood as ways to create informative priors for the parameters of each module from unlabeled and weakly labeled data. In the proposed work they will be demonstrated on a challenging mobile robot navigation problem.

Lab Meeting September 29, 2008(ZhenYu): Screen-Camera Calibration using a Spherical Mirror

Authors: Yannick Francken, Chris Hermans, Philippe Bekaert
Abstract:

Recent developments in the consumer market have indicated that the average user of a personal computer is likely to also own a webcam. With the emergence of this new user group will come a new set of applications, which will require a user-friendly way to calibrate the position of the camera with respect to the location of the screen.
This paper presents a fully automatic method to calibrate a screen-camera setup, using a single moving spherical mirror. Unlike other methods, our algorithm needs no user intervention other then moving around a spherical mirror. In addition, if the user provides the algorithm with the exact radius of the sphere in millimeters, the scale of the computed solution is uniquely defined.

[Link]

CfP NIPS 2008 workshop "Learning over Empirical Hypothesis Spaces"

NIPS-2008 Workshop on "Learning over Empirical Hypothesis Spaces"

Call for Contributions
http://www.kuleuven.be/wehys/
Whistler, BC, Canada
December 13, 2008

Important Dates:
- Deadline: October 31, 2008,
- Notification: November 7, 2008

Workshop Chairs:

. Maria-Florina Balcan
. Shai Ben-David
. Avrim Blum
. Kristiaan Pelckmans
. John Shawe-Taylor

Contact:
Wehys08@gmail.com

Scope:

This workshop aims at collecting theoretical insights in the design of data-dependent learning strategies. Specifically we are interested in how far learned prediction rules may be characterized in terms of the observations themselves. This amounts to capturing how well data can be used to construct structured hypothesis spaces for risk minimization strategies - termed empirical hypothesis spaces. Classical analysis of learning algorithms requires the user to define a proper hypothesis space before seeing the data. In practice however, one often decides on the proper learning strategy or the form of the prediction rules of interest after inspection of the data (see e.g. [5, 7]). This theoretical gap constitutes exactly the scope of this workshop. A main theme is then the extent to which prior knowledge or additional (unlabeled) samples can or should be used to improve learning curves. ...(read further - )


Tentative Program:

One day divided in to four sessions, two morning, two afternoon with coffee between. Each session would have one invited contributor talking for 45 mins followed by 15 mins discussion, except the first where there would be two 45 min tutorial presentations. The sessions would each have an additional part:
* Session 1 Tutorials by S. Ben-David and A. Blum;
* Session 2 Invited talk plus two contributed 15 mins presentations (posters to be shown in the afternoon);
* Session 3 Invited talk plus spotlight (2min) presentations for posters with poster session following during coffee break;
* Session 4 Invited talk followed by discussion aimed at identifying 10 key open questions.

Both John Langford and Csaba Szepesvari will present their take on the problem. The final program will be announced soon.

Call for contributions:

We solicit discussions and insights (controversial or otherwise) into any of the following topics:
1.
Relations between the luckiness framework, compatibility functions and empirically defined regularization strategies in general.
2.
Luckiness and compatibility can be seen as defining a prior in terms of the (unknown but fixed) distribution generating the data. To what extent can this approach be generalised while still ensuring effective learning?
3.
Models of prior knowledge that capture both complexity and distribution dependence for powerful learning.
4.
Theoretical analysis of the use of additional (empirical) side information in the form of unlabeled data or data labeled by related problems
5.
Examples of proper or natural luckiness or compatibility functions in practical learning tasks. How could, for example, luckiness be defined in the context of collaborative filtering?
6.
The effect of (empirical) preprocessing of the data not involving the labels as for example in PCA, other data-dependent transformations or cleaning, as well as using label information as for example in PLS or in feature selection and construction based on the training sample.
7.
Empirically defined theoretical measures such as Rademacher complexity or sparsity coefficients and their relevance for analysing empirical hypothesis spaces.

This workshop is intended for researchers interested in the theoretical underpinnings of learning algorithms which do not comply to the standard learning theoretical assumptions.

Submissions should be in the form of a 2-page abstract (i) summarizing a formal result, (ii) a discussion of its relevance to the workshop and (iii) pointers to the relevant literature. The abstract can be supported by an additional paper (either published or technical report), that contain detailed proofs of any assertions. We especially encourage contributions which describe how to bring in results from other formal frameworks.

CFP: NIPS 2008 Workshop on Analyzing Graphs: Theory and Applications

#####################################################################

CALL FOR PAPERS

Analyzing Graphs: Theory and Methods

a workshop in conjunction with

22nd Annual Conference on Neural Information Processing Systems
(NIPS 2008)

December 12, 2008 Whistler, BC, Canada

http://research.yahoo.com/workshops/nipsgraphs2008/

Deadline for Submissions: Friday, October 31, 2008
Notification of Decision: Friday, November 10, 2008

#####################################################################


Overview:

Recent research in machine learning and statistics has seen the proliferation of computational methods for analyzing graphs and networks. These methods support progress in many application areas, including the social sciences, biology, medicine, neuroscience, physics, finance, and economics.

This workshop will address statistical, methodological and computational issues that arise when modeling and analyzing graphs. The workshop aims to bring together researchers from applied disciplines such as sociology, economics, medicine and biology with researchers from mathematics, physics, statistics and computer
science. Different communities use diverse ideas and mathematical tools; our goal is to foster cross-disciplinary collaborations and intellectual exchange.

Presentations will include novel graph models, the application of established models to new domains, theoretical and computational issues, limitations of current graph methods and directions for future research.


Online Submissions:
-------------------
We welcome the following types of papers:

1. Research papers that introduce new models or apply established models to novel domains,

2. Research papers that explore theoretical and computational issues, or

3. Position papers that discuss shortcomings and desiderata of current approaches, or propose new directions for future research.

All submissions will be peer-reviewed; exceptional work will be considered for oral presentation. We encourage authors to emphasize the role of learning and its relevance to the application domains at hand. In addition, we hope to identify current successes in the area, and will therefore consider papers that apply previously proposed models to novel domains and data sets.

Submissions should be 4-to-8 pages long, and adhere to NIPS format (http://nips.cc/PaperInformation/StyleFiles). Please email your submissions to: nipsgraphs2008@yahoo.com

Deadline for Submissions: Friday, October 31 2008
Notification of Decision: Friday, November 10 2008


Format

This is a one-day workshop. The program will feature invited talks, poster sessions, poster spotlights, and a panel discussion. All submissions will be peer-reviewed; exceptional work will be considered for oral presentation.

Publication:

Accepted papers will be distributed on a CD and made available for download. We are negotiating the publication of the accepted papers in print form.


Organizers:
-----------
Edo Airoldi, Princeton University, eairoldi@princeton.edu
David Blei, Princeton University, blei@cs.princeton.edu
Jake Hofman, Yahoo! Research, hofman@yahoo-inc.com
Tony Jebara, Columbia University, jebara@cs.columbia.edu
Eric Xing, Carnegie Mellon University, epxing@cs.cmu.edu


Program Committee
-----------------
David Banks (Duke University)
Peter Bearman (Columbia University)
Joseph Blitzstein (Harvard University)
Kathleen Carley (Carnegie Mellon University)
Aaron Clauset (Santa Fe Institute)
William Cohen (Carnegie Mellon University)
Stephen Fienberg (Carnegie Mellon University)
Lise Getoor (University of Maryland)
Peter Hoff (University of Washington)
Eric Horvitz (Microsoft Research)
Alan Karr (National Institute of Statistical Sciences)
Jure Leskovec (Carnegie Mellon University)
Kevin Murphy (University of British Columbia)
Eugene Stanley (Boston University)
Lyle Ungar (Universitoy of Pennsylvania)
Chris Wiggins (Columbia University)

Lab Meeting September 29, 2008(Chung-Han): A Mobile Vision System for Robust Multi-Person Tracking

Andreas Ess, Bastian Leibe, Konrad Schindler, Luc Van Gool
in IEEE Conference on Computer Vision and Pattern Recognition (CVPR'08)

Abstract:
We present a mobile vision system for multi-person tracking in busy environments. Specifically, the system integrates continuous visual odometry computation with tracking-by-detection in order to track pedestrians in spite of frequent occlusions and egomotion of the camera rig. To achieve reliable performance under real-world conditions, it has long been advocated to extract and combine as much visual in-formation as possible. We propose a way to closely integrate the vision modules for visual odometry, pedestrian detection, depth estimation, and tracking. The integration naturally leads to several cognitive feedback loops between the modules. Among others, we propose a novel feedback connection from the object detector to visual odometry which utilizes the semantic knowledge of detection to stabilize localization. Feedback loops always carry the danger that erroneous feedback from one module is amplified and causes the entire system to become instable. We therefore incorporate automatic failure detection and recovery, allowing the system to continue when a module becomes unreliable.The approach is experimentally evaluated on several long and difficult video sequences from busy inner-city locations.Our results show that the proposed integration makes it possible to deliver stable tracking performance in scenes of previously infeasible complexity.

Full Text : Link

Lab Meeting September 29, 2008 (Any): SCAPE: Shape Completion and Animation of People

D. Anguelov, P.Srinivasan, D.Koller, S.Thrun, J. Rodgers, J.Davis. SCAPE: Shape Completion and Animation of People. Proceedings of the SIGGRAPH Conference, 2005.

Abstract—We introduce the SCAPE method (Shape Completion and Animation for PEople)—a data-driven method for building a human shape model that spans variation in both subject shape and pose. The method is based on a representation that incorporates both articulated and non-rigid deformations. We learn a pose deformation model that derives the non-rigid surface deformation as a function of the pose of the articulated skeleton. We also learn a separate model of variation based on body shape. Our two models can be combined to produce 3D surface models with realistic muscle deformation for different people in different poses, when neither appear in the training set. We show how the model can be used for shape completion — generating a complete surface mesh given a limited set of markers specifying the target shape. We present applications of shape completion to partial view completion and motion capture animation. In particular, our method is capable of constructing a high-quality animated surface model of a moving person, with realistic muscle deformation, using just a single static scan and a marker motion capture sequence of the person.

Full text

SCAPE project page

Lab Meeting September 22, 2008 (Yu-chun): Human Adaptation to a Miniature Robot: Precursors of Mutual Adaptation

Yasser Mohammad and Toyoaki Nishida
Robot and Human Interactive Communication (IEEE Ro-Man 2008)

Abstract:
Mutual adaptation is an important phenomenon in human-human communications. Traditionally HRI research was more interested in investigating adaptation of the robot to the human using machine learning techniques but the possibility of utilizing the natural ability of humans to adapt to other humans and artifacts including robots is recently becoming increasingly attractive. This paper presents some of the results from an experiment conducted to investigate the interaction patterns and effectiveness of motion cues as a feedback modality between a human operator and a miniature robot in a confined collaborative navigation task. The results presented in this paper show evidence of human adaptation to the robot and moreover suggest that the adaptation rate is not constant or continuous in time but is discontinuous and nonlinear. The results also show evidence of a starting exploration stage before the adaptation with duration dependent on the expectations of the human regarding the capabilities of the robot in the given task. The paper investigates how to utilize these and related findings for building robots not only capable of adapting to human operators but can also help those operators adapt to them.

[Link]

Lab Meeting September 22,2008(Shao-Chen): Closing the Loop in Scene Interpretation

Title: Closing the Loop in Scene Interpretation
Authors: D. Hoiem, A. A. Efros, and M. Hebert.

Abstract:
Image understanding involves analyzing many different aspects of the scene. In this paper, we are concerned with how these tasks can be combined in a way that improves the performance of each of them. Inspired by Barrow and Tenenbaum, we present a flexible framework for interfacing scene analysis processes using intrinsic images. Each intrinsic image is a registered map describing one characteristic of the scene. We apply this framework to develop an integrated 3D scene understanding system with estimates of surface orientations, occlusion boundaries, objects, camera viewpoint, and relative depth. Our experiments on a set of 300 outdoor images demonstrate that these tasks reinforce each other, and we illustrate a coherent scene understanding with automatically reconstructed 3D models.

link

CMU ECE talk:From Single Images To Camera Networks: Modeling and Inference Strategies

From Single Images To Camera Networks: Modeling and Inference Strategies

Amit Roy-Chowdhury
UC Riverside
Sep 19 2008

Abstract:
The complexity of vision systems can be represented along many parameters, one of them being the amount of data that is processed. On one end of this spectrum is a single image, while on the other end is a large camera network. In this talk, I will focus on these two ends of the spectrum, analyze their unique requirements and inter-relationships. In the first part, we will discuss mathematical models of image appearance. In my research, I have tried to address the question on how valid are some of the commonly used models, like linear, bilinear, multilinear, locally linear. Given the physical laws of object motion, surface properties and image formation, can we derive some of these models from first principles? We will see that, under certain mathematical assumptions, we can indeed derive some of these models and that this analysis provides new insights into problems of tracking and recognition. In the second part of the talk, I will discuss our current work on scene analysis in camera networks. I will first describe a multi-objective optimization framework that is able to hold tracks of multiple targets over space and time by adapting between delay and accuracy requirements. Then, I will describe our recent work on cooperative control of a camera network using game theory. The importance of a good understanding of the properties of single images in analyzing data over a camera network will be highlighted.

Bio:
Amit K. Roy-Chowdhury is an Assistant Professor of Electrical Engineering and a Cooperating Faculty in the Dept. of Computer Science at the University of California, Riverside. He completed his PhD in 2002 from the University of Maryland, College Park, where he also worked as a Research Associate in 2003. Previous to that, he received his Masters in Systems Science and Automation from the Indian Institute of Science, Bangalore. His research interests are in the broad areas of image processing and analysis, computer vision, video communications and statistical methods for signal processing, pattern recognition and machine learning. His current research projects include network-centric scene analysis in camera networks, physics-based mathematical modeling of image appearance, activity modeling and recognition, face and gait recognition, biological video analysis and distributed video compression. Dr. Roy-Chowdhury has over seventy papers in peer-reviewed journals, conferences and edited books. He is an author of the book titled "Recognition of Humans and Their Activities Using Video". He is an Associate Editor of the IAPR journal Machine Vision and Applications and is regular reviewer of the major journals and conference proceedings in his area.

CMU talk: Kernelized Sorting

Speaker: Le Song
Title: Kernelized Sorting
Date: Monday September 22

Abstract:
Matching pairs of objects is a fundamental operation of unsupervised learning. For instance, we might want to match a photo with a textual description of a person. In those cases it is desirable to have a compatibility function which determines how one set may be translated into the other. For many such instances we may be able to design a compatibility score based on prior knowledge or to observe one based on the co-occurrence of such objects.

In some cases, however, such a match may not exist or it may not be given to us beforehand. That is, while we may have a good understanding of two sources, we may not understand the mapping between the two spaces. For instance, we might have two collections of documents purportedly covering the same content, written in two different languages. Can we determine the correspondence between these two sets of documents without using a dictionary?

We will present a method which is able to perform such matching WITHOUT the need of a cross-domain similarity measure and we shall show that if such measures exist it generalizes normal sorting. Our method relies on the fact that one may estimate the dependence between sets of random variables even without knowing the cross-domain mapping. Various criteria are available. We choose the Hilbert Schmidt Independence Criterion between two sets and we maximize over the permutation group to find a good match. As a side-effect we obtain an explicit representation of the covariance.

We will demonstrate this kernelized sorting using various examples, including image layout, image matching, data attribute matching and multilingual document matching.

Special Google/VASC Seminar:

Check this out if you want to see what Google is thinking about for computer vision research.

Jay Yagnik,
Head of Computer Vision and Audio Understanding Research
Google, Inc.

When: Wednesday, September 17, 12:00 p.m.

Abstract:
In most recognition / retrieval problems dealing with large image/video datasets users are often looking for searching / browsing around semantics of the data. Standard computer vision algorithms deal with features extracted from pixels and attempt to perform a mapping to predict semantics. Approaches looking at just pixels are inherently limited in this regard and give rise to what we call the "semantic gap", i.e the disconnect between the semantic concepts natural to users and the pixel based predictions. One possible solution here is to rely on the large collection of public web pages where we have images and surrounding text that is potentially relevant to the inherent semantics of the image. I'll present a special case of this class of solutions around learning to recognize people. Named entity recognition style text parsing can give us hints from the text about what phrases might be people names. Retaining all the possible associations between names and faces would give us a very weak training set. I'll talk about a machine learning formulation that we refer to as consistency learning, that can effectively train models from such weak training sets and use them for robust recognition. The training procedure is inherently parallel and scales to really large sets. We verify this by large scale experiments with more than 86M face models involved for more than 200K people.

Bio: Jay Yagnik is Head of Computer Vision and Audio Understanding Research at Google Inc. His interests include machine learning, scalable matching, graph information propagation, image representation and recognition, temporal information mining, statistics. He is an alumni of the Indian Institute of Science and Nirma Institute of Technology. Prior to Google he worked on criminal identification through beard-mustache invariant facial recognition, machine learning for predicting protein function and more at the Super Education and Research Center at IISc Bangalore.

Lab Meeting September 15th, 2008 (swem): Face Alignment via Boosted Ranking Model

Title: Face Alignment via Boosted Ranking Model
Author: Hao Wu, Xiaoming Liu and Gianfranco Doretto

Abstract:

Face alignment seeks to deform a face model to match it with the features of the image of a face by optimizing an appropriate cost function. We propose a new face model that is aligned by maximizing a score function, which we learn from training data, and that we impose to be concave. We show that this problem can be reduced to learning a classifier that is able to say whether or not by switching from one alignment to a new one, the model is approaching the correct fitting. This relates to the ranking problem where a number of instances need to be ordered. For training the model, we propose to extend GentleBoost [23] to ranklearning. Extensive experimentation shows the superiority of this approach to other learning paradigms, and demonstrates that this model exceeds the alignment performance of the state-of-the-art.

Link

Lab Meeting September 15th, 2008 (Alan): Pedestrian Detection in Crowed Scenes

Title: Pedestrian Detection in Crowed Scenes
Authors: B. Leibe, E. Seemann, and B. Schiele.
in IEEE Conference on Computer Vision and Pattern Recognition (CVPR'05), San Diego, CA, pp. 878-885, 2005

Abstract:
In this paper, we address the problem of detecting pedestrians in crowded real-world scenes with severe overlaps. Our basic premise is that this problem is too difficult for any type of model or feature alone. Instead, we present a novel algorithm that integrates evidence in multiple iterations and from different sources. The core part of our method is the combination of local and global cues via a probabilistic top-down segmentation. Altogether, this approach allows to examine and compare object hypotheses with high precision down to the pixel level. Qualitative and quantitative results on a large data set confirm that our method is able to reliably detect pedestrians in crowded scenes, even when they overlap and partially occlude each other. In addition, the flexible nature of our approach allows it to operate on very small training sets.

Full Text: Link

IROS Workshop: Robotics Challenges for Machine Learning II

IROS Workshop: Robotics Challenges for Machine Learning II
http://www.learning-robots.de/TC/IROS2008

Objectives and Topics
There is an increasing interest in machine learning and statistics within the robotics community. At the same time, there has been a growth in the learning community in using robots as motivating applications for new algorithms and formalisms. Rapid progress requires researchers from both disciplines to come together and agree on the challenges, problem formulations, and solution techniques. Specific themes of the workshop include:

• learning models of robots, tasks or environments
  
• learning plans and control policies by imitation and reinforcement learning
  
• representations which facilitate learning, such as low-dimensional embeddings of movements
  
• learning representations and task abstractions by unsupervised learning
  
• probabilistic inference of task parameters from multi-modal sensory information
  
• integration of learning into control architectures.
  

This workshop will also serve to kick-off the new IEEE Technical Committee (TC) on Robot Learning.

This workshop is directly related to what we are doing now. Check out the related papers.  -Bob

CfP: Autonomous Robots - Special Issue on Robot Learning

===================================================

Call For Papers: Autonomous Robots - Special Issue on Robot Learning
===================================================
Quick Facts
=========
Editors: Jan Peters, Max Planck Institute for Biological Cybernetics,
               Andrew Y. Ng, Stanford University
Journal: Autonomous Robots
Submission Deadline: November 8, 2008
Author Notification: March 1, 2009
Revised Manuscripts: June 1, 2009
Approximate Publication Date: 4th Quarter, 2009

Abstract
======
Creating autonomous robots that can learn to act in unpredictable environments has been a long standing goal of robotics, artificial intelligence, and the cognitive sciences. In contrast, current commercially available industrial and service robots mostly execute fixed tasks and exhibit little adaptability. To bridge this gap, machine learning offers a myriad set of methods some of which have already been applied with great success to robotics problems. Machine learning is also likely play an increasingly important role in robotics as we take robots out of research labs and factory floors, into the unstructured environments inhabited by humans and into other
natural environments.

To carry out increasingly difficult and diverse sets of tasks, future robots will need to make proper use of perceptual stimuli such as vision, lidar, proprioceptive sensing and tactile feedback, and translate these into appropriate motor commands. In order to close this complex loop from perception to action, machine learning will be needed in various stages such as scene understanding, sensory-based action generation, high-level plan generation, and torque level motor control. Among the important problems hidden in these steps are robotic perception, perceptuo-action coupling, imitation learning, movement decomposition, probabilistic planning, motor primitive learning, reinforcement learning, model learning, motor control, and
many others.

Driven by high-profile competitions such as RoboCup and the DARPA Challenges, as well as the growing number of robot learning research programs funded by governments around the world (e.g., FP7-ICT, the euCognition initiative, DARPA Legged Locomotion and LAGR programs), interest in robot learning has reached an unprecedented high point. The interest in machine learning and statistics within robotics has increased substantially; and, robot applications have also become important for motivating new algorithms and formalisms in the machine learning community.

In this Autonomous Robots Special Issue on Robot Learning, we intend to outline recent successes in the application of domain-driven machine learning methods to robotics. Examples of topics of interest include, but are not limited to:

• learning models of robots, task or environments

• learning deep hierarchies or levels of representations from sensor & motor representations to task abstractions
• learning plans and control policies by imitation, apprenticeship and reinforcement learning
• finding low-dimensional embeddings of movement as implicit generative models
• integrating learning with control architectures
• methods for probabilistic inference from multi-modal sensory information (e.g., proprioceptive, tactile, vision)
• structured spatio-temporal representations designed for robot learning
• probabilistic inference in non-linear, non-Gaussian stochastic systems (e.g., for planning as well as for optimal or adaptive control)

From several recent workshops, it has become apparent that there is a significant body of novel work on these topics. The special issue will only focus on high quality articles based on sound theoretical development as well as evaluations on real robot systems.

Lab Meeting September 8th, 2008 (slyfox): Improving Localization Robustness in Monocular SLAM Using a High-Speed Camera

Title: Improving Localization Robustness in Monocular SLAM Using a High-Speed Camera

Authors: Peter Gemeiner, Andrew J. Davison, and Markus Vincze

Abstract:

In the robotics community localization and mapping of an unknown environment is a well-studied problem. To solve this problem in real-time using visual input, a standard
monocular Simultaneous Localization and Mapping (SLAM) algorithm can be used. This algorithm is very stable when smooth motion is expected, but in case of erratic or sudden movements, the camera pose typically gets lost. To improve robustness in Monocular SLAM (MonoSLAM) we propose to use a camera with faster readout speed to obtain a frame rate of 200Hz. We further present an extended MonoSLAM motion model, which can handle movements with significant jitter. In this work the improved localization and mapping have been evaluated against ground truth, which is reconstructed from off-line vision. To explain the benefits of using a high frame rate vision input in MonoSLAM framework, we performed epeatable experiments with a high-speed camera mounted onto a robotic arm. Due to the dense visual information MonoSLAM can faster shrink localization and mapping uncertainties and can operate under fast, erratic, or sudden movements. The extended motion model can provide additional robustness against significant handheld jitter when throwing or shaking the camera.

Link:
RSS2008 Paper
http://www.doc.ic.ac.uk/~ajd/Publications/gemeiner_etal_rss2008.pdf
http://www.doc.ic.ac.uk/~ajd/Movies/RSS2008_video_0071_high.avi

CMU RI Seminar: Engineering Self-Organizing Systems

Joint Intelligence Seminar (http://www.cs.cmu.edu/~iseminar) /
Robotics Institute Seminar

September 12, 2008

Title: Engineering Self-Organizing Systems

Radhika Nagpal, Harvard University

Biological systems, from embryos to ant colonies, achieve tremendous mileage by using vast numbers of cheap and unreliable components to achieve complex goals reliably. We are rapidly building embedded systems with similar characteristics, from self-assembling modular robots to vast sensor networks. How do we engineer robust collective behavior?

In this talk, I will describe two projects from my group where we have used inspiration from nature, both cells and social insects, to design decentralized algorithms for programmable self-assembly. In the first project, we use insights from social insects to design algorithms for collective construction by simple mobile robots. In the second project we use insights from multicellular tissues to design a modular robot that can form complex environmentally-adaptive shapes. In both cases we can achieve "global-to-local compilation": the agents rely on simple and local interactions that provably self-organize a wide class of user-specified global goals. Finally, time permitting, I will show an example of "local-to-global" phenomena that happens in real tissue self-assembly.


Bio:
Radhika Nagpal is an Assistant Professor of Computer Science at Harvard University since 2004. She received her PhD degree in Computer Science from MIT, and spent a year as a research fellow at Harvard Medical School. She is a recipient of the 2005 Microsoft New Faculty Fellowship award and the 2007 NSF Career award. Her research interests are biologically-inspired engineering principles for multi-agent systems and modelling multicellular biology.

Her student, Chih-Han Yu, gave a talk here on July 7, 2008. -Bob

CMU VASC Seminar: Metric Learning for Image Alignment and Classification

VASC Seminar
Monday, September 8, 2008

Metric Learning for Image Alignment and Classification
Minh Hoai Nguyen
Robotics Institute, Carnegie Mellon University

Abstract:

What constitutes good metrics to encode and compare? This talk will address this fundamental question that concerns computer vision scientists. We will show how to learn metrics that are optimal for image alignment with Active Appearance Models (AAMs), and image classification using Support Vector Machines (SVMs). Traditionally, feature extraction/selection and metric learning methods have been inferred independently of model estimation (e.g. SVM, AAM). Independently learning features and model parameters may result in the loss of information that is relevant to the alignment or classification process. Rather, we propose a convex framework for jointly learning image metrics and model parameters. To illustrate the benefits of our approach, this talk is divided in two parts. In the first part, we will discuss the problem of learning image metrics to avoid local minima in template alignment and AAMs. We learn a cost function that explicitly optimizes the occurrence of local minima at and only at the places corresponding to the correct alignment parameters. In the second part of the talk, we will consider the problem of building a fast classifier for facial feature detection. We will show how to jointly learn SVM parameters together with a subset of the pixels that are relevant for classification. This work is done in collaboration with Joan Perez and Fernando De la Torre.


Bio:

Minh Hoai Nguyen received his B.E. in Software Engineering from University of New South Wales, Australia in 2005. He has been a Ph.D. student in Carnegie Mellon University's Robotics Institute since 2006 and is advised by Fernando de la Torre. His research interests are in the area of computer vision and machine learning, especially at the intersection of the two. He is particularly interested in using data-driven techniques to learn representations of images (e.g. pixel selection, non-linear pixel combination) that are optimal for classification, clustering, visual tracking, and modeling.

Swem: add this to your reading list. -Bob

Lab Meeting September 8th, 2008 (fish60): High Performance Outdoor Navigation from Overhead Data using Imitation Learning

David Silver, J. Andrew Bagnell, Anthony StentzRobotics Institute, Carnegie Mellon UniversityPittsburgh, Pennsylvania USA
Robotics Science and Systems, June, 2008

High performance, long-distance autonomous navigationis a central problem for field robotics. Recently, a class of machine learning techniques have been developed that rely upon expert human demonstration to develop a function mapping overhead data to traversal cost. In this work, we extend these methods to automate interpretation of overhead data. We address key challenges, including interpolation-based planners, non-linear approximation techniques, and imperfect expert demonstration, necessary to apply these methods for learning to search for effective terrain interpretations.

Link

CFP Abstract for "Experimental Design for Real-World Systems

CALL FOR PAPERS

Experimental Design for Real-World Systems

AAAI Spring 2009 Symposium, March 23-25, Palo Alto, CA

Submission deadline: October 3, 2008

As more artificial intelligence (AI) research is fielded in real-world applications, the evaluation of systems designed for human-machine interaction becomes critical. AI research often intersects with other areas of study, including human-robot interaction, human-computer interaction, assistive technology, and ethics. Designing experiments to test hypotheses at the intersections of multiple research fields can be incredibly challenging. Many commonalities and differences already exist in experimental design for real-world systems. For example, the fields of human-robot interaction and human-computer interaction are two fields that have both joint and discrete goals. They look to evaluate very different aspects of design, interface, and interaction. In some instances, these two fields can share aspects of experimental design, while, in others, the experimental design must be fundamentally different.

We will provide a forum for researchers from many disciplines to discuss experiment design and the evaluation of real-world systems. We invite researchers from all applicable fields of human-machine interaction. We also invite researchers from allied fields, such as psychology, anthropology, design, human-computer interaction, human-robot interaction, rehabilitation and clinical care, assistive technology, and other related disciplines.

This symposium will focus on a wide variety of topics that address the challenges of experiment design for real-world systems including:
* the design of system evaluations,
* successes and failures in system evaluations,
* survey design for user studies,
* understanding the role technology plays in society,
* ethics of human subject studies,
* evaluating the use of machines as interventions,
* the uses of quantitative and qualitative data,
* and other related topics.

Format and Submissions

We will have a mix of plenary speakers, short presentations, and break-out groups. We will also have a poster session. Short presentations and posters are invited to submit an abstract (< 3 pages) on experiments conducted during their research, focused on the experimental methodology, especially those with unusual and effective methodologies. Submission formatting details at http://robotics.usc.edu/~dfseifer/aaai-expdesign/. Email submissions to aaai-sss-2009@cs.uml.edu.

Important Dates

* Call for Papers Due: October 3, 2008
* Authors Notified: November 2008
* Camera Ready Due: January 9, 2009

Organizing Committee

David Feil-Seifer (USC), Heidy Maldonado (Stanford), Bilge Mutlu (CMU), Leila Takayama (Stanford), Katherine Tsui (UMass Lowell)

Program Committee

Jenny Burke (USF), Kerstin Dautenhahn (Hertfordshire), Gert Jan Gelderbloom (VILANS), Maja Mataric (USC), Aaron Steinfeld (CMU), Holly Yanco (UMass Lowell)

Lab Meeting September 8th, 2008 (Jeff):On Handling Uncertainty in the Fundamental Matrix for Scene and Motion Adaptive Pose Recovery

Title: On Handling Uncertainty in the Fundamental Matrix for Scene and Motion Adaptive Pose Recovery

Authors: Sreenivas R. Sukumar, Hamparsum Bozdogan, David L. Page, Andreas F. Koschan and Mongi A. Abidi

Abstract:

The estimation of the fundamental matrix is the key step in feature-based camera ego-motion estimation for applications in scene modeling and vehicle navigation. In this paper, we present a new method of analyzing and further reducing the risk in the fundamental matrix due to
the choice of a particular feature detector, the choice of the matching algorithm, the motion model, iterative hypothesis generation and verification paradigms. Our scheme makes use of model-selection theory to guide the switch to optimal methods for fundamental matrix
estimation within the hypothesis-and-test architecture. We demonstrate our proposed method for vision-based robot localization in large-scale environments where the environment is constantly changing and navigation within the environment is unpredictable.

Link:
CVPR2008 Paper
http://imaging.utk.edu/publications/papers/2008/CVPR08_ss.pdf

CMU Intelligence Seminar: Predicting Neural Representations of Word Meanings

Intelligence Seminar (http://www.cs.cmu.edu/~iseminar)

September 9, 2008

Title: Predicting Neural Representations of Word Meanings

Tom M. Mitchell
E. Fredkin Professor
Machine Learning Department
Carnegie Mellon University

How does the human brain represent meanings of words and pictures in terms of the neural activity observable through fMRI brain imaging? This talk will present our research using machine learning methods to study this question. One line of our research has involved training classifiers that identify which word a person is thinking about, based on the image of their fMRI brain activity. A more recent line involves developing a generative computational model capable of predicting the neural activity associated with arbitrary English words, including words for which we do not yet have brain image data. This computational model was trained using a combination of fMRI data associated with several dozen concrete nouns, together with statistics gathered from a trillion-word text corpus. Once trained, the model predicts fMRI activation for any other concrete noun appearing in the tera-word text corpus, with highly significant accuracies over the 60 nouns for which we currently have fMRI data.

This work is based on a collaboration with a number of researchers, including my primary collaborator Marcel Just.

Bio:
Tom M. Mitchell is the E. Fredkin Professor and head of the Machine Learning Department at Carnegie Mellon University. Mitchell is a past President of the American Association of Artificial Intelligence (AAAI), past Chair of the American Association for the Advancement of Science (AAAS) section on Information, Computing and Communication, and a recent member of the US National Research Council's Computer Science and Telecommunications Board. His general research interests lie in machine learning, artificial intelligence, and cognitive neuroscience. Mitchell believes the field of machine learning will be the fastest growing branch of computer science during the 21st century.

CMU thesis defense: Dynamics of Large Networks

Title: Dynamics of Large Networks
Speaker: Jure Leskovec
Date: September 3, 2008


Abstract:
A basic premise behind the study of large networks is that interaction leads to complex collective behavior. In our work we found interesting and counterintuitive patterns for time evolving networks, which change some of the basic assumptions that were made in the past. We then develop network models, fit such models to real networks, and use them to generate realistic graphs or give formal explanations about their properties.

Another important aspect of our research is the study information diffusion and the spread of influence in a large person-to-person product recommendation network and its effect on purchases. We also model the propagation of information on the blogosphere, and propose algorithms to efficiently find influential nodes in the network.

A central topic of our thesis is also the analysis of large datasets as certain network properties only emerge and thus become visible when dealing with lots of data. We analyze the world's social and communication network of Microsoft Instant Messenger with 240 million people and 255 billion conversations. We also made interesting and counterintuitive observations about network community structure that suggest that only small network clusters exist, and that they merge and vanish as they grow.

To view a draft of the thesis see:
http://www.cs.cmu.edu/~jure/pubs/thesis/jure-thesis.pdf

COMMITTEE:
Christos Faloutsos
Avim Blum
Jon Lafferty
Jon Kleinberg

Lab Meeting September 1, 2008 (Casey): Prior Data and Kernel Conditional Random Fields for Obstacle Detection

Authors: Carlos Vallespi, Anthony (Tony) Stentz
Robotics: Science and Systems 2008

Abstract—We consider the task of training an obstacle detection(OD) system based on a monocular color camera usingminimal supervision. We train it to match the performance of asystem that uses a laser rangefinder to estimate the presenceof obstacles by size and shape. However, the lack of rangedata in the image cannot be compensated by the extraction oflocal features alone. Thus, we investigate contextual techniquesbased on Conditional Random Fields (CRFs) that can exploitthe global context of the image, and we compare them to aconventional learning approach. Furthermore, we describe aprocedure for introducing prior data in the OD system to increaseits performance in “familiar” terrains. Finally, we performexperiments using sequences of images taken from a vehicle for autonomous vehicle navigation applications.

[Paper Link]