Header logo is


2010


no image
Robust nonparametric detection of objects in noisy images

Langovoy, M., Wittich, O.

(2010-049), EURANDOM, Technische Universiteit Eindhoven, September 2010 (techreport)

Abstract
We propose a novel statistical hypothesis testing method for detection of objects in noisy images. The method uses results from percolation theory and random graph theory. We present an algorithm that allows to detect objects of unknown shapes in the presence of nonparametric noise of unknown level and of unknown distribution. No boundary shape constraints are imposed on the object, only a weak bulk condition for the object's interior is required. The algorithm has linear complexity and exponential accuracy and is appropriate for real-time systems. In this paper, we develop further the mathematical formalism of our method and explore im- portant connections to the mathematical theory of percolation and statistical physics. We prove results on consistency and algorithmic complexity of our testing procedure. In addition, we address not only an asymptotic behavior of the method, but also a nite sample performance of our test.

ei

PDF [BibTex]

2010


PDF [BibTex]


no image
Statistical image analysis and percolation theory

Davies, P., Langovoy, M., Wittich, O.

73rd Annual Meeting of the Institute of Mathematical Statistics (IMS), August 2010 (talk)

Abstract
We develop a novel method for detection of signals and reconstruction of images in the presence of random noise. The method uses results from percolation theory. We specifically address the problem of detection of objects of unknown shapes in the case of nonparametric noise. The noise density is unknown and can be heavy-tailed. We view the object detection problem as hypothesis testing for discrete statistical inverse problems. We present an algorithm that allows to detect objects of various shapes in noisy images. We prove results on consistency and algorithmic complexity of our procedures.

ei

Web [BibTex]

Web [BibTex]


no image
Large Scale Variational Inference and Experimental Design for Sparse Generalized Linear Models

Seeger, M., Nickisch, H.

Max Planck Institute for Biological Cybernetics, August 2010 (techreport)

Abstract
Many problems of low-level computer vision and image processing, such as denoising, deconvolution, tomographic reconstruction or super-resolution, can be addressed by maximizing the posterior distribution of a sparse linear model (SLM). We show how higher-order Bayesian decision-making problems, such as optimizing image acquisition in magnetic resonance scanners, can be addressed by querying the SLM posterior covariance, unrelated to the density's mode. We propose a scalable algorithmic framework, with which SLM posteriors over full, high-resolution images can be approximated for the first time, solving a variational optimization problem which is convex iff posterior mode finding is convex. These methods successfully drive the optimization of sampling trajectories for real-world magnetic resonance imaging through Bayesian experimental design, which has not been attempted before. Our methodology provides new insight into similarities and differences between sparse reconstruction and approximate Bayesian inference, and has important implications for compressive sensing of real-world images.

ei

Web [BibTex]


no image
Cooperative Cuts for Image Segmentation

Jegelka, S., Bilmes, J.

(UWEETR-1020-0003), University of Washington, Washington DC, USA, August 2010 (techreport)

Abstract
We propose a novel framework for graph-based cooperative regularization that uses submodular costs on graph edges. We introduce an efficient iterative algorithm to solve the resulting hard discrete optimization problem, and show that it has a guaranteed approximation factor. The edge-submodular formulation is amenable to the same extensions as standard graph cut approaches, and applicable to a range of problems. We apply this method to the image segmentation problem. Specifically, Here, we apply it to introduce a discount for homogeneous boundaries in binary image segmentation on very difficult images, precisely, long thin objects and color and grayscale images with a shading gradient. The experiments show that significant portions of previously truncated objects are now preserved.

ei

Web [BibTex]

Web [BibTex]


no image
Statistical image analysis and percolation theory

Langovoy, M., Wittich, O.

28th European Meeting of Statisticians (EMS), August 2010 (talk)

ei

PDF Web [BibTex]

PDF Web [BibTex]


no image
Fast algorithms for total-variationbased optimization

Barbero, A., Sra, S.

(194), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, August 2010 (techreport)

Abstract
We derive a number of methods to solve efficiently simple optimization problems subject to a totalvariation (TV) regularization, under different norms of the TV operator and both for the case of 1-dimensional and 2-dimensional data. In spite of the non-smooth, non-separable nature of the TV terms considered, we show that a dual formulation with strong structure can be derived. Taking advantage of this structure we develop adaptions of existing algorithms from the optimization literature, resulting in efficient methods for the problem at hand. Experimental results show that for 1-dimensional data the proposed methods achieve convergence within good accuracy levels in practically linear time, both for L1 and L2 norms. For the more challenging 2-dimensional case a performance of order O(N2 log2 N) for N x N inputs is achieved when using the L2 norm. A final section suggests possible extensions and lines of further work.

ei

PDF [BibTex]

PDF [BibTex]


no image
Reinforcement Learning by Relative Entropy Policy Search

Peters, J., Mülling, K., Altun, Y.

30th International Workshop on Bayesian Inference and Maximum Entropy Methods in Science and Engineering (MaxEnt 2010), 30, pages: 69, July 2010 (poster)

Abstract
Policy search is a successful approach to reinforcement learning. However, policy improvements often result in the loss of information. Hence, it has been marred by premature convergence and implausible solutions. As first suggested in the context of covariant policy gradients, many of these problems may be addressed by constraining the information loss. In this book chapter, we continue this path of reasoning and suggest the Relative Entropy Policy Search (REPS) method. The resulting method differs significantly from previous policy gradient approaches and yields an exact update step. It works well on typical reinforcement learning benchmark problems. We will also present a real-world applications where a robot employs REPS to learn how to return balls in a game of table tennis.

ei

PDF [BibTex]

PDF [BibTex]


no image
Cooperative Cuts: Graph Cuts with Submodular Edge Weights

Jegelka, S., Bilmes, J.

24th European Conference on Operational Research (EURO XXIV), July 2010 (talk)

Abstract
We introduce cooperative cut, a minimum cut problem whose cost is a submodular function on sets of edges: the cost of an edge that is added to a cut set depends on the edges in the set. Applications are e.g. in probabilistic graphical models and image processing. We prove NP hardness and a polynomial lower bound on the approximation factor, and upper bounds via four approximation algorithms based on different techniques. Our additional heuristics have attractive practical properties, e.g., to rely only on standard min-cut. Both our algorithms and heuristics appear to do well in practice.

ei

PDF Web [BibTex]

PDF Web [BibTex]


no image
A Maximum Entropy Approach to Semi-supervised Learning

Erkan, A., Altun, Y.

30th International Workshop on Bayesian Inference and Maximum Entropy Methods in Science and Engineering (MaxEnt 2010), 30, pages: 80, July 2010 (poster)

Abstract
Maximum entropy (MaxEnt) framework has been studied extensively in supervised learning. Here, the goal is to find a distribution p that maximizes an entropy function while enforcing data constraints so that the expected values of some (pre-defined) features with respect to p match their empirical counterparts approximately. Using different entropy measures, different model spaces for p and different approximation criteria for the data constraints yields a family of discriminative supervised learning methods (e.g., logistic regression, conditional random fields, least squares and boosting). This framework is known as the generalized maximum entropy framework. Semi-supervised learning (SSL) has emerged in the last decade as a promising field that combines unlabeled data along with labeled data so as to increase the accuracy and robustness of inference algorithms. However, most SSL algorithms to date have had trade-offs, e.g., in terms of scalability or applicability to multi-categorical data. We extend the generalized MaxEnt framework to develop a family of novel SSL algorithms. Extensive empirical evaluation on benchmark data sets that are widely used in the literature demonstrates the validity and competitiveness of the proposed algorithms.

ei

PDF PDF [BibTex]

PDF PDF [BibTex]


no image
The effect of positioning aids on PET quantification following MR-based attenuation correction (AC) in PET/MR imaging

Mantlik, F., Hofmann, M., Kupferschläger, J., Werner, M., Pichler, B., Beyer, T.

Journal of Nuclear Medicine, 51(Supplement 2):1418 , June 2010 (poster)

Abstract
Objectives: We study the quantitative effect of not accounting for the attenuation of patient positioning aids in combined PET/MR imaging. Methods: Positioning aids cannot be detected with conventional MR sequences. We mimic this effect using PET/CT data (Biograph HiRez16) with the foams removed from CT images prior to using them for CT-AC. PET/CT data were acquired using standard parameters (phantoms/patients): 120/140 kVp, 30/250 mAs, 5 mm slices, OSEM (4i, 8s, 5 mm filter) following CT-AC. First, a uniform 68Ge-cylinder was positioned centrally in the PET/CT and fixed with a vacuum mattress (10 cm thick). Second, the same cylinder was placed in 3 positioning aids from the PET/MR (BrainPET-3T). Third, 5 head/neck patients who were fixed in a vacuum mattress were selected. In all 3 studies PET recon post CT-AC based on measured CT images was used as the reference (mCT-AC). The PET/MR set-up was mimicked by segmenting the foam inserts from the measured CT images and setting their voxel values to -1000 HU (air). PET images were reconstructed using CT-AC with the segmented CT images (sCT-AC). PET images with mCT- and sCT-AC were compared. Results: sCT-AC underestimated PET voxel values in the phantom by 6.7% on average compared to mCT-AC with the vacuum mattress in place. 5% of the PET voxels were underestimated by >=10%. Not accounting for MR positioning aids during AC led to an underestimation of 2.8% following sCT-AC, with 5% of the PET voxels being underestimated by >=7% wrt mCT-AC. Preliminary evaluation of the patient data indicates a slightly higher bias from not accounting for patient positioning aids (mean: -9.1%, 5% percentile: -11.2%). Conclusions: A considerable and regionally variable underestimation of the PET activity following AC is observed when positioning aids are not accounted for. This bias may become relevant in neurological activation or dementia studies with PET/MR

ei

Web [BibTex]

Web [BibTex]


no image
Multi-task Learning for Zero Training Brain-Computer Interfaces

Alamgir, M., Grosse-Wentrup, M., Altun, Y.

4th International BCI Meeting, June 2010 (poster)

Abstract
Brain-computer interfaces (BCIs) are limited in their applicability in everyday settings by the current necessity to record subject-specific calibration data prior to actual use of the BCI for communication. In this work, we utilize the framework of multitask learning to construct a BCI that can be used without any subject-specific calibration process, i.e., with zero training data. In BCIs based on EEG or MEG, the predictive function of a subject's intention is commonly modeled as a linear combination of some features derived from spatial and spectral recordings. The coefficients of this combination correspond to the importance of the features for predicting the intention of the subject. These coefficients are usually learned separately for each subject due to inter-subject variability. Principle feature characteristics, however, are known to remain invariant across subject. For example, it is well known that in motor imagery paradigms spectral power in the mu- and beta frequency ranges (roughly 8-14 Hz and 20-30 Hz, respectively) over sensorimotor areas provides most information on a subject's intention. Based on this assumption, we define the intention prediction function as a combination of subject-invariant and subject-specific models, and propose a machine learning method that infers these models jointly using data from multiple subjects. This framework leads to an out-of-the-box intention predictor, where the subject-invariant model can be employed immediately for a subject with no prior data. We present a computationally efficient method to further improve this BCI to incorporate subject-specific variations as such data becomes available. To overcome the problem of high dimensional feature spaces in this context, we further present a new method for finding the relevance of different recording channels according to actions performed by subjects. Usually, the BCI feature representation is a concatenation of spectral features extracted from different channels. This representation, however, is redundant, as recording channels at different spatial locations typically measure overlapping sources within the brain due to volume conduction. We address this problem by assuming that the relevance of different spectral bands is invariant across channels, while learning different weights for each recording electrode. This framework allows us to significantly reduce the feature space dimensionality without discarding potentially useful information. Furthermore, the resulting out-of-the-box BCI can be adapted to different experimental setups, for example EEG caps with different numbers of channels, as long as there exists a mapping across channels in different setups. We demonstrate the feasibility of our approach on a set of experimental EEG data recorded during a standard two-class motor imagery paradigm from a total of ten healthy subjects. Specifically, we show that satisfactory classification results can be achieved with zero training data, and that combining prior recordings with subject-specific calibration data substantially outperforms using subject-specific data only.

ei

Web [BibTex]


no image
Causal Influence of Gamma Oscillations on Performance in Brain-Computer Interfaces

Grosse-Wentrup, M., Hill, J., Schölkopf, B.

4th International BCI Meeting0, June 2010 (poster)

Abstract
Background and Objective: While machine learning approaches have led to tremendous advances in brain-computer interfaces (BCIs) in recent years (cf. [1]), there still exists a large variation in performance across subjects. Furthermore, a significant proportion of subjects appears incapable of achieving above chance-level classification accuracy [2], which to date includes all subjects in a completely locked-in state that have been trained in BCI control. Understanding the reasons for this variation in performance arguably constitutes one of the most fundamental open questions in research on BCIs. Methods & Results Using a machine learning approach, we derive a trial-wise measure of how well EEG recordings can be classified as either left- or right-hand motor imagery. Specifically, we train a support vector machine (SVM) on log-bandpower features (7-40 Hz) derived from EEG channels after spatial filtering with a surface Laplacian, and then compute the trial-wise distance of the output of the SVM from the separating hyperplane using a cross-validation procedure. We then correlate this trial-wise performance measure, computed on EEG recordings of ten healthy subjects, with log-bandpower in the gamma frequency range (55-85 Hz), and demonstrate that it is positively correlated with frontal- and occipital gamma-power and negatively correlated with centro-parietal gamma-power. This correlation is shown to be highly significant on the group level as well as in six out of ten subjects on the single-subject level. We then utilize the framework for causal inference developed by Pearl, Spirtes and others [3,4] to present evidence that gamma-power is not only correlated with BCI performance but does indeed exert a causal influence on it. Discussion and Conclusions Our results indicate that successful execution of motor imagery, and hence reliable communication by means of a BCI based on motor imagery, requires a volitional shift of gamma-power from centro-parietal to frontal and occipital regions. As such, our results provide the first non-trivial explanation for the variation in BCI performance across and within subjects. As this topographical alteration in gamma-power is likely to correspond to a specific attentional shift, we propose to provide subjects with feedback on their topographical distribution of gamma-power in order to establish the attentional state required for successful execution of motor imagery.

ei

Web [BibTex]


no image
Solving large-scale nonnegative least-squares

Sra, S.

16th Conference of the International Linear Algebra Society (ILAS 2010), 16, pages: 19, June 2010, based on Joint work with Dongmin Kim and Inderjit Dhillon (poster)

Abstract
We study the fundamental problem of nonnegative least squares. This problem was apparently introduced by Lawson and Hanson [1] under the name NNLS. As is evident from its name, NNLS seeks least-squares solutions that are also nonnegative. Owing to its wide-applicability numerous algorithms have been derived for NNLS, beginning from the active-set approach of Lawson and Hanson [1] leading up to the sophisticated interior-point method of Bellavia et al. [2]. We present a new algorithm for NNLS that combines projected subgradients with the non-monotonic gradient descent idea of Barzilai and Borwein [3]. Our resulting algorithm is called BBSG, and we guarantee its convergence by exploiting properties of NNLS in conjunction with projected subgradients. BBSG is surprisingly simple and scales well to large problems. We substantiate our claims by empirically evaluating BBSG and comparing it with established convex solvers and specialized NNLS algorithms. The numerical results suggest that BBSG is a practical method for solving large-scale NNLS problems.

ei

PDF Web [BibTex]

PDF Web [BibTex]


no image
Simultaneous PET/MRI for the evaluation of hemato-oncological diseases with lower extremity manifestations

Sauter, A., Horger, M., Boss, A., Kolb, A., Mantlik, F., Kanz, L., Pfannenberg, C., Stegger, L., Claussen, C., Pichler, B.

Journal of Nuclear Medicine, 51(Supplement 2):1001 , June 2010 (poster)

Abstract
Objectives: The study purpose is the evaluation of patients, suffering from hemato-oncological disease with complications at the lower extremities, using simultaneous PET/MRI. Methods: Until now two patients (chronic active graft-versus-host-disease [GvHD], B-non Hodgkin lymphoma [B-NHL]) before and after therapy were examined in a 3-Tesla-BrainPET/MRI hybrid system following F-18-FDG-PET/CT. Simultaneous static PET (1200 sec.) and MRI scans (T1WI, T2WI, post-CA) were acquired. Results: Initial results show the feasibility of using hybrid PET/MRI-technology for musculoskeletal imaging of the lower extremities. Simultaneous PET and MRI could be acquired in diagnostic quality. Before treatment our patient with GvHD had a high fascia and muscle FDG uptake, possibly due to muscle encasement. T2WI and post gadolinium T1WI revealed a fascial thickening and signs of inflammation. After therapy with steroids followed by imatinib the patient’s symptoms improved while, the muscular FDG uptake droped whereas the MRI signal remained unchanged. We assume that fascial elasticity improved during therapy despite persistance of fascial thickening. The examination of the second patient with B-NHL manifestation in the tibia showed a significant signal and uptake decrease in the bone marrow and surrounding lesions in both, MRI and PET after therapy with rituximab. The lack of residual FDG-uptake proved superior to MRI information alone helping for exclusion of vital tumor. Conclusions: Combined PET/MRI is a powerful tool to monitor diseases requiring high soft tissue contrast along with molecular information from the FDG uptake.

ei

Web [BibTex]

Web [BibTex]


no image
Solving Large-Scale Nonnegative Least Squares

Sra, S.

16th Conference of the International Linear Algebra Society (ILAS), June 2010 (talk)

Abstract
We study the fundamental problem of nonnegative least squares. This problem was apparently introduced by Lawson and Hanson [1] under the name NNLS. As is evident from its name, NNLS seeks least-squares solutions that are also nonnegative. Owing to its wide-applicability numerous algorithms have been derived for NNLS, beginning from the active-set approach of Lawson and Han- son [1] leading up to the sophisticated interior-point method of Bellavia et al. [2]. We present a new algorithm for NNLS that combines projected subgradients with the non-monotonic gradient descent idea of Barzilai and Borwein [3]. Our resulting algorithm is called BBSG, and we guarantee its convergence by ex- ploiting properties of NNLS in conjunction with projected subgradients. BBSG is surprisingly simple and scales well to large problems. We substantiate our claims by empirically evaluating BBSG and comparing it with established con- vex solvers and specialized NNLS algorithms. The numerical results suggest that BBSG is a practical method for solving large-scale NNLS problems.

ei

PDF PDF [BibTex]

PDF PDF [BibTex]


no image
Gaussian Mixture Modeling with Gaussian Process Latent Variable Models

Nickisch, H., Rasmussen, C.

Max Planck Institute for Biological Cybernetics, June 2010 (techreport)

Abstract
Density modeling is notoriously difficult for high dimensional data. One approach to the problem is to search for a lower dimensional manifold which captures the main characteristics of the data. Recently, the Gaussian Process Latent Variable Model (GPLVM) has successfully been used to find low dimensional manifolds in a variety of complex data. The GPLVM consists of a set of points in a low dimensional latent space, and a stochastic map to the observed space. We show how it can be interpreted as a density model in the observed space. However, the GPLVM is not trained as a density model and therefore yields bad density estimates. We propose a new training strategy and obtain improved generalisation performance and better density estimates in comparative evaluations on several benchmark data sets.

ei

Web [BibTex]

Web [BibTex]


no image
Matrix Approximation Problems

Sra, S.

EU Regional School: Rheinisch-Westf{\"a}lische Technische Hochschule Aachen, May 2010 (talk)

ei

PDF AVI [BibTex]

PDF AVI [BibTex]


no image
BCI2000 and Python

Hill, NJ.

Invited lecture at the 7th International BCI2000 Workshop, Pacific Grove, CA, USA, May 2010 (talk)

Abstract
A tutorial, with exercises, on how to integrate your own Python code with the BCI2000 realtime software package.

ei

PDF [BibTex]

PDF [BibTex]


no image
Extending BCI2000 Functionality with Your Own C++ Code

Hill, NJ.

Invited lecture at the 7th International BCI2000 Workshop, Pacific Grove, CA, USA, May 2010 (talk)

Abstract
A tutorial, with exercises, on how to use BCI2000 C++ framework to write your own real-time signal-processing modules.

ei

[BibTex]

[BibTex]


no image
Generalized Proximity and Projection with Norms and Mixed-norms

Sra, S.

(192), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, May 2010 (techreport)

Abstract
We discuss generalized proximity operators (GPO) and their associated generalized projection problems. On inputs of size n, we show how to efficiently apply GPOs and generalized projections for separable norms and distance-like functions to accuracy e in O(n log(1/e)) time. We also derive projection algorithms that run theoretically in O(n log n log(1/e)) time but can for suitable parameter ranges empirically outperform the O(n log(1/e)) projection method. The proximity and projection tasks are either separable, and solved directly, or are reduced to a single root-finding step. We highlight that as a byproduct, our analysis also yields an O(n log(1/e)) (weakly linear-time) procedure for Euclidean projections onto the l1;1-norm ball; previously only an O(n log n) method was known. We provide empirical evaluation to illustrate the performance of our methods, noting that for the l1;1-norm projection, our implementation is more than two orders of magnitude faster than the previously known method.

ei

PDF [BibTex]

PDF [BibTex]


no image
Solving large-scale nonnegative least squares using an adaptive non-monotonic method

Sra, S., Kim, D., Dhillon, I.

24th European Conference on Operational Research (EURO 2010), 24, pages: 223, April 2010 (poster)

Abstract
We present an efficient algorithm for large-scale non-negative least-squares (NNLS). We solve NNLS by extending the unconstrained quadratic optimization method of Barzilai and Borwein (BB) to handle nonnegativity constraints. Our approach is simple yet efficient. It differs from other constrained BB variants as: (i) it uses a specific subset of variables for computing BB steps; and (ii) it scales these steps adaptively to ensure convergence. We compare our method with both established convex solvers and specialized NNLS methods, and observe highly competitive empirical performance.

ei

PDF [BibTex]

PDF [BibTex]


no image
Sparse regression via a trust-region proximal method

Kim, D., Sra, S., Dhillon, I.

24th European Conference on Operational Research (EURO 2010), 24, pages: 278, April 2010 (poster)

Abstract
We present a method for sparse regression problems. Our method is based on the nonsmooth trust-region framework that minimizes a sum of smooth convex functions and a nonsmooth convex regularizer. By employing a separable quadratic approximation to the smooth part, the method enables the use of proximity operators, which in turn allow tackling the nonsmooth part efficiently. We illustrate our method by implementing it for three important sparse regression problems. In experiments with synthetic and real-world large-scale data, our method is seen to be competitive, robust, and scalable.

ei

PDF [BibTex]

PDF [BibTex]


no image
Machine-Learning Methods for Decoding Intentional Brain States

Hill, NJ.

Symposium "Non-Invasive Brain Computer Interfaces: Current Developments and Applications" (BIOMAG), March 2010 (talk)

Abstract
Brain-computer interfaces (BCI) work by making the user perform a specific mental task, such as imagining moving body parts or performing some other covert mental activity, or attending to a particular stimulus out of an array of options, in order to encode their intention into a measurable brain signal. Signal-processing and machine-learning techniques are then used to decode the measured signal to identify the encoded mental state and hence extract the user‘s initial intention. The high-noise high-dimensional nature of brain-signals make robust decoding techniques a necessity. Generally, the approach has been to use relatively simple feature extraction techniques, such as template matching and band-power estimation, coupled to simple linear classifiers. This has led to a prevailing view among applied BCI researchers that (sophisticated) machine-learning is irrelevant since “it doesn‘t matter what classifier you use once your features are extracted.” Using examples from our own MEG and EEG experiments, I‘ll demonstrate how machine-learning principles can be applied in order to improve BCI performance, if they are formulated in a domain-specific way. The result is a type of data-driven analysis that is more than “just” classification, and can be used to find better feature extractors.

ei

PDF Web [BibTex]

PDF Web [BibTex]


no image
Cooperative Cuts: Graph Cuts with Submodular Edge Weights

Jegelka, S., Bilmes, J.

(189), Max Planck Institute for Biological Cybernetics, Tuebingen, Germany, March 2010 (techreport)

Abstract
We introduce a problem we call Cooperative cut, where the goal is to find a minimum-cost graph cut but where a submodular function is used to define the cost of a subsets of edges. That means, the cost of an edge that is added to the current cut set C depends on the edges in C. This generalization of the cost in the standard min-cut problem to a submodular cost function immediately makes the problem harder. Not only do we prove NP hardness even for nonnegative submodular costs, but also show a lower bound of Omega(|V|^(1/3)) on the approximation factor for the problem. On the positive side, we propose and compare four approximation algorithms with an overall approximation factor of min { |V|/2, |C*|, O( sqrt(|E|) log |V|), |P_max|}, where C* is the optimal solution, and P_max is the longest s, t path across the cut between given s, t. We also introduce additional heuristics for the problem which have attractive properties from the perspective of practical applications and implementations in that existing fast min-cut libraries may be used as subroutines. Both our approximation algorithms, and our heuristics, appear to do well in practice.

ei

PDF [BibTex]

PDF [BibTex]


no image
PAC-Bayesian Analysis in Unsupervised Learning

Seldin, Y.

Foundations and New Trends of PAC Bayesian Learning Workshop, March 2010 (talk)

ei

PDF Web [BibTex]

PDF Web [BibTex]


no image
PAC-Bayesian Bounds for Discrete Density Estimation and Co-clustering Analysis

Seldin, Y., Tishby, N.

Workshop "Foundations and New Trends of PAC Bayesian Learning", 2010, March 2010 (poster)

Abstract
We applied PAC-Bayesian framework to derive gen- eralization bounds for co-clustering1. The analysis yielded regularization terms that were absent in the preceding formulations of this task. The bounds sug- gested that co-clustering should optimize a trade-off between its empirical performance and the mutual in- formation that the cluster variables preserve on row and column indices. Proper regularization enabled us to achieve state-of-the-art results in prediction of the missing ratings in the MovieLens collaborative filtering dataset. In addition a PAC-Bayesian bound for discrete den- sity estimation was derived. We have shown that the PAC-Bayesian bound for classification is a spe- cial case of the PAC-Bayesian bound for discrete den- sity estimation. We further introduced combinatorial priors to PAC-Bayesian analysis. The combinatorial priors are more appropriate for discrete domains, as opposed to Gaussian priors, the latter of which are suitable for continuous domains. It was shown that combinatorial priors lead to regularization terms in the form of mutual information.

ei

PDF Web [BibTex]

PDF Web [BibTex]


no image
Learning Motor Primitives for Robotics

Kober, J., Peters, J.

EVENT Lab: Reinforcement Learning in Robotics and Virtual Reality, January 2010 (talk)

Abstract
The acquisition and self-improvement of novel motor skills is among the most important problems in robotics. Motor primitives offer one of the most promising frameworks for the application of machine learning techniques in this context. Employing the Dynamic Systems Motor primitives originally introduced by Ijspeert et al. (2003), appropriate learning algorithms for a concerted approach of both imitation and reinforcement learning are presented. Using these algorithms new motor skills, i.e., Ball-in-a-Cup, Ball-Paddling and Dart-Throwing, are learned.

ei

[BibTex]

[BibTex]


no image
Learning Continuous Grasp Affordances by Sensorimotor Exploration

Detry, R., Baseski, E., Popovic, M., Touati, Y., Krüger, N., Kroemer, O., Peters, J., Piater, J.

In From Motor Learning to Interaction Learning in Robots, pages: 451-465, Studies in Computational Intelligence ; 264, (Editors: Sigaud, O. and Peters, J.), Springer, Berlin, Germany, January 2010 (inbook)

Abstract
We develop means of learning and representing object grasp affordances probabilistically. By grasp affordance, we refer to an entity that is able to assess whether a given relative object-gripper configuration will yield a stable grasp. These affordances are represented with grasp densities, continuous probability density functions defined on the space of 3D positions and orientations. Grasp densities are registered with a visual model of the object they characterize. They are exploited by aligning them to a target object using visual pose estimation. Grasp densities are refined through experience: A robot “plays” with an object by executing grasps drawn randomly for the object’s grasp density. The robot then uses the outcomes of these grasps to build a richer density through an importance sampling mechanism. Initial grasp densities, called hypothesis densities, are bootstrapped from grasps collected using a motion capture system, or from grasps generated from the visual model of the object. Refined densities, called empirical densities, represent affordances that have been confirmed through physical experience. The applicability of our method is demonstrated by producing empirical densities for two object with a real robot and its 3-finger hand. Hypothesis densities are created from visual cues and human demonstration.

ei

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Imitation and Reinforcement Learning for Motor Primitives with Perceptual Coupling

Kober, J., Mohler, B., Peters, J.

In From Motor Learning to Interaction Learning in Robots, pages: 209-225, Studies in Computational Intelligence ; 264, (Editors: Sigaud, O. and Peters, J.), Springer, Berlin, Germany, January 2010 (inbook)

Abstract
Traditional motor primitive approaches deal largely with open-loop policies which can only deal with small perturbations. In this paper, we present a new type of motor primitive policies which serve as closed-loop policies together with an appropriate learning algorithm. Our new motor primitives are an augmented version version of the dynamical system-based motor primitives [Ijspeert et al(2002)Ijspeert, Nakanishi, and Schaal] that incorporates perceptual coupling to external variables. We show that these motor primitives can perform complex tasks such as Ball-in-a-Cup or Kendama task even with large variances in the initial conditions where a skilled human player would be challenged. We initialize the open-loop policies by imitation learning and the perceptual coupling with a handcrafted solution. We first improve the open-loop policies and subsequently the perceptual coupling using a novel reinforcement learning method which is particularly well-suited for dynamical system-based motor primitives.

ei

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
From Motor Learning to Interaction Learning in Robots

Sigaud, O., Peters, J.

In From Motor Learning to Interaction Learning in Robots, pages: 1-12, Studies in Computational Intelligence ; 264, (Editors: Sigaud, O. and Peters, J.), Springer, Berlin, Germany, January 2010 (inbook)

Abstract
The number of advanced robot systems has been increasing in recent years yielding a large variety of versatile designs with many degrees of freedom. These robots have the potential of being applicable in uncertain tasks outside wellstructured industrial settings. However, the complexity of both systems and tasks is often beyond the reach of classical robot programming methods. As a result, a more autonomous solution for robot task acquisition is needed where robots adaptively adjust their behaviour to the encountered situations and required tasks. Learning approaches pose one of the most appealing ways to achieve this goal. However, while learning approaches are of high importance for robotics, we cannot simply use off-the-shelf methods from the machine learning community as these usually do not scale into the domains of robotics due to excessive computational cost as well as a lack of scalability. Instead, domain appropriate approaches are needed. In this book, we focus on several core domains of robot learning. For accurate task execution, we need motor learning capabilities. For fast learning of the motor tasks, imitation learning offers the most promising approach. Self improvement requires reinforcement learning approaches that scale into the domain of complex robots. Finally, for efficient interaction of humans with robot systems, we will need a form of interaction learning. This chapter provides a general introduction to these issues and briefly presents the contributions of the subsequent chapters to the corresponding research topics.

ei

Web DOI [BibTex]

Web DOI [BibTex]


no image
Real-Time Local GP Model Learning

Nguyen-Tuong, D., Seeger, M., Peters, J.

In From Motor Learning to Interaction Learning in Robots, 264, pages: 193-207, Studies in Computational Intelligence, (Editors: Sigaud, O. and Peters, J.), Springer, Berlin, Germany, January 2010 (inbook)

Abstract
For many applications in robotics, accurate dynamics models are essential. However, in some applications, e.g., in model-based tracking control, precise dynamics models cannot be obtained analytically for sufficiently complex robot systems. In such cases, machine learning offers a promising alternative for approximating the robot dynamics using measured data. However, standard regression methods such as Gaussian process regression (GPR) suffer from high computational complexity which prevents their usage for large numbers of samples or online learning to date. In this paper, we propose an approximation to the standard GPR using local Gaussian processes models inspired by [Vijayakumar et al(2005)Vijayakumar, D’Souza, and Schaal, Snelson and Ghahramani(2007)]. Due to reduced computational cost, local Gaussian processes (LGP) can be applied for larger sample-sizes and online learning. Comparisons with other nonparametric regressions, e.g., standard GPR, support vector regression (SVR) and locally weighted proje ction regression (LWPR), show that LGP has high approximation accuracy while being sufficiently fast for real-time online learning.

ei

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Machine Learning Methods for Automatic Image Colorization

Charpiat, G., Bezrukov, I., Hofmann, M., Altun, Y., Schölkopf, B.

In Computational Photography: Methods and Applications, pages: 395-418, Digital Imaging and Computer Vision, (Editors: Lukac, R.), CRC Press, Boca Raton, FL, USA, 2010 (inbook)

Abstract
We aim to color greyscale images automatically, without any manual intervention. The color proposition could then be interactively corrected by user-provided color landmarks if necessary. Automatic colorization is nontrivial since there is usually no one-to-one correspondence between color and local texture. The contribution of our framework is that we deal directly with multimodality and estimate, for each pixel of the image to be colored, the probability distribution of all possible colors, instead of choosing the most probable color at the local level. We also predict the expected variation of color at each pixel, thus defining a non-uniform spatial coherency criterion. We then use graph cuts to maximize the probability of the whole colored image at the global level. We work in the L-a-b color space in order to approximate the human perception of distances between colors, and we use machine learning tools to extract as much information as possible from a dataset of colored examples. The resulting algorithm is fast, designed to be more robust to texture noise, and is above all able to deal with ambiguity, in contrary to previous approaches.

ei

PDF Web [BibTex]

PDF Web [BibTex]


no image
Learning the Reward Model of Dialogue POMDPs

Boularias, A., Chinaei, H., Chaib-Draa, B.

NIPS Workshop on Machine Learning for Assistive Technology (MLAT-2010), 2010 (poster)

ei

[BibTex]

[BibTex]


no image
Approaches Based on Support Vector Machine to Classification of Remote Sensing Data

Bruzzone, L., Persello, C.

In Handbook of Pattern Recognition and Computer Vision, pages: 329-352, (Editors: Chen, C.H.), ICP, London, UK, 2010 (inbook)

Abstract
This chapter presents an extensive and critical review on the use of kernel methods and in particular of support vector machines (SVMs) in the classification of remote-sensing (RS) data. The chapter recalls the mathematical formulation and the main theoretical concepts related to SVMs, and discusses the motivations at the basis of the use of SVMs in remote sensing. A review on the main applications of SVMs in classification of remote sensing is given, presenting a literature survey on the use of SVMs for the analysis of different kinds of RS images. In addition, the most recent methodological developments related to SVM-based classification techniques in RS are illustrated by focusing on semisupervised, domain adaptation, and context sensitive approaches. Finally, the most promising research directions on SVM in RS are identified and discussed.

ei

Web [BibTex]

Web [BibTex]


no image
Information-theoretic inference of common ancestors

Steudel, B., Ay, N.

Computing Research Repository (CoRR), abs/1010.5720, pages: 18, 2010 (techreport)

ei

Web [BibTex]

Web [BibTex]


no image
Erste Erfahrungen bei der Beurteilung hämato-onkologischer Krankheitsmanifestationen an den Extremitäten mit einem PET/MRT-Hybridsystem.

Sauter, A., Boss, A., Kolb, A., Mantlik, F., Bethge, W., Kanz, L., Pfannenberg, C., Stegger, L., Pichler, B., Claussen, C., Horger, M.

Thieme Verlag, Stuttgart, Germany, 91. Deutscher R{\"o}ntgenkongress, 2010 (poster)

ei

Web DOI [BibTex]

Web DOI [BibTex]


no image
Nanohandling robot cells

Fatikow, Sergej, Wich, Thomas, Dahmen, Christian, Jasper, Daniel, Stolle, Christian, Eichhorn, Volkmar, Hagemann, Saskia, Weigel-Jech, Michael

In Handbook of Nanophysics: Nanomedicine and Nanorobotics, pages: 1-31, CRC Press, 2010 (incollection)

pi

[BibTex]

[BibTex]


no image
Locally weighted regression for control

Ting, J., Vijayakumar, S., Schaal, S.

In Encyclopedia of Machine Learning, pages: 613-624, (Editors: Sammut, C.;Webb, G. I.), Springer, 2010, clmc (inbook)

Abstract
This is article addresses two topics: learning control and locally weighted regression.

am

link (url) [BibTex]

link (url) [BibTex]


no image
Physisorption in porous materials

Hirscher, M., Panella, B.

In Handbook of Hydrogen Storage, pages: 39-62, WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim, 2010 (incollection)

mms

[BibTex]

[BibTex]


no image
Adsorption technologies

Schmitz, B., Hirscher, M.

In Hydrogen and Fuel Cells, pages: 431-445, WILEY-VCH, Weinheim, 2010 (incollection)

mms

[BibTex]

[BibTex]


no image
Atomic-Force-Microscopy-Based Nanomanipulation Systems

Onal, C. D., Ozcan, O., Sitti, M.

In Handbook of Nanophysics: Nanomedicine and Nanorobotics, pages: 1-15, CRC Press, 2010 (incollection)

pi

[BibTex]

[BibTex]


no image
Taming the Beast: Guided Self-organization of Behavior in Autonomous Robots

Martius, G., Herrmann, J. M.

In From Animals to Animats 11, 6226, pages: 50-61, LNCS, Springer, 2010 (incollection)

al

link (url) DOI [BibTex]

link (url) DOI [BibTex]

2009


no image
Learning an Interactive Segmentation System

Nickisch, H., Kohli, P., Rother, C.

Max Planck Institute for Biological Cybernetics, December 2009 (techreport)

Abstract
Many successful applications of computer vision to image or video manipulation are interactive by nature. However, parameters of such systems are often trained neglecting the user. Traditionally, interactive systems have been treated in the same manner as their fully automatic counterparts. Their performance is evaluated by computing the accuracy of their solutions under some fixed set of user interactions. This paper proposes a new evaluation and learning method which brings the user in the loop. It is based on the use of an active robot user - a simulated model of a human user. We show how this approach can be used to evaluate and learn parameters of state-of-the-art interactive segmentation systems. We also show how simulated user models can be integrated into the popular max-margin method for parameter learning and propose an algorithm to solve the resulting optimisation problem.

ei

Web [BibTex]

2009


Web [BibTex]


no image
Machine Learning for Brain-Computer Interfaces

Hill, NJ.

Mini-Symposia on Assistive Machine Learning for People with Disabilities at NIPS (AMD), December 2009 (talk)

Abstract
Brain-computer interfaces (BCI) aim to be the ultimate in assistive technology: decoding a user‘s intentions directly from brain signals without involving any muscles or peripheral nerves. Thus, some classes of BCI potentially offer hope for users with even the most extreme cases of paralysis, such as in late-stage Amyotrophic Lateral Sclerosis, where nothing else currently allows communication of any kind. Other lines in BCI research aim to restore lost motor function in as natural a way as possible, reconnecting and in some cases re-training motor-cortical areas to control prosthetic, or previously paretic, limbs. Research and development are progressing on both invasive and non-invasive fronts, although BCI has yet to make a breakthrough to widespread clinical application. The high-noise high-dimensional nature of brain-signals, particularly in non-invasive approaches and in patient populations, make robust decoding techniques a necessity. Generally, the approach has been to use relatively simple feature extraction techniques, such as template matching and band-power estimation, coupled to simple linear classifiers. This has led to a prevailing view among applied BCI researchers that (sophisticated) machine-learning is irrelevant since "it doesn‘t matter what classifier you use once you‘ve done your preprocessing right and extracted the right features." I shall show a few examples of how this runs counter to both the empirical reality and the spirit of what needs to be done to bring BCI into clinical application. Along the way I‘ll highlight some of the interesting problems that remain open for machine-learners.

ei

PDF Web Web [BibTex]

PDF Web Web [BibTex]


no image
PAC-Bayesian Approach to Formulation of Clustering Objectives

Seldin, Y.

NIPS Workshop on "Clustering: Science or Art? Towards Principled Approaches", December 2009 (talk)

Abstract
Clustering is a widely used tool for exploratory data analysis. However, the theoretical understanding of clustering is very limited. We still do not have a well-founded answer to the seemingly simple question of "how many clusters are present in the data?", and furthermore a formal comparison of clusterings based on different optimization objectives is far beyond our abilities. The lack of good theoretical support gives rise to multiple heuristics that confuse the practitioners and stall development of the field. We suggest that the ill-posed nature of clustering problems is caused by the fact that clustering is often taken out of its subsequent application context. We argue that one does not cluster the data just for the sake of clustering it, but rather to facilitate the solution of some higher level task. By evaluation of the clustering‘s contribution to the solution of the higher level task it is possible to compare different clusterings, even those obtained by different optimization objectives. In the preceding work it was shown that such an approach can be applied to evaluation and design of co-clustering solutions. Here we suggest that this approach can be extended to other settings, where clustering is applied.

ei

PDF Web Web [BibTex]

PDF Web Web [BibTex]


no image
Semi-supervised Kernel Canonical Correlation Analysis of Human Functional Magnetic Resonance Imaging Data

Shelton, JA.

Women in Machine Learning Workshop (WiML), December 2009 (talk)

Abstract
Kernel Canonical Correlation Analysis (KCCA) is a general technique for subspace learning that incorporates principal components analysis (PCA) and Fisher linear discriminant analysis (LDA) as special cases. By finding directions that maximize correlation, KCCA learns representations tied more closely to underlying process generating the the data and can ignore high-variance noise directions. However, for data where acquisition in a given modality is expensive or otherwise limited, KCCA may suffer from small sample effects. We propose to use semi-supervised Laplacian regularization to utilize data that are present in only one modality. This manifold learning approach is able to find highly correlated directions that also lie along the data manifold, resulting in a more robust estimate of correlated subspaces. Functional magnetic resonance imaging (fMRI) acquired data are naturally amenable to subspace techniques as data are well aligned and such data of the human brain are a particularly interesting candidate. In this study we implemented various supervised and semi-supervised versions of KCCA on human fMRI data, with regression to single and multivariate labels (corresponding to video content subjects viewed during the image acquisition). In each variate condition, Laplacian regularization improved performance whereas the semi-supervised variants of KCCA yielded the best performance. We additionally analyze the weights learned by the regression in order to infer brain regions that are important during different types of visual processing.

ei

PDF Web [BibTex]

PDF Web [BibTex]


no image
An Incremental GEM Framework for Multiframe Blind Deconvolution, Super-Resolution, and Saturation Correction

Harmeling, S., Sra, S., Hirsch, M., Schölkopf, B.

(187), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, November 2009 (techreport)

Abstract
We develop an incremental generalized expectation maximization (GEM) framework to model the multiframe blind deconvolution problem. A simplistic version of this problem was recently studied by Harmeling etal~cite{harmeling09}. We solve a more realistic version of this problem which includes the following major features: (i) super-resolution ability emph{despite} noise and unknown blurring; (ii) saturation-correction, i.e., handling of overexposed pixels that can otherwise confound the image processing; and (iii) simultaneous handling of color channels. These features are seamlessly integrated into our incremental GEM framework to yield simple but efficient multiframe blind deconvolution algorithms. We present technical details concerning critical steps of our algorithms, especially to highlight how all operations can be written using matrix-vector multiplications. We apply our algorithm to real-world images from astronomy and super resolution tasks. Our experimental results show that our methods yield improve d resolution and deconvolution at the same time.

ei

PDF [BibTex]

PDF [BibTex]


no image
Efficient Filter Flow for Space-Variant Multiframe Blind Deconvolution

Hirsch, M., Sra, S., Schölkopf, B., Harmeling, S.

(188), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, November 2009 (techreport)

Abstract
Ultimately being motivated by facilitating space-variant blind deconvolution, we present a class of linear transformations, that are expressive enough for space-variant filters, but at the same time especially designed for efficient matrix-vector-multiplications. Successful results on astronomical imaging through atmospheric turbulences and on noisy magnetic resonance images of constantly moving objects demonstrate the practical significance of our approach.

ei

PDF [BibTex]

PDF [BibTex]


no image
Event-Related Potentials in Brain-Computer Interfacing

Hill, NJ.

Invited lecture on the bachelor & masters course "Introduction to Brain-Computer Interfacing", October 2009 (talk)

Abstract
An introduction to event-related potentials with specific reference to their use in brain-computer interfacing applications and research.

ei

PDF [BibTex]

PDF [BibTex]