Header logo is


2009


no image
Does Cognitive Science Need Kernels?

Jäkel, F., Schölkopf, B., Wichmann, F.

Trends in Cognitive Sciences, 13(9):381-388, September 2009 (article)

Abstract
Kernel methods are among the most successful tools in machine learning and are used in challenging data analysis problems in many disciplines. Here we provide examples where kernel methods have proven to be powerful tools for analyzing behavioral data, especially for identifying features in categorization experiments. We also demonstrate that kernel methods relate to perceptrons and exemplar models of categorization. Hence, we argue that kernel methods have neural and psychological plausibility, and theoretical results concerning their behavior are therefore potentially relevant for human category learning. In particular, we believe kernel methods have the potential to provide explanations ranging from the implementational via the algorithmic to the computational level.

ei

PDF Web DOI [BibTex]

2009


PDF Web DOI [BibTex]


no image
Robot Learning

Peters, J., Morimoto, J., Tedrake, R., Roy, N.

IEEE Robotics and Automation Magazine, 16(3):19-20, September 2009 (article)

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. As a result, 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. Considerable evidence of this exists in the use of learning in high-profile competitions such as RoboCup and the Defense Advanced Research Projects Agency (DARPA) challenges, and the growing number of research programs funded by governments around the world.

ei

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Kernel Methods in Computer Vision

Lampert, CH.

Foundations and Trends in Computer Graphics and Vision, 4(3):193-285, September 2009 (article)

Abstract
Over the last years, kernel methods have established themselves as powerful tools for computer vision researchers as well as for practitioners. In this tutorial, we give an introduction to kernel methods in computer vision from a geometric perspective, introducing not only the ubiquitous support vector machines, but also less known techniques for regression, dimensionality reduction, outlier detection and clustering. Additionally, we give an outlook on very recent, non-classical techniques for the prediction of structure data, for the estimation of statistical dependency and for learning the kernel function itself. All methods are illustrated with examples of successful application from the recent computer vision research literature.

ei

Web DOI [BibTex]

Web DOI [BibTex]


no image
Fast Kernel-Based Independent Component Analysis

Shen, H., Jegelka, S., Gretton, A.

IEEE Transactions on Signal Processing, 57(9):3498-3511, September 2009 (article)

Abstract
Recent approaches to independent component analysis (ICA) have used kernel independence measures to obtain highly accurate solutions, particularly where classical methods experience difficulty (for instance, sources with near-zero kurtosis). FastKICA (fast HSIC-based kernel ICA) is a new optimization method for one such kernel independence measure, the Hilbert-Schmidt Independence Criterion (HSIC). The high computational efficiency of this approach is achieved by combining geometric optimization techniques, specifically an approximate Newton-like method on the orthogonal group, with accurate estimates of the gradient and Hessian based on an incomplete Cholesky decomposition. In contrast to other efficient kernel-based ICA algorithms, FastKICA is applicable to any twice differentiable kernel function. Experimental results for problems with large numbers of sources and observations indicate that FastKICA provides more accurate solutions at a given cost than gradient descent on HSIC. Comparing with other recently published ICA methods, FastKICA is competitive in terms of accuracy, relatively insensitive to local minima when initialized far from independence, and more robust towards outliers. An analysis of the local convergence properties of FastKICA is provided.

ei

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Guest editorial: Special issue on robot learning, Part B

Peters, J., Ng, A.

Autonomous Robots, 27(2):91-92, August 2009 (article)

ei

PDF PDF DOI [BibTex]

PDF PDF DOI [BibTex]


no image
Policy Search for Motor Primitives

Peters, J., Kober, J.

KI - Zeitschrift K{\"u}nstliche Intelligenz, 23(3):38-40, August 2009 (article)

Abstract
Many motor skills in humanoid robotics can be learned using parametrized motor primitives from demonstrations. However, most interesting motor learning problems require self-improvement often beyond the reach of current reinforcement learning methods due to the high dimensionality of the state-space. We develop an EM-inspired algorithm applicable to complex motor learning tasks. We compare this algorithm to several well-known parametrized policy search methods and show that it outperforms them. We apply it to motor learning problems and show that it can learn the complex Ball-in-a-Cup task using a real Barrett WAM robot arm.

ei

Web [BibTex]

Web [BibTex]


no image
A neurophysiologically plausible population code model for human contrast discrimination

Goris, R., Wichmann, F., Henning, G.

Journal of Vision, 9(7):1-22, July 2009 (article)

Abstract
The pedestal effect is the improvement in the detectability of a sinusoidal grating in the presence of another grating of the same orientation, spatial frequency, and phase—usually called the pedestal. Recent evidence has demonstrated that the pedestal effect is differently modified by spectrally flat and notch-filtered noise: The pedestal effect is reduced in flat noise but virtually disappears in the presence of notched noise (G. B. Henning & F. A. Wichmann, 2007). Here we consider a network consisting of units whose contrast response functions resemble those of the cortical cells believed to underlie human pattern vision and demonstrate that, when the outputs of multiple units are combined by simple weighted summation—a heuristic decision rule that resembles optimal information combination and produces a contrast-dependent weighting profile—the network produces contrast-discrimination data consistent with psychophysical observations: The pedestal effect is present without noise, reduced in broadband noise, but almost disappears in notched noise. These findings follow naturally from the normalization model of simple cells in primary visual cortex, followed by response-based pooling, and suggest that in processing even low-contrast sinusoidal gratings, the visual system may combine information across neurons tuned to different spatial frequencies and orientations.

ei

Web DOI [BibTex]

Web DOI [BibTex]


no image
Falsificationism and Statistical Learning Theory: Comparing the Popper and Vapnik-Chervonenkis Dimensions

Corfield, D., Schölkopf, B., Vapnik, V.

Journal for General Philosophy of Science, 40(1):51-58, July 2009 (article)

Abstract
We compare Karl Popper’s ideas concerning the falsifiability of a theory with similar notions from the part of statistical learning theory known as VC-theory. Popper’s notion of the dimension of a theory is contrasted with the apparently very similar VC-dimension. Having located some divergences, we discuss how best to view Popper’s work from the perspective of statistical learning theory, either as a precursor or as aiming to capture a different learning activity.

ei

PDF DOI [BibTex]

PDF DOI [BibTex]


no image
Guest editorial: Special issue on robot learning, Part A

Peters, J., Ng, A.

Autonomous Robots, 27(1):1-2, July 2009 (article)

ei

PDF PDF DOI [BibTex]

PDF PDF DOI [BibTex]


no image
A Geometric Approach to Confidence Sets for Ratios: Fieller’s Theorem, Generalizations, and Bootstrap

von Luxburg, U., Franz, V.

Statistica Sinica, 19(3):1095-1117, July 2009 (article)

Abstract
We present a geometric method to determine confidence sets for the ratio E(Y)/E(X) of the means of random variables X and Y. This method reduces the problem of constructing confidence sets for the ratio of two random variables to the problem of constructing confidence sets for the means of one-dimensional random variables. It is valid in a large variety of circumstances. In the case of normally distributed random variables, the so constructed confidence sets coincide with the standard Fieller confidence sets. Generalizations of our construction lead to definitions of exact and conservative confidence sets for very general classes of distributions, provided the joint expectation of (X,Y) exists and the linear combinations of the form aX + bY are well-behaved. Finally, our geometric method allows to derive a very simple bootstrap approach for constructing conservative confidence sets for ratios which perform favorably in certain situations, in particular in the asymmetric heavy-tailed regime.

ei

PDF PDF Web [BibTex]


no image
Center-surround patterns emerge as optimal predictors for human saccade targets

Kienzle, W., Franz, M., Schölkopf, B., Wichmann, F.

Journal of Vision, 9(5:7):1-15, May 2009 (article)

Abstract
The human visual system is foveated, that is, outside the central visual field resolution and acuity drop rapidly. Nonetheless much of a visual scene is perceived after only a few saccadic eye movements, suggesting an effective strategy for selecting saccade targets. It has been known for some time that local image structure at saccade targets influences the selection process. However, the question of what the most relevant visual features are is still under debate. Here we show that center-surround patterns emerge as the optimal solution for predicting saccade targets from their local image structure. The resulting model, a one-layer feed-forward network, is surprisingly simple compared to previously suggested models which assume much more complex computations such as multi-scale processing and multiple feature channels. Nevertheless, our model is equally predictive. Furthermore, our findings are consistent with neurophysiological hardware in the superior colliculus. Bottom-up visual saliency may thus not be computed cortically as has been thought previously.

ei

PDF DOI [BibTex]


no image
Influence of Different Assignment Conditions on the Determination of Symmetric Homo-dimeric Structures with ARIA

Bardiaux, B., Bernard, A., Rieping, W., Habeck, M., Malliavin, TE., Nilges, M.

Proteins, 75(3):569-585, May 2009 (article)

Abstract
The ambiguous restraint for iterative assignment (ARIA) approach for NMR structure calculation is evaluated for symmetric homodimeric proteins by assessing the effect of several data analysis and assignment methods on the structure quality. In particular, we study the effects of network anchoring and spin-diffusion correction. The spin-diffusion correction improves the protein structure quality systematically, whereas network anchoring enhances the assignment efficiency by speeding up the convergence and coping with highly ambiguous data. For some homodimeric folds, network anchoring has been proved essential for unraveling both chain and proton assignment ambiguities.

ei

Web DOI [BibTex]

Web DOI [BibTex]


no image
Beamforming in Noninvasive Brain-Computer Interfaces

Grosse-Wentrup, M., Liefhold, C., Gramann, K., Buss, M.

IEEE Transactions on Biomedical Engineering, 56(4):1209-1219, April 2009 (article)

Abstract
Spatial filtering (SF) constitutes an integral part of building EEG-based brain–computer interfaces (BCIs). Algorithms frequently used for SF, such as common spatial patterns (CSPs) and independent component analysis, require labeled training data for identifying filters that provide information on a subject‘s intention, which renders these algorithms susceptible to overfitting on artifactual EEG components. In this study, beamforming is employed to construct spatial filters that extract EEG sources originating within predefined regions of interest within the brain. In this way, neurophysiological knowledge on which brain regions are relevant for a certain experimental paradigm can be utilized to construct unsupervised spatial filters that are robust against artifactual EEG components. Beamforming is experimentally compared with CSP and Laplacian spatial filtering (LP) in a two-class motor-imagery paradigm. It is demonstrated that beamforming outperforms CSP and LP on noisy datasets, while CSP and beamforming perform almost equally well on datasets with few artifactual trials. It is concluded that beamforming constitutes an alternative method for SF that might be particularly useful for BCIs used in clinical settings, i.e., in an environment where artifact-free datasets are difficult to obtain.

ei

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Constructing Sparse Kernel Machines Using Attractors

Lee, D., Jung, K., Lee, J.

IEEE Transactions on Neural Networks, 20(4):721-729, April 2009 (article)

Abstract
In this brief, a novel method that constructs a sparse kernel machine is proposed. The proposed method generates attractors as sparse solutions from a built-in kernel machine via a dynamical system framework. By readjusting the corresponding coefficients and bias terms, a sparse kernel machine that approximates a conventional kernel machine is constructed. The simulation results show that the constructed sparse kernel machine improves the efficiency of testing phase while maintaining comparable test error.

ei

Web DOI [BibTex]

Web DOI [BibTex]


no image
Optimal construction of k-nearest-neighbor graphs for identifying noisy clusters

Maier, M., Hein, M., von Luxburg, U.

Theoretical Computer Science, 410(19):1749-1764, April 2009 (article)

Abstract
We study clustering algorithms based on neighborhood graphs on a random sample of data points. The question we ask is how such a graph should be constructed in order to obtain optimal clustering results. Which type of neighborhood graph should one choose, mutual k-nearest-neighbor or symmetric k-nearest-neighbor? What is the optimal parameter k? In our setting, clusters are defined as connected components of the t-level set of the underlying probability distribution. Clusters are said to be identified in the neighborhood graph if connected components in the graph correspond to the true underlying clusters. Using techniques from random geometric graph theory, we prove bounds on the probability that clusters are identified successfully, both in a noise-free and in a noisy setting. Those bounds lead to several conclusions. First, k has to be chosen surprisingly high (rather of the order n than of the order logn) to maximize the probability of cluster identification. Secondly, the major difference between the mutual and the symmetric k-nearest-neighbor graph occurs when one attempts to detect the most significant cluster only.

ei

PDF PDF DOI [BibTex]


no image
Overlap and refractory effects in a Brain-Computer Interface speller based on the visual P300 Event-Related Potential

Martens, S., Hill, N., Farquhar, J., Schölkopf, B.

Journal of Neural Engineering, 6(2):1-9, April 2009 (article)

Abstract
We reveal the presence of refractory and overlap effects in the event-related potentials in visual P300 speller datasets, and we show their negative impact on the performance of the system. This finding has important implications for how to encode the letters that can be selected for communication. However, we show that such effects are dependent on stimulus parameters: an alternative stimulus type based on apparent motion suffers less from the refractory effects and leads to an improved letter prediction performance.

ei

PDF DOI [BibTex]


no image
Nearest Neighbor Clustering: A Baseline Method for Consistent Clustering with Arbitrary Objective Functions

Bubeck, S., von Luxburg, U.

Journal of Machine Learning Research, 10, pages: 657-698, March 2009 (article)

Abstract
Clustering is often formulated as a discrete optimization problem. The objective is to find, among all partitions of the data set, the best one according to some quality measure. However, in the statistical setting where we assume that the finite data set has been sampled from some underlying space, the goal is not to find the best partition of the given sample, but to approximate the true partition of the underlying space. We argue that the discrete optimization approach usually does not achieve this goal, and instead can lead to inconsistency. We construct examples which provably have this behavior. As in the case of supervised learning, the cure is to restrict the size of the function classes under consideration. For appropriate “small” function classes we can prove very general consistency theorems for clustering optimization schemes. As one particular algorithm for clustering with a restricted function space we introduce “nearest neighbor clustering”. Similar to the k-nearest neighbor classifier in supervised learning, this algorithm can be seen as a general baseline algorithm to minimize arbitrary clustering objective functions. We prove that it is statistically consistent for all commonly used clustering objective functions.

ei

PDF Web [BibTex]


no image
Protein Functional Class Prediction With a Combined Graph

Shin, H., Tsuda, K., Schölkopf, B.

Expert Systems with Applications, 36(2):3284-3292, March 2009 (article)

Abstract
In bioinformatics, there exist multiple descriptions of graphs for the same set of genes or proteins. For instance, in yeast systems, graph edges can represent different relationships such as protein–protein interactions, genetic interactions, or co-participation in a protein complex, etc. Relying on similarities between nodes, each graph can be used independently for prediction of protein function. However, since different graphs contain partly independent and partly complementary information about the problem at hand, one can enhance the total information extracted by combining all graphs. In this paper, we propose a method for integrating multiple graphs within a framework of semi-supervised learning. The method alternates between minimizing the objective function with respect to network output and with respect to combining weights. We apply the method to the task of protein functional class prediction in yeast. The proposed method performs significantly better than the same algorithm trained on any singl e graph.

ei

Web DOI [BibTex]

Web DOI [BibTex]


no image
Gaussian Process Dynamic Programming

Deisenroth, M., Rasmussen, C., Peters, J.

Neurocomputing, 72(7-9):1508-1524, March 2009 (article)

Abstract
Reinforcement learning (RL) and optimal control of systems with contin- uous states and actions require approximation techniques in most interesting cases. In this article, we introduce Gaussian process dynamic programming (GPDP), an approximate value-function based RL algorithm. We consider both a classic optimal control problem, where problem-specific prior knowl- edge is available, and a classic RL problem, where only very general priors can be used. For the classic optimal control problem, GPDP models the unknown value functions with Gaussian processes and generalizes dynamic programming to continuous-valued states and actions. For the RL problem, GPDP starts from a given initial state and explores the state space using Bayesian active learning. To design a fast learner, available data has to be used efficiently. Hence, we propose to learn probabilistic models of the a priori unknown transition dynamics and the value functions on the fly. In both cases, we successfully apply the resulting continuous-valued controllers to the under-actuated pendulum swing up and analyze the performances of the suggested algorithms. It turns out that GPDP uses data very efficiently and can be applied to problems, where classic dynamic programming would be cumbersome.

ei

PDF PDF DOI [BibTex]

PDF PDF DOI [BibTex]


no image
Towards quantitative PET/MRI: a review of MR-based attenuation correction techniques

Hofmann, M., Pichler, B., Schölkopf, B., Beyer, T.

European Journal of Nuclear Medicine and Molecular Imaging, 36(Supplement 1):93-104, March 2009 (article)

Abstract
Introduction Positron emission tomography (PET) is a fully quantitative technology for imaging metabolic pathways and dynamic processes in vivo. Attenuation correction of raw PET data is a prerequisite for quantification and is typically based on separate transmission measurements. In PET/CT attenuation correction, however, is performed routinely based on the available CT transmission data. Objective Recently, combined PET/magnetic resonance (MR) has been proposed as a viable alternative to PET/CT. Current concepts of PET/MRI do not include CT-like transmission sources and, therefore, alternative methods of PET attenuation correction must be found. This article reviews existing approaches to MR-based attenuation correction (MR-AC). Most groups have proposed MR-AC algorithms for brain PET studies and more recently also for torso PET/MR imaging. Most MR-AC strategies require the use of complementary MR and transmission images, or morphology templates generated from transmission images. We review and discuss these algorithms and point out challenges for using MR-AC in clinical routine. Discussion MR-AC is work-in-progress with potentially promising results from a template-based approach applicable to both brain and torso imaging. While efforts are ongoing in making clinically viable MR-AC fully automatic, further studies are required to realize the potential benefits of MR-based motion compensation and partial volume correction of the PET data.

ei

PDF DOI [BibTex]

PDF DOI [BibTex]


no image
Generating Spike Trains with Specified Correlation Coefficients

Macke, J., Berens, P., Ecker, A., Tolias, A., Bethge, M.

Neural Computation, 21(2):397-423, February 2009 (article)

Abstract
Spike trains recorded from populations of neurons can exhibit substantial pairwise correlations between neurons and rich temporal structure. Thus, for the realistic simulation and analysis of neural systems, it is essential to have efficient methods for generating artificial spike trains with specified correlation structure. Here we show how correlated binary spike trains can be simulated by means of a latent multivariate gaussian model. Sampling from the model is computationally very efficient and, in particular, feasible even for large populations of neurons. The entropy of the model is close to the theoretical maximum for a wide range of parameters. In addition, this framework naturally extends to correlations over time and offers an elegant way to model correlated neural spike counts with arbitrary marginal distributions.

ei

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Automatic detection of preclinical neurodegeneration: Presymptomatic Huntington disease

Klöppel, S., Chu, C., Tan, G., Draganski, B., Johnson, H., Paulsen, J., Kienzle, W., Tabrizi, S., Ashburner, J., Frackowiak, R.

Neurology, 72(5):426-431, February 2009 (article)

Abstract
Background: Treatment of neurodegenerative diseases is likely to be most beneficial in the very early, possibly preclinical stages of degeneration. We explored the usefulness of fully automatic structural MRI classification methods for detecting subtle degenerative change. The availability of a definitive genetic test for Huntington disease (HD) provides an excellent metric for judging the performance of such methods in gene mutation carriers who are free of symptoms. Methods: Using the gray matter segment of MRI scans, this study explored the usefulness of a multivariate support vector machine to automatically identify presymptomatic HD gene mutation carriers (PSCs) in the absence of any a priori information. A multicenter data set of 96 PSCs and 95 age- and sex-matched controls was studied. The PSC group was subclassified into three groups based on time from predicted clinical onset, an estimate that is a function of DNA mutation size and age. Results: Subjects with at least a 33% chance of developing unequivocal signs of HD in 5 years were correctly assigned to the PSC group 69% of the time. Accuracy improved to 83% when regions affected by the disease were selected a priori for analysis. Performance was at chance when the probability of developing symptoms in 5 years was less than 10%. Conclusions: Presymptomatic Huntington disease gene mutation carriers close to estimated diagnostic onset were successfully separated from controls on the basis of single anatomic scans, without additional a priori information. Prior information is required to allow separation when degenerative changes are either subtle or variable.

ei

Web [BibTex]

Web [BibTex]


no image
Enumeration of condition-dependent dense modules in protein interaction networks

Georgii, E., Dietmann, S., Uno, T., Pagel, P., Tsuda, K.

Bioinformatics, 25(7):933-940, February 2009 (article)

Abstract
Motivation: Modern systems biology aims at understanding how the different molecular components of a biological cell interact. Often, cellular functions are performed by complexes consisting of many different proteins. The composition of these complexes may change according to the cellular environment, and one protein may be involved in several different processes. The automatic discovery of functional complexes from protein interaction data is challenging. While previous approaches use approximations to extract dense modules, our approach exactly solves the problem of dense module enumeration. Furthermore, constraints from additional information sources such as gene expression and phenotype data can be integrated, so we can systematically mine for dense modules with interesting profiles. Results: Given a weighted protein interaction network, our method discovers all protein sets that satisfy a user-defined minimum density threshold. We employ a reverse search strategy, which allows us to exploit the density criterion in an efficient way. Our experiments show that the novel approach is feasible and produces biologically meaningful results. In comparative validation studies using yeast data, the method achieved the best overall prediction performance with respect to confirmed complexes. Moreover, by enhancing the yeast network with phenotypic and phylogenetic profiles and the human network with tissue-specific expression data, we identified condition-dependent complex variants.

ei

Web DOI [BibTex]

Web DOI [BibTex]


no image
Prototype Classification: Insights from Machine Learning

Graf, A., Bousquet, O., Rätsch, G., Schölkopf, B.

Neural Computation, 21(1):272-300, January 2009 (article)

Abstract
We shed light on the discrimination between patterns belonging to two different classes by casting this decoding problem into a generalized prototype framework. The discrimination process is then separated into two stages: a projection stage that reduces the dimensionality of the data by projecting it on a line and a threshold stage where the distributions of the projected patterns of both classes are separated. For this, we extend the popular mean-of-class prototype classification using algorithms from machine learning that satisfy a set of invariance properties. We report a simple yet general approach to express different types of linear classification algorithms in an identical and easy-to-visualize formal framework using generalized prototypes where these prototypes are used to express the normal vector and offset of the hyperplane. We investigate nonmargin classifiers such as the classical prototype classifier, the Fisher classifier, and the relevance vector machine. We then study hard and soft margin cl assifiers such as the support vector machine and a boosted version of the prototype classifier. Subsequently, we relate mean-of-class prototype classification to other classification algorithms by showing that the prototype classifier is a limit of any soft margin classifier and that boosting a prototype classifier yields the support vector machine. While giving novel insights into classification per se by presenting a common and unified formalism, our generalized prototype framework also provides an efficient visualization and a principled comparison of machine learning classification.

ei

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
The DICS repository: module-assisted analysis of disease-related gene lists

Dietmann, S., Georgii, E., Antonov, A., Tsuda, K., Mewes, H.

Bioinformatics, 25(6):830-831, January 2009 (article)

Abstract
The DICS database is a dynamic web repository of computationally predicted functional modules from the human protein–protein interaction network. It provides references to the CORUM, DrugBank, KEGG and Reactome pathway databases. DICS can be accessed for retrieving sets of overlapping modules and protein complexes that are significantly enriched in a gene list, thereby providing valuable information about the functional context.

ei

Web DOI [BibTex]

Web DOI [BibTex]


no image
mGene: accurate SVM-based gene finding with an application to nematode genomes

Schweikert, G., Zien, A., Zeller, G., Behr, J., Dieterich, C., Ong, C., Philips, P., De Bona, F., Hartmann, L., Bohlen, A., Krüger, N., Sonnenburg, S., Rätsch, G.

Genome Research, 19(11):2133-43, 2009 (article)

Abstract
We present a highly accurate gene-prediction system for eukaryotic genomes, called mGene. It combines in an unprecedented manner the flexibility of generalized hidden Markov models (gHMMs) with the predictive power of modern machine learning methods, such as Support Vector Machines (SVMs). Its excellent performance was proved in an objective competition based on the genome of the nematode Caenorhabditis elegans. Considering the average of sensitivity and specificity, the developmental version of mGene exhibited the best prediction performance on nucleotide, exon, and transcript level for ab initio and multiple-genome gene-prediction tasks. The fully developed version shows superior performance in 10 out of 12 evaluation criteria compared with the other participating gene finders, including Fgenesh++ and Augustus. An in-depth analysis of mGene's genome-wide predictions revealed that approximately 2200 predicted genes were not contained in the current genome annotation. Testing a subset of 57 of these genes by RT-PCR and sequencing, we confirmed expression for 24 (42%) of them. mGene missed 300 annotated genes, out of which 205 were unconfirmed. RT-PCR testing of 24 of these genes resulted in a success rate of merely 8%. These findings suggest that even the gene catalog of a well-studied organism such as C. elegans can be substantially improved by mGene's predictions. We also provide gene predictions for the four nematodes C. briggsae, C. brenneri, C. japonica, and C. remanei. Comparing the resulting proteomes among these organisms and to the known protein universe, we identified many species-specific gene inventions. In a quality assessment of several available annotations for these genomes, we find that mGene's predictions are most accurate.

ei

DOI [BibTex]

DOI [BibTex]


no image
Structure and activity of the N-terminal substrate recognition domains in proteasomal ATPases

Djuranovic, S., Hartmann, MD., Habeck, M., Ursinus, A., Zwickl, P., Martin, J., Lupas, AN., Zeth, K.

Molecular Cell, 34(5):580-590, 2009 (article)

Abstract
The proteasome forms the core of the protein quality control system in archaea and eukaryotes and also occurs in one bacterial lineage, the Actinobacteria. Access to its proteolytic compartment is controlled by AAA ATPases, whose N-terminal domains (N domains) are thought to mediate substrate recognition. The N domains of an archaeal proteasomal ATPase, Archaeoglobus fulgidus PAN, and of its actinobacterial homolog, Rhodococcus erythropolis ARC, form hexameric rings, whose subunits consist of an N-terminal coiled coil and a C-terminal OB domain. In ARC-N, the OB domains are duplicated and form separate rings. PAN-N and ARC-N can act as chaperones, preventing the aggregation of heterologous proteins in vitro, and this activity is preserved in various chimeras, even when these include coiled coils and OB domains from unrelated proteins. The structures suggest a molecular mechanism for substrate processing based on concerted radial motions of the coiled coils relative to the OB rings.

ei

DOI [BibTex]

DOI [BibTex]


no image
Discussion of: Brownian Distance Covariance

Gretton, A., Fukumizu, K., Sriperumbudur, B.

The Annals of Applied Statistics, 3(4):1285-1294, 2009 (article)

ei

[BibTex]

[BibTex]


no image
Non-linear System Identification: Visual Saliency Inferred from Eye-Movement Data

Wichmann, F., Kienzle, W., Schölkopf, B., Franz, M.

Journal of Vision, 9(8):article 32, 2009 (article)

Abstract
For simple visual patterns under the experimenter's control we impose which information, or features, an observer can use to solve a given perceptual task. For natural vision tasks, however, there are typically a multitude of potential features in a given visual scene which the visual system may be exploiting when analyzing it: edges, corners, contours, etc. Here we describe a novel non-linear system identification technique based on modern machine learning methods that allows the critical features an observer uses to be inferred directly from the observer's data. The method neither requires stimuli to be embedded in noise nor is it limited to linear perceptive fields (classification images). We demonstrate our technique by deriving the critical image features observers fixate in natural scenes (bottom-up visual saliency). Unlike previous studies where the relevant structure is determined manually—e.g. by selecting Gabors as visual filters—we do not make any assumptions in this regard, but numerically infer number and properties them from the eye-movement data. We show that center-surround patterns emerge as the optimal solution for predicting saccade targets from local image structure. The resulting model, a one-layer feed-forward network with contrast gain-control, is surprisingly simple compared to previously suggested saliency models. Nevertheless, our model is equally predictive. Furthermore, our findings are consistent with neurophysiological hardware in the superior colliculus. Bottom-up visual saliency may thus not be computed cortically as has been thought previously.

ei

Web DOI [BibTex]


no image
mGene.web: a web service for accurate computational gene finding

Schweikert, G., Behr, J., Zien, A., Zeller, G., Ong, C., Sonnenburg, S., Rätsch, G.

Nucleic Acids Research, 37, pages: W312-6, 2009 (article)

Abstract
We describe mGene.web, a web service for the genome-wide prediction of protein coding genes from eukaryotic DNA sequences. It offers pre-trained models for the recognition of gene structures including untranslated regions in an increasing number of organisms. With mGene.web, users have the additional possibility to train the system with their own data for other organisms on the push of a button, a functionality that will greatly accelerate the annotation of newly sequenced genomes. The system is built in a highly modular way, such that individual components of the framework, like the promoter prediction tool or the splice site predictor, can be used autonomously. The underlying gene finding system mGene is based on discriminative machine learning techniques and its high accuracy has been demonstrated in an international competition on nematode genomes. mGene.web is available at http://www.mgene.org/web, it is free of charge and can be used for eukaryotic genomes of small to moderate size (several hundred Mbp).

ei

DOI [BibTex]

DOI [BibTex]

2008


no image
Modelling contrast discrimination data suggest both the pedestal effect and stochastic resonance to be caused by the same mechanism

Goris, R., Wagemans, J., Wichmann, F.

Journal of Vision, 8(15):1-21, November 2008 (article)

Abstract
Computational models of spatial vision typically make use of a (rectified) linear filter, a nonlinearity and dominant late noise to account for human contrast discrimination data. Linear–nonlinear cascade models predict an improvement in observers' contrast detection performance when low, subthreshold levels of external noise are added (i.e., stochastic resonance). Here, we address the issue whether a single contrast gain-control model of early spatial vision can account for both the pedestal effect, i.e., the improved detectability of a grating in the presence of a low-contrast masking grating, and stochastic resonance. We measured contrast discrimination performance without noise and in both weak and moderate levels of noise. Making use of a full quantitative description of our data with few parameters combined with comprehensive model selection assessments, we show the pedestal effect to be more reduced in the presence of weak noise than in moderate noise. This reduction rules out independent, additive sources of performance improvement and, together with a simulation study, supports the parsimonious explanation that a single mechanism underlies the pedestal effect and stochastic resonance in contrast perception.

ei

Web DOI [BibTex]


no image
gBoost: A Mathematical Programming Approach to Graph Classification and Regression

Saigo, H., Nowozin, S., Kadowaki, T., Kudo, T., Tsuda, K.

Machine Learning, 75(1):69-89, November 2008 (article)

Abstract
Graph mining methods enumerate frequently appearing subgraph patterns, which can be used as features for subsequent classification or regression. However, frequent patterns are not necessarily informative for the given learning problem. We propose a mathematical programming boosting method (gBoost) that progressively collects informative patterns. Compared to AdaBoost, gBoost can build the prediction rule with fewer iterations. To apply the boosting method to graph data, a branch-and-bound pattern search algorithm is developed based on the DFS code tree. The constructed search space is reused in later iterations to minimize the computation time. Our method can learn more efficiently than the simpler method based on frequent substructure mining, because the output labels are used as an extra information source for pruning the search space. Furthermore, by engineering the mathematical program, a wide range of machine learning problems can be solved without modifying the pattern search algorithm.

ei

PDF DOI [BibTex]

PDF DOI [BibTex]


no image
Machine Learning for Motor Skills in Robotics

Peters, J.

K{\"u}nstliche Intelligenz, 2008(4):41-43, November 2008 (article)

Abstract
Autonomous robots that can adapt to novel situations has been a long standing vision of robotics, artificial intelligence, and the cognitive sciences. Early approaches to this goal during the heydays of artificial intelligence research in the late 1980s, however, made it clear that an approach purely based on reasoning or human insights would not be able to model all the perceptuomotor tasks of future robots. Instead, new hope was put in the growing wake of machine learning that promised fully adaptive control algorithms which learn both by observation and trial-and-error. However, to date, learning techniques have yet to fulfill this promise as only few methods manage to scale into the high-dimensional domains of manipulator and humanoid robotics and usually scaling was only achieved in precisely pre-structured domains. We have investigated the ingredients for a general approach to motor skill learning in order to get one step closer towards human-like performance. For doing so, we study two major components for such an approach, i.e., firstly, a theoretically well-founded general approach to representing the required control structures for task representation and execution and, secondly, appropriate learning algorithms which can be applied in this setting.

ei

PDF Web [BibTex]

PDF Web [BibTex]


no image
Kernels, Regularization and Differential Equations

Steinke, F., Schölkopf, B.

Pattern Recognition, 41(11):3271-3286, November 2008 (article)

Abstract
Many common machine learning methods such as Support Vector Machines or Gaussian process inference make use of positive definite kernels, reproducing kernel Hilbert spaces, Gaussian processes, and regularization operators. In this work these objects are presented in a general, unifying framework, and interrelations are highlighted. With this in mind we then show how linear stochastic differential equation models can be incorporated naturally into the kernel framework. And vice versa, many kernel machines can be interpreted in terms of differential equations. We focus especially on ordinary differential equations, also known as dynamical systems, and it is shown that standard kernel inference algorithms are equivalent to Kalman filter methods based on such models. In order not to cloud qualitative insights with heavy mathematical machinery, we restrict ourselves to finite domains, implying that differential equations are treated via their corresponding finite difference equations.

ei

PDF DOI [BibTex]

PDF DOI [BibTex]


no image
Mixture Models for Protein Structure Ensembles

Hirsch, M., Habeck, M.

Bioinformatics, 24(19):2184-2192, October 2008 (article)

ei

Web DOI [BibTex]

Web DOI [BibTex]


no image
Structure of the human voltage-dependent anion channel

Bayrhuber, M., Meins, T., Habeck, M., Becker, S., Giller, K., Villinger, S., Vonrhein, C., Griesinger, C., Zweckstetter, M., Zeth, K.

Proceedings of the National Academy of Sciences of the United States of America, 105(40):15370-15375, October 2008 (article)

Abstract
The voltage-dependent anion channel (VDAC), also known as mitochondrial porin, is the most abundant protein in the mitochondrial outer membrane (MOM). VDAC is the channel known to guide the metabolic flux across the MOM and plays a key role in mitochondrially induced apoptosis. Here, we present the 3D structure of human VDAC1, which was solved conjointly by NMR spectroscopy and x-ray crystallography. Human VDAC1 (hVDAC1) adopts a β-barrel architecture composed of 19 β-strands with an α-helix located horizontally midway within the pore. Bioinformatic analysis indicates that this channel architecture is common to all VDAC proteins and is adopted by the general import pore TOM40 of mammals, which is also located in the MOM.

ei

Web DOI [BibTex]

Web DOI [BibTex]


no image
MRI-Based Attenuation Correction for PET/MRI: A Novel Approach Combining Pattern Recognition and Atlas Registration

Hofmann, M., Steinke, F., Scheel, V., Charpiat, G., Farquhar, J., Aschoff, P., Brady, M., Schölkopf, B., Pichler, B.

Journal of Nuclear Medicine, 49(11):1875-1883, October 2008 (article)

Abstract
For quantitative PET information, correction of tissue photon attenuation is mandatory. Generally in conventional PET, the attenuation map is obtained from a transmission scan, which uses a rotating radionuclide source, or from the CT scan in a combined PET/CT scanner. In the case of PET/MRI scanners currently under development, insufficient space for the rotating source exists; the attenuation map can be calculated from the MR image instead. This task is challenging because MR intensities correlate with proton densities and tissue-relaxation properties, rather than with attenuation-related mass density. METHODS: We used a combination of local pattern recognition and atlas registration, which captures global variation of anatomy, to predict pseudo-CT images from a given MR image. These pseudo-CT images were then used for attenuation correction, as the process would be performed in a PET/CT scanner. RESULTS: For human brain scans, we show on a database of 17 MR/CT image pairs that our method reliably enables e stimation of a pseudo-CT image from the MR image alone. On additional datasets of MRI/PET/CT triplets of human brain scans, we compare MRI-based attenuation correction with CT-based correction. Our approach enables PET quantification with a mean error of 3.2% for predefined regions of interest, which we found to be clinically not significant. However, our method is not specific to brain imaging, and we show promising initial results on 1 whole-body animal dataset. CONCLUSION: This method allows reliable MRI-based attenuation correction for human brain scans. Further work is necessary to validate the method for whole-body imaging.

ei

Web DOI [BibTex]

Web DOI [BibTex]


no image
Support Vector Machines and Kernels for Computational Biology

Ben-Hur, A., Ong, C., Sonnenburg, S., Schölkopf, B., Rätsch, G.

PLoS Computational Biology, 4(10: e1000173):1-10, October 2008 (article)

ei

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Approximations for Binary Gaussian Process Classification

Nickisch, H., Rasmussen, C.

Journal of Machine Learning Research, 9, pages: 2035-2078, October 2008 (article)

Abstract
We provide a comprehensive overview of many recent algorithms for approximate inference in Gaussian process models for probabilistic binary classification. The relationships between several approaches are elucidated theoretically, and the properties of the different algorithms are corroborated by experimental results. We examine both 1) the quality of the predictive distributions and 2) the suitability of the different marginal likelihood approximations for model selection (selecting hyperparameters) and compare to a gold standard based on MCMC. Interestingly, some methods produce good predictive distributions although their marginal likelihood approximations are poor. Strong conclusions are drawn about the methods: The Expectation Propagation algorithm is almost always the method of choice unless the computational budget is very tight. We also extend existing methods in various ways, and provide unifying code implementing all approaches.

ei

PDF PDF [BibTex]

PDF PDF [BibTex]


no image
Accurate NMR Structures Through Minimization of an Extended Hybrid Energy

Nilges, M., Bernard, A., Bardiaux, B., Malliavin, T., Habeck, M., Rieping, W.

Structure, 16(9):1305-1312, September 2008 (article)

Abstract
The use of generous distance bounds has been the hallmark of NMR structure determination. However, bounds necessitate the estimation of data quality before the calculation, reduce the information content, introduce human bias, and allow for major errors in the structures. Here, we propose a new rapid structure calculation scheme based on Bayesian analysis. The minimization of an extended energy function, including a new type of distance restraint and a term depending on the data quality, results in an estimation of the data quality in addition to coordinates. This allows for the determination of the optimal weight on the experimental information. The resulting structures are of better quality and closer to the X–ray crystal structure of the same molecule. With the new calculation approach, the analysis of discrepancies from the target distances becomes meaningful. The strategy may be useful in other applications—for example, in homology modeling.

ei

PDF DOI [BibTex]

PDF DOI [BibTex]


no image
Similarity, Kernels, and the Triangle Inequality

Jäkel, F., Schölkopf, B., Wichmann, F.

Journal of Mathematical Psychology, 52(5):297-303, September 2008 (article)

Abstract
Similarity is used as an explanatory construct throughout psychology and multidimensional scaling (MDS) is the most popular way to assess similarity. In MDS, similarity is intimately connected to the idea of a geometric representation of stimuli in a perceptual space. Whilst connecting similarity and closeness of stimuli in a geometric representation may be intuitively plausible, Tversky and Gati [Tversky, A., Gati, I. (1982). Similarity, separability, and the triangle inequality. Psychological Review, 89(2), 123–154] have reported data which are inconsistent with the usual geometric representations that are based on segmental additivity. We show that similarity measures based on Shepard’s universal law of generalization [Shepard, R. N. (1987). Toward a universal law of generalization for psychologica science. Science, 237(4820), 1317–1323] lead to an inner product representation in a reproducing kernel Hilbert space. In such a space stimuli are represented by their similarity to all other stimuli. This representation, based on Shepard’s law, has a natural metric that does not have additive segments whilst still retaining the intuitive notion of connecting similarity and distance between stimuli. Furthermore, this representation has the psychologically appealing property that the distance between stimuli is bounded.

ei

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Comparison of Pattern Recognition Methods in Classifying High-resolution BOLD Signals Obtained at High Magnetic Field in Monkeys

Ku, S., Gretton, A., Macke, J., Logothetis, N.

Magnetic Resonance Imaging, 26(7):1007-1014, September 2008 (article)

Abstract
Pattern recognition methods have shown that functional magnetic resonance imaging (fMRI) data can reveal significant information about brain activity. For example, in the debate of how object categories are represented in the brain, multivariate analysis has been used to provide evidence of a distributed encoding scheme [Science 293:5539 (2001) 2425–2430]. Many follow-up studies have employed different methods to analyze human fMRI data with varying degrees of success [Nature reviews 7:7 (2006) 523–534]. In this study, we compare four popular pattern recognition methods: correlation analysis, support-vector machines (SVM), linear discriminant analysis (LDA) and Gaussian naïve Bayes (GNB), using data collected at high field (7 Tesla) with higher resolution than usual fMRI studies. We investigate prediction performance on single trials and for averages across varying numbers of stimulus presentations. The performance of the various algorithms depends on the nature of the brain activity being categorized: for several tasks, many of the methods work well, whereas for others, no method performs above chance level. An important factor in overall classification performance is careful preprocessing of the data, including dimensionality reduction, voxel selection and outlier elimination.

ei

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
A Single-shot Measurement of the Energy of Product States in a Translation Invariant Spin Chain Can Replace Any Quantum Computation

Janzing, D., Wocjan, P., Zhang, S.

New Journal of Physics, 10(093004):1-18, September 2008 (article)

Abstract
In measurement-based quantum computation, quantum algorithms are implemented via sequences of measurements. We describe a translationally invariant finite-range interaction on a one-dimensional qudit chain and prove that a single-shot measurement of the energy of an appropriate computational basis state with respect to this Hamiltonian provides the output of any quantum circuit. The required measurement accuracy scales inverse polynomially with the size of the simulated quantum circuit. This shows that the implementation of energy measurements on generic qudit chains is as hard as the realization of quantum computation. Here, a ‘measurement‘ is any procedure that samples from the spectral measurement induced by the observable and the state under consideration. As opposed to measurement-based quantum computation, the post-measurement state is irrelevant.

ei

PDF DOI [BibTex]


no image
Voluntary Brain Regulation and Communication with ECoG-Signals

Hinterberger, T., Widmann, G., Lal, T., Hill, J., Tangermann, M., Rosenstiel, W., Schölkopf, B., Elger, C., Birbaumer, N.

Epilepsy and Behavior, 13(2):300-306, August 2008 (article)

Abstract
Brain–computer interfaces (BCIs) can be used for communication in writing without muscular activity or for learning to control seizures by voluntary regulation of brain signals such as the electroencephalogram (EEG). Three of five patients with epilepsy were able to spell their names with electrocorticogram (ECoG) signals derived from motor-related areas within only one or two training sessions. Imagery of finger or tongue movements was classified with support-vector classification of autoregressive coefficients derived from the ECoG signals. After training of the classifier, binary classification responses were used to select letters from a computer-generated menu. Offline analysis showed increased theta activity in the unsuccessful patients, whereas the successful patients exhibited dominant sensorimotor rhythms that they could control. The high spatial resolution and increased signal-to-noise ratio in ECoG signals, combined with short training periods, may offer an alternative for communication in complete paralysis, locked-in syndrome, and motor restoration.

ei

PDF DOI [BibTex]

PDF DOI [BibTex]


no image
Multi-class Common Spatial Pattern and Information Theoretic Feature Extraction

Grosse-Wentrup, M., Buss, M.

IEEE Transactions on Biomedical Engineering, 55(8):1991-2000, August 2008 (article)

Abstract
We address two shortcomings of the common spatial patterns (CSP) algorithm for spatial filtering in the context of brain--computer interfaces (BCIs) based on electroencephalography/magnetoencephalography (EEG/MEG): First, the question of optimality of CSP in terms of the minimal achievable classification error remains unsolved. Second, CSP has been initially proposed for two-class paradigms. Extensions to multiclass paradigms have been suggested, but are based on heuristics. We address these shortcomings in the framework of information theoretic feature extraction (ITFE). We show that for two-class paradigms, CSP maximizes an approximation of mutual information of extracted EEG/MEG components and class labels. This establishes a link between CSP and the minimal classification error. For multiclass paradigms, we point out that CSP by joint approximate diagonalization (JAD) is equivalent to independent component analysis (ICA), and provide a method to choose those independent components (ICs) that approximately maximize mutual information of ICs and class labels. This eliminates the need for heuristics in multiclass CSP, and allows incorporating prior class probabilities. The proposed method is applied to the dataset IIIa of the third BCI competition, and is shown to increase the mean classification accuracy by 23.4% in comparison to multiclass CSP.

ei

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
At-TAX: A Whole Genome Tiling Array Resource for Developmental Expression Analysis and Transcript Identification in Arabidopsis thaliana

Laubinger, S., Zeller, G., Henz, S., Sachsenberg, T., Widmer, C., Naouar, N., Vuylsteke, M., Schölkopf, B., Rätsch, G., Weigel, D.

Genome Biology, 9(7: R112):1-16, July 2008 (article)

Abstract
Gene expression maps for model organisms, including Arabidopsis thaliana, have typically been created using gene-centric expression arrays. Here, we describe a comprehensive expression atlas, Arabidopsis thaliana Tiling Array Express (At-TAX), which is based on whole-genome tiling arrays. We demonstrate that tiling arrays are accurate tools for gene expression analysis and identified more than 1,000 unannotated transcribed regions. Visualizations of gene expression estimates, transcribed regions, and tiling probe measurements are accessible online at the At-TAX homepage.

ei

PDF DOI [BibTex]

PDF DOI [BibTex]


no image
Graphical Analysis of NMR Structural Quality and Interactive Contact Map of NOE Assignments in ARIA

Bardiaux, B., Bernard, A., Rieping, W., Habeck, M., Malliavin, T., Nilges, M.

BMC Structural Biology, 8(30):1-5, June 2008 (article)

Abstract
BACKGROUND: The Ambiguous Restraints for Iterative Assignment (ARIA) approach is widely used for NMR structure determination. It is based on simultaneously calculating structures and assigning NOE through an iterative protocol. The final solution consists of a set of conformers and a list of most probable assignments for the input NOE peak list. RESULTS: ARIA was extended with a series of graphical tools to facilitate a detailed analysis of the intermediate and final results of the ARIA protocol. These additional features provide (i) an interactive contact map, serving as a tool for the analysis of assignments, and (ii) graphical representations of structure quality scores and restraint statistics. The interactive contact map between residues can be clicked to obtain information about the restraints and their contributions. Profiles of quality scores are plotted along the protein sequence, and contact maps provide information of the agreement with the data on a residue pair level. CONCLUSIONS: The g raphical tools and outputs described here significantly extend the validation and analysis possibilities of NOE assignments given by ARIA as well as the analysis of the quality of the final structure ensemble. These tools are included in the latest version of ARIA, which is available at http://aria.pasteur.fr. The Web site also contains an installation guide, a user manual and example calculations.

ei

PDF DOI [BibTex]

PDF DOI [BibTex]


no image
Kernel Methods in Machine Learning

Hofmann, T., Schölkopf, B., Smola, A.

Annals of Statistics, 36(3):1171-1220, June 2008 (article)

Abstract
We review machine learning methods employing positive definite kernels. These methods formulate learning and estimation problems in a reproducing kernel Hilbert space (RKHS) of functions defined on the data domain, expanded in terms of a kernel. Working in linear spaces of function has the benefit of facilitating the construction and analysis of learning algorithms while at the same time allowing large classes of functions. The latter include nonlinear functions as well as functions defined on nonvectorial data.

ei

PDF PDF DOI [BibTex]

PDF PDF DOI [BibTex]


no image
Cross-validation Optimization for Large Scale Structured Classification Kernel Methods

Seeger, M.

Journal of Machine Learning Research, 9, pages: 1147-1178, June 2008 (article)

Abstract
We propose a highly efficient framework for penalized likelihood kernel methods applied to multi-class models with a large, structured set of classes. As opposed to many previous approaches which try to decompose the fitting problem into many smaller ones, we focus on a Newton optimization of the complete model, making use of model structure and linear conjugate gradients in order to approximate Newton search directions. Crucially, our learning method is based entirely on matrix-vector multiplication primitives with the kernel matrices and their derivatives, allowing straightforward specialization to new kernels, and focusing code optimization efforts to these primitives only. Kernel parameters are learned automatically, by maximizing the cross-validation log likelihood in a gradient-based way, and predictive probabilities are estimated. We demonstrate our approach on large scale text classification tasks with hierarchical structure on thousands of classes, achieving state-of-the-art results in an order of magnitude less time than previous work.

ei

PDF PDF [BibTex]

PDF PDF [BibTex]


no image
Reinforcement Learning of Motor Skills with Policy Gradients

Peters, J., Schaal, S.

Neural Networks, 21(4):682-697, May 2008 (article)

ei

PDF Web DOI [BibTex]