Header logo is ei


2006


no image
Global Biclustering of Microarray Data

Wolf, T., Brors, B., Hofmann, T., Georgii, E.

In ICDMW 2006, pages: 125-129, (Editors: Tsumoto, S. , C. W. Clifton, N. Zhong, X. Wu, J. Liu, B. W. Wah, Y.-M. Cheung), IEEE Computer Society, Los Alamitos, CA, USA, Sixth IEEE International Conference on Data Mining, December 2006 (inproceedings)

Abstract
We consider the problem of simultaneously clustering genes and conditions of a gene expression data matrix. A bicluster is defined as a subset of genes that show similar behavior within a subset of conditions. Finding biclusters can be useful for revealing groups of genes involved in the same molecular process as well as groups of conditions where this process takes place. Previous work either deals with local, bicluster-based criteria or assumes a very specific structure of the data matrix (e.g. checkerboard or block-diagonal) [11]. In contrast, our goal is to find a set of flexibly arranged biclusters which is optimal in regard to a global objective function. As this is a NP-hard combinatorial problem, we describe several techniques to obtain approximate solutions. We benchmarked our approach successfully on the Alizadeh B-cell lymphoma data set [1].

Web DOI [BibTex]

2006

Web DOI [BibTex]


no image
Conformal Multi-Instance Kernels

Blaschko, M., Hofmann, T.

In NIPS 2006 Workshop on Learning to Compare Examples, pages: 1-6, NIPS Workshop on Learning to Compare Examples, December 2006 (inproceedings)

Abstract
In the multiple instance learning setting, each observation is a bag of feature vectors of which one or more vectors indicates membership in a class. The primary task is to identify if any vectors in the bag indicate class membership while ignoring vectors that do not. We describe here a kernel-based technique that defines a parametric family of kernels via conformal transformations and jointly learns a discriminant function over bags together with the optimal parameter settings of the kernel. Learning a conformal transformation effectively amounts to weighting regions in the feature space according to their contribution to classification accuracy; regions that are discriminative will be weighted higher than regions that are not. This allows the classifier to focus on regions contributing to classification accuracy while ignoring regions that correspond to vectors found both in positive and in negative bags. We show how parameters of this transformation can be learned for support vector machines by posing the problem as a multiple kernel learning problem. The resulting multiple instance classifier gives competitive accuracy for several multi-instance benchmark datasets from different domains.

PDF Web [BibTex]

PDF Web [BibTex]


no image
Some observations on the pedestal effect or dipper function

Henning, B., Wichmann, F.

Journal of Vision, 6(13):50, 2006 Fall Vision Meeting of the Optical Society of America, December 2006 (poster)

Abstract
The pedestal effect is the large improvement in the detectabilty of a sinusoidal “signal” grating observed when the signal is added to a masking or “pedestal” grating of the same spatial frequency, orientation, and phase. We measured the pedestal effect in both broadband and notched noise - noise from which a 1.5-octave band centred on the signal frequency had been removed. Although the pedestal effect persists in broadband noise, it almost disappears in the notched noise. Furthermore, the pedestal effect is substantial when either high- or low-pass masking noise is used. We conclude that the pedestal effect in the absence of notched noise results principally from the use of information derived from channels with peak sensitivities at spatial frequencies different from that of the signal and pedestal. The spatial-frequency components of the notched noise above and below the spatial frequency of the signal and pedestal prevent the use of information about changes in contrast carried in channels tuned to spatial frequencies that are very much different from that of the signal and pedestal. Thus the pedestal or dipper effect measured without notched noise is not a characteristic of individual spatial-frequency tuned channels.

Web DOI [BibTex]

Web DOI [BibTex]


no image
A New Projected Quasi-Newton Approach for the Nonnegative Least Squares Problem

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

(TR-06-54), Univ. of Texas, Austin, December 2006 (techreport)

PDF [BibTex]

PDF [BibTex]


no image
Information-theoretic Metric Learning

Davis, J., Kulis, B., Sra, S., Dhillon, I.

In NIPS 2006 Workshop on Learning to Compare Examples, pages: 1-5, NIPS Workshop on Learning to Compare Examples, December 2006 (inproceedings)

Abstract
We formulate the metric learning problem as that of minimizing the differential relative entropy between two multivariate Gaussians under constraints on the Mahalanobis distance function. Via a surprising equivalence, we show that this problem can be solved as a low-rank kernel learning problem. Specifically, we minimize the Burg divergence of a low-rank kernel to an input kernel, subject to pairwise distance constraints. Our approach has several advantages over existing methods. First, we present a natural information-theoretic formulation for the problem. Second, the algorithm utilizes the methods developed by Kulis et al. [6], which do not involve any eigenvector computation; in particular, the running time of our method is faster than most existing techniques. Third, the formulation offers insights into connections between metric learning and kernel learning.

PDF Web [BibTex]

PDF Web [BibTex]


no image
Pattern Mining in Frequent Dynamic Subgraphs

Borgwardt, KM., Kriegel, H-P., Wackersreuther, P.

In pages: 818-822, (Editors: Clifton, C.W.), IEEE Computer Society, Los Alamitos, CA, USA, Sixth International Conference on Data Mining (ICDM), December 2006 (inproceedings)

Abstract
Graph-structured data is becoming increasingly abundant in many application domains. Graph mining aims at finding interesting patterns within this data that represent novel knowledge. While current data mining deals with static graphs that do not change over time, coming years will see the advent of an increasing number of time series of graphs. In this article, we investigate how pattern mining on static graphs can be extended to time series of graphs. In particular, we are considering dynamic graphs with edge insertions and edge deletions over time. We define frequency in this setting and provide algorithmic solutions for finding frequent dynamic subgraph patterns. Existing subgraph mining algorithms can be easily integrated into our framework to make them handle dynamic graphs. Experimental results on real-world data confirm the practical feasibility of our approach.

Web DOI [BibTex]

Web DOI [BibTex]


no image
Probabilistic inference for solving (PO)MDPs

Toussaint, M., Harmeling, S., Storkey, A.

(934), School of Informatics, University of Edinburgh, December 2006 (techreport)

PDF [BibTex]

PDF [BibTex]


no image
Minimal Logical Constraint Covering Sets

Sinz, F., Schölkopf, B.

(155), Max Planck Institute for Biological Cybernetics, Tübingen, December 2006 (techreport)

Abstract
We propose a general framework for computing minimal set covers under class of certain logical constraints. The underlying idea is to transform the problem into a mathematical programm under linear constraints. In this sense it can be seen as a natural extension of the vector quantization algorithm proposed by Tipping and Schoelkopf. We show which class of logical constraints can be cast and relaxed into linear constraints and give an algorithm for the transformation.

PDF [BibTex]

PDF [BibTex]


no image
3DString: a feature string kernel for 3D object classification on voxelized data

Assfalg, J., Borgwardt, KM., Kriegel, H-P.

In pages: 198-207, (Editors: Yu, P.S. , V.J. Tsotras, E.A. Fox, B. Liu), ACM Press, New York, NY, USA, 15th ACM International Conference on Information and Knowledge Management (CIKM), November 2006 (inproceedings)

Abstract
Classification of 3D objects remains an important task in many areas of data management such as engineering, medicine or biology. As a common preprocessing step in current approaches to classification of voxelized 3D objects, voxel representations are transformed into a feature vector description.In this article, we introduce an approach of transforming 3D objects into feature strings which represent the distribution of voxels over the voxel grid. Attractively, this feature string extraction can be performed in linear runtime with respect to the number of voxels. We define a similarity measure on these feature strings that counts common k-mers in two input strings, which is referred to as the spectrum kernel in the field of kernel methods. We prove that on our feature strings, this similarity measure can be computed in time linear to the number of different characters in these strings. This linear runtime behavior makes our kernel attractive even for large datasets that occur in many application domains. Furthermore, we explain that our similarity measure induces a metric which allows to combine it with an M-tree for handling of large volumes of data. Classification experiments on two published benchmark datasets show that our novel approach is competitive with the best state-of-the-art methods for 3D object classification.

DOI [BibTex]

DOI [BibTex]


no image
Prediction of Protein Function from Networks

Shin, H., Tsuda, K.

In Semi-Supervised Learning, pages: 361-376, Adaptive Computation and Machine Learning, (Editors: Chapelle, O. , B. Schölkopf, A. Zien), MIT Press, Cambridge, MA, USA, November 2006 (inbook)

Abstract
In computational biology, it is common to represent domain knowledge using graphs. Frequently there exist multiple graphs for the same set of nodes, representing information from different sources, and no single graph is sufficient to predict class labels of unlabelled nodes reliably. One way to enhance reliability is to integrate multiple graphs, since individual graphs are partly independent and partly complementary to each other for prediction. In this chapter, we describe an algorithm to assign weights to multiple graphs within graph-based semi-supervised learning. Both predicting class labels and searching for weights for combining multiple graphs are formulated into one convex optimization problem. The graph-combining method is applied to functional class prediction of yeast proteins.When compared with individual graphs, the combined graph with optimized weights performs significantly better than any single graph.When compared with the semidefinite programming-based support vector machine (SDP/SVM), it shows comparable accuracy in a remarkably short time. Compared with a combined graph with equal-valued weights, our method could select important graphs without loss of accuracy, which implies the desirable property of integration with selectivity.

Web [BibTex]

Web [BibTex]


no image
Adapting Spatial Filter Methods for Nonstationary BCIs

Tomioka, R., Hill, J., Blankertz, B., Aihara, K.

In IBIS 2006, pages: 65-70, 2006 Workshop on Information-Based Induction Sciences, November 2006 (inproceedings)

Abstract
A major challenge in applying machine learning methods to Brain-Computer Interfaces (BCIs) is to overcome the possible nonstationarity in the data from the datablock the method is trained on and that the method is applied to. Assuming the joint distributions of the whitened signal and the class label to be identical in two blocks, where the whitening is done in each block independently, we propose a simple adaptation formula that is applicable to a broad class of spatial filtering methods including ICA, CSP, and logistic regression classifiers. We characterize the class of linear transformations for which the above assumption holds. Experimental results on 60 BCI datasets show improved classification accuracy compared to (a) fixed spatial filter approach (no adaptation) and (b) fixed spatial pattern approach (proposed by Hill et al., 2006 [1]).

PDF [BibTex]

PDF [BibTex]


no image
Discrete Regularization

Zhou, D., Schölkopf, B.

In Semi-supervised Learning, pages: 237-250, Adaptive computation and machine learning, (Editors: O Chapelle and B Schölkopf and A Zien), MIT Press, Cambridge, MA, USA, November 2006 (inbook)

Abstract
Many real-world machine learning problems are situated on finite discrete sets, including dimensionality reduction, clustering, and transductive inference. A variety of approaches for learning from finite sets has been proposed from different motivations and for different problems. In most of those approaches, a finite set is modeled as a graph, in which the edges encode pairwise relationships among the objects in the set. Consequently many concepts and methods from graph theory are adopted. In particular, the graph Laplacian is widely used. In this chapter we present a systemic framework for learning from a finite set represented as a graph. We develop discrete analogues of a number of differential operators, and then construct a discrete analogue of classical regularization theory based on those discrete differential operators. The graph Laplacian based approaches are special cases of this general discrete regularization framework. An important thing implied in this framework is that we have a wide choices of regularization on graph in addition to the widely-used graph Laplacian based one.

PDF Web [BibTex]

PDF Web [BibTex]


no image
New Methods for the P300 Visual Speller

Biessmann, F.

(1), (Editors: Hill, J. ), Max-Planck Institute for Biological Cybernetics, Tübingen, Germany, November 2006 (techreport)

PDF [BibTex]

PDF [BibTex]


no image
Optimizing Spatial Filters for BCI: Margin- and Evidence-Maximization Approaches

Farquhar, J., Hill, N., Schölkopf, B.

Challenging Brain-Computer Interfaces: MAIA Workshop 2006, pages: 1, November 2006 (poster)

Abstract
We present easy-to-use alternatives to the often-used two-stage Common Spatial Pattern + classifier approach for spatial filtering and classification of Event-Related Desychnronization signals in BCI. We report two algorithms that aim to optimize the spatial filters according to a criterion more directly related to the ability of the algorithms to generalize to unseen data. Both are based upon the idea of treating the spatial filter coefficients as hyperparameters of a kernel or covariance function. We then optimize these hyper-parameters directly along side the normal classifier parameters with respect to our chosen learning objective function. The two objectives considered are margin maximization as used in Support-Vector Machines and the evidence maximization framework used in Gaussian Processes. Our experiments assessed generalization error as a function of the number of training points used, on 9 BCI competition data sets and 5 offline motor imagery data sets measured in Tubingen. Both our approaches sho w consistent improvements relative to the commonly used CSP+linear classifier combination. Strikingly, the improvement is most significant in the higher noise cases, when either few trails are used for training, or with the most poorly performing subjects. This a reversal of the usual "rich get richer" effect in the development of CSP extensions, which tend to perform best when the signal is strong enough to accurately find their additional parameters. This makes our approach particularly suitable for clinical application where high levels of noise are to be expected.

PDF PDF [BibTex]

PDF PDF [BibTex]


no image
Geometric Analysis of Hilbert Schmidt Independence criterion based ICA contrast function

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

(PA006080), National ICT Australia, Canberra, Australia, October 2006 (techreport)

Web [BibTex]

Web [BibTex]


no image
A Linear Programming Approach for Molecular QSAR analysis

Saigo, H., Kadowaki, T., Tsuda, K.

In MLG 2006, pages: 85-96, (Editors: Gärtner, T. , G. C. Garriga, T. Meinl), International Workshop on Mining and Learning with Graphs, September 2006, Best Paper Award (inproceedings)

Abstract
Small molecules in chemistry can be represented as graphs. In a quantitative structure-activity relationship (QSAR) analysis, the central task is to find a regression function that predicts the activity of the molecule in high accuracy. Setting a QSAR as a primal target, we propose a new linear programming approach to the graph-based regression problem. Our method extends the graph classification algorithm by Kudo et al. (NIPS 2004), which is a combination of boosting and graph mining. Instead of sequential multiplicative updates, we employ the linear programming boosting (LP) for regression. The LP approach allows to include inequality constraints for the parameter vector, which turns out to be particularly useful in QSAR tasks where activity values are sometimes unavailable. Furthermore, the efficiency is improved significantly by employing multiple pricing.

PDF PDF [BibTex]

PDF PDF [BibTex]


no image
Incremental Aspect Models for Mining Document Streams

Surendran, A., Sra, S.

In PKDD 2006, pages: 633-640, (Editors: Fürnkranz, J. , T. Scheffer, M. Spiliopoulou), Springer, Berlin, Germany, 10th European Conference on Principles and Practice of Knowledge Discovery in Databases, September 2006 (inproceedings)

Abstract
In this paper we introduce a novel approach for incrementally building aspect models, and use it to dynamically discover underlying themes from document streams. Using the new approach we present an application which we call “query-line tracking” i.e., we automatically discover and summarize different themes or stories that appear over time, and that relate to a particular query. We present evaluation on news corpora to demonstrate the strength of our method for both query-line tracking, online indexing and clustering.

Web DOI [BibTex]

Web DOI [BibTex]


no image
Learning Eye Movements

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

Sensory Coding And The Natural Environment, 2006, pages: 1, September 2006 (poster)

Abstract
The human visual system samples images through saccadic eye movements which rapidly change the point of fixation. Although the selection of eye movement targets depends on numerous top-down mechanisms, a number of recent studies have shown that low-level image features such as local contrast or edges play an important role. These studies typically used predefined image features which were afterwards experimentally verified. Here, we follow a complementary approach: instead of testing a set of candidate image features, we infer these hypotheses from the data, using methods from statistical learning. To this end, we train a non-linear classifier on fixated vs. randomly selected image patches without making any physiological assumptions. The resulting classifier can be essentially characterized by a nonlinear combination of two center-surround receptive fields. We find that the prediction performance of this simple model on our eye movement data is indistinguishable from the physiologically motivated model of Itti & Koch (2000) which is far more complex. In particular, we obtain a comparable performance without using any multi-scale representations, long-range interactions or oriented image features.

Web [BibTex]

Web [BibTex]


no image
PALMA: Perfect Alignments using Large Margin Algorithms

Rätsch, G., Hepp, B., Schulze, U., Ong, C.

In GCB 2006, pages: 104-113, (Editors: Huson, D. , O. Kohlbacher, A. Lupas, K. Nieselt, A. Zell), Gesellschaft für Informatik, Bonn, Germany, German Conference on Bioinformatics, September 2006 (inproceedings)

Abstract
Despite many years of research on how to properly align sequences in the presence of sequencing errors, alternative splicing and micro-exons, the correct alignment of mRNA sequences to genomic DNA is still a challenging task. We present a novel approach based on large margin learning that combines kernel based splice site predictions with common sequence alignment techniques. By solving a convex optimization problem, our algorithm -- called PALMA -- tunes the parameters of the model such that the true alignment scores higher than all other alignments. In an experimental study on the alignments of mRNAs containing artificially generated micro-exons, we show that our algorithm drastically outperforms all other methods: It perfectly aligns all 4358 sequences on an hold-out set, while the best other method misaligns at least 90 of them. Moreover, our algorithm is very robust against noise in the query sequence: when deleting, inserting, or mutating up to 50% of the query sequence, it still aligns 95% of all sequences correctly, while other methods achieve less than 36% accuracy. For datasets, additional results and a stand-alone alignment tool see http://www.fml.mpg.de/raetsch/projects/palma.

PDF Web [BibTex]

PDF Web [BibTex]


no image
Graph Based Semi-Supervised Learning with Sharper Edges

Shin, H., Hill, N., Rätsch, G.

In ECML 2006, pages: 401-412, (Editors: Fürnkranz, J. , T. Scheffer, M. Spiliopoulou), Springer, Berlin, Germany, 17th European Conference on Machine Learning (ECML), September 2006 (inproceedings)

Abstract
In many graph-based semi-supervised learning algorithms, edge weights are assumed to be fixed and determined by the data points‘ (often symmetric)relationships in input space, without considering directionality. However, relationships may be more informative in one direction (e.g. from labelled to unlabelled) than in the reverse direction, and some relationships (e.g. strong weights between oppositely labelled points) are unhelpful in either direction. Undesirable edges may reduce the amount of influence an informative point can propagate to its neighbours -- the point and its outgoing edges have been ``blunted.‘‘ We present an approach to ``sharpening‘‘ in which weights are adjusted to meet an optimization criterion wherever they are directed towards labelled points. This principle can be applied to a wide variety of algorithms. In the current paper, we present one ad hoc solution satisfying the principle, in order to show that it can improve performance on a number of publicly available benchmark data sets.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Robust MEG Source Localization of Event Related Potentials: Identifying Relevant Sources by Non-Gaussianity

Breun, P., Grosse-Wentrup, M., Utschick, W., Buss, M.

In DAGM 2006, pages: 394-403, (Editors: Franke, K. , K.-R. Müller, B. Nickolay, R. Schäfer), Springer, Berlin, Germany, 28th Annual Symposium of the German Association for Pattern Recognition, September 2006 (inproceedings)

Abstract
Independent Component Analysis (ICA) is a frequently used preprocessing step in source localization of MEG and EEG data. By decomposing the measured data into maximally independent components (ICs), estimates of the time course and the topographies of neural sources are obtained. In this paper, we show that when using estimated source topographies for localization, correlations between neural sources introduce an error into the obtained source locations. This error can be avoided by reprojecting ICs onto the observation space, but requires the identification of relevant ICs. For Event Related Potentials (ERPs), we identify relevant ICs by estimating their non-Gaussianity. The efficacy of the approach is tested on auditory evoked potentials (AEPs) recorded by MEG. It is shown that ten trials are sufficient for reconstructing all important characteristics of the AEP, and source localization of the reconstructed ERP yields the same focus of activity as the average of 250 trials.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Finite-Horizon Optimal State-Feedback Control of Nonlinear Stochastic Systems Based on a Minimum Principle

Deisenroth, MP., Ohtsuka, T., Weissel, F., Brunn, D., Hanebeck, UD.

In MFI 2006, pages: 371-376, (Editors: Hanebeck, U. D.), IEEE Service Center, Piscataway, NJ, USA, 6th IEEE International Conference on Multisensor Fusion and Integration, September 2006 (inproceedings)

Abstract
In this paper, an approach to the finite-horizon optimal state-feedback control problem of nonlinear, stochastic, discrete-time systems is presented. Starting from the dynamic programming equation, the value function will be approximated by means of Taylor series expansion up to second-order derivatives. Moreover, the problem will be reformulated, such that a minimum principle can be applied to the stochastic problem. Employing this minimum principle, the optimal control problem can be rewritten as a two-point boundary-value problem to be solved at each time step of a shrinking horizon. To avoid numerical problems, the two-point boundary-value problem will be solved by means of a continuation method. Thus, the curse of dimensionality of dynamic programming is avoided, and good candidates for the optimal state-feedback controls are obtained. The proposed approach will be evaluated by means of a scalar example system.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Uniform Convergence of Adaptive Graph-Based Regularization

Hein, M.

In COLT 2006, pages: 50-64, (Editors: Lugosi, G. , H.-U. Simon), Springer, Berlin, Germany, 19th Annual Conference on Learning Theory, September 2006 (inproceedings)

Abstract
The regularization functional induced by the graph Laplacian of a random neighborhood graph based on the data is adaptive in two ways. First it adapts to an underlying manifold structure and second to the density of the data-generating probability measure. We identify in this paper the limit of the regularizer and show uniform convergence over the space of Hoelder functions. As an intermediate step we derive upper bounds on the covering numbers of Hoelder functions on compact Riemannian manifolds, which are of independent interest for the theoretical analysis of manifold-based learning methods.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Efficient Large Scale Linear Programming Support Vector Machines

Sra, S.

In ECML 2006, pages: 767-774, (Editors: Fürnkranz, J. , T. Scheffer, M. Spiliopoulou), Springer, Berlin, Germany, 17th European Conference on Machine Learning, September 2006 (inproceedings)

Abstract
This paper presents a decomposition method for efficiently constructing ℓ1-norm Support Vector Machines (SVMs). The decomposition algorithm introduced in this paper possesses many desirable properties. For example, it is provably convergent, scales well to large datasets, is easy to implement, and can be extended to handle support vector regression and other SVM variants. We demonstrate the efficiency of our algorithm by training on (dense) synthetic datasets of sizes up to 20 million points (in ℝ32). The results show our algorithm to be several orders of magnitude faster than a previously published method for the same task. We also present experimental results on real data sets—our method is seen to be not only very fast, but also highly competitive against the leading SVM implementations.

Web DOI [BibTex]

Web DOI [BibTex]


no image
Regularised CSP for Sensor Selection in BCI

Farquhar, J., Hill, N., Lal, T., Schölkopf, B.

In Proceedings of the 3rd International Brain-Computer Interface Workshop and Training Course 2006, pages: 14-15, (Editors: GR Müller-Putz and C Brunner and R Leeb and R Scherer and A Schlögl and S Wriessnegger and G Pfurtscheller), Verlag der Technischen Universität Graz, Graz, Austria, 3rd International Brain-Computer Interface Workshop and Training Course, September 2006 (inproceedings)

Abstract
The Common Spatial Pattern (CSP) algorithm is a highly successful method for efficiently calculating spatial filters for brain signal classification. Spatial filtering can improve classification performance considerably, but demands that a large number of electrodes be mounted, which is inconvenient in day-to-day BCI usage. The CSP algorithm is also known for its tendency to overfit, i.e. to learn the noise in the training set rather than the signal. Both problems motivate an approach in which spatial filters are sparsified. We briefly sketch a reformulation of the problem which allows us to do this, using 1-norm regularisation. Focusing on the electrode selection issue, we present preliminary results on EEG data sets that suggest that effective spatial filters may be computed with as few as 10--20 electrodes, hence offering the potential to simplify the practical realisation of BCI systems significantly.

PDF PDF [BibTex]

PDF PDF [BibTex]


no image
Time-Dependent Demixing of Task-Relevant EEG Signals

Hill, N., Farquhar, J., Lal, T., Schölkopf, B.

In Proceedings of the 3rd International Brain-Computer Interface Workshop and Training Course 2006, pages: 20-21, (Editors: GR Müller-Putz and C Brunner and R Leeb and R Scherer and A Schlögl and S Wriessnegger and G Pfurtscheller), Verlag der Technischen Universität Graz, Graz, Austria, 3rd International Brain-Computer Interface Workshop and Training Course, September 2006 (inproceedings)

Abstract
Given a spatial filtering algorithm that has allowed us to identify task-relevant EEG sources, we present a simple approach for monitoring the activity of these sources while remaining relatively robust to changes in other (task-irrelevant) brain activity. The idea is to keep spatial *patterns* fixed rather than spatial filters, when transferring from training to test sessions or from one time window to another. We show that a fixed spatial pattern (FSP) approach, using a moving-window estimate of signal covariances, can be more robust to non-stationarity than a fixed spatial filter (FSF) approach.

PDF PDF [BibTex]

PDF PDF [BibTex]


no image
Transductive Gaussian Process Regression with Automatic Model Selection

Le, Q., Smola, A., Gärtner, T., Altun, Y.

In Machine Learning: ECML 2006, pages: 306-317, (Editors: Fürnkranz, J. , T. Scheffer, M. Spiliopoulou), Springer, Berlin, Germany, 17th European Conference on Machine Learning (ECML), September 2006 (inproceedings)

Abstract
n contrast to the standard inductive inference setting of predictive machine learning, in real world learning problems often the test instances are already available at training time. Transductive inference tries to improve the predictive accuracy of learning algorithms by making use of the information contained in these test instances. Although this description of transductive inference applies to predictive learning problems in general, most transductive approaches consider the case of classification only. In this paper we introduce a transductive variant of Gaussian process regression with automatic model selection, based on approximate moment matching between training and test data. Empirical results show the feasibility and competitiveness of this approach.

Web DOI [BibTex]

Web DOI [BibTex]


no image
A Sober Look at Clustering Stability

Ben-David, S., von Luxburg, U., Pal, D.

In COLT 2006, pages: 5-19, (Editors: Lugosi, G. , H.-U. Simon), Springer, Berlin, Germany, 19th Annual Conference on Learning Theory, September 2006 (inproceedings)

Abstract
Stability is a common tool to verify the validity of sample based algorithms. In clustering it is widely used to tune the parameters of the algorithm, such as the number k of clusters. In spite of the popularity of stability in practical applications, there has been very little theoretical analysis of this notion. In this paper we provide a formal definition of stability and analyze some of its basic properties. Quite surprisingly, the conclusion of our analysis is that for large sample size, stability is fully determined by the behavior of the objective function which the clustering algorithm is aiming to minimize. If the objective function has a unique global minimizer, the algorithm is stable, otherwise it is unstable. In particular we conclude that stability is not a well-suited tool to determine the number of clusters - it is determined by the symmetries of the data which may be unrelated to clustering parameters. We prove our results for center-based clusterings and for spectral clustering, and support our conclusions by many examples in which the behavior of stability is counter-intuitive.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Information Marginalization on Subgraphs

Huang, J., Zhu, T., Rereiner, R., Zhou, D., Schuurmans, D.

In ECML/PKDD 2006, pages: 199-210, (Editors: Fürnkranz, J. , T. Scheffer, M. Spiliopoulou), Springer, Berlin, Germany, 10th European Conference on Principles and Practice of Knowledge Discovery in Databases, September 2006 (inproceedings)

Abstract
Real-world data often involves objects that exhibit multiple relationships; for example, ‘papers’ and ‘authors’ exhibit both paper-author interactions and paper-paper citation relationships. A typical learning problem requires one to make inferences about a subclass of objects (e.g. ‘papers’), while using the remaining objects and relations to provide relevant information. We present a simple, unified mechanism for incorporating information from multiple object types and relations when learning on a targeted subset. In this scheme, all sources of relevant information are marginalized onto the target subclass via random walks. We show that marginalized random walks can be used as a general technique for combining multiple sources of information in relational data. With this approach, we formulate new algorithms for transduction and ranking in relational data, and quantify the performance of new schemes on real world data—achieving good results in many problems.

Web DOI [BibTex]

Web DOI [BibTex]


no image
Bayesian Active Learning for Sensitivity Analysis

Pfingsten, T.

In ECML 2006, pages: 353-364, (Editors: Fürnkranz, J. , T. Scheffer, M. Spiliopoulou), Springer, Berlin, Germany, 17th European Conference on Machine Learning, September 2006 (inproceedings)

Abstract
Designs of micro electro-mechanical devices need to be robust against fluctuations in mass production. Computer experiments with tens of parameters are used to explore the behavior of the system, and to compute sensitivity measures as expectations over the input distribution. Monte Carlo methods are a simple approach to estimate these integrals, but they are infeasible when the models are computationally expensive. Using a Gaussian processes prior, expensive simulation runs can be saved. This Bayesian quadrature allows for an active selection of inputs where the simulation promises to be most valuable, and the number of simulation runs can be reduced further. We present an active learning scheme for sensitivity analysis which is rigorously derived from the corresponding Bayesian expected loss. On three fully featured, high dimensional physical models of electro-mechanical sensors, we show that the learning rate in the active learning scheme is significantly better than for passive learning.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
A tutorial on spectral clustering

von Luxburg, U.

(149), Max Planck Institute for Biological Cybernetics, Tübingen, August 2006 (techreport)

Abstract
In recent years, spectral clustering has become one of the most popular modern clustering algorithms. It is simple to implement, can be solved efficiently by standard linear algebra software, and very often outperforms traditional clustering algorithms such as the k-means algorithm. Nevertheless, on the first glance spectral clustering looks a bit mysterious, and it is not obvious to see why it works at all and what it really does. This article is a tutorial introduction to spectral clustering. We describe different graph Laplacians and their basic properties, present the most common spectral clustering algorithms, and derive those algorithms from scratch by several different approaches. Advantages and disadvantages of the different spectral clustering algorithms are discussed.

PDF [BibTex]

PDF [BibTex]


no image
Towards the Inference of Graphs on Ordered Vertexes

Zien, A., Raetsch, G., Ong, C.

(150), Max Planck Institute for Biological Cybernetics, Tübingen, August 2006 (techreport)

Abstract
We propose novel methods for machine learning of structured output spaces. Specifically, we consider outputs which are graphs with vertices that have a natural order. We consider the usual adjacency matrix representation of graphs, as well as two other representations for such a graph: (a) decomposing the graph into a set of paths, (b) converting the graph into a single sequence of nodes with labeled edges. For each of the three representations, we propose an encoding and decoding scheme. We also propose an evaluation measure for comparing two graphs.

PDF [BibTex]

PDF [BibTex]


no image
Semi-supervised Hyperspectral Image Classification with Graphs

Bandos, T., Zhou, D., Camps-Valls, G.

In IGARSS 2006, pages: 3883-3886, IEEE Computer Society, Los Alamitos, CA, USA, IEEE International Conference on Geoscience and Remote Sensing, August 2006 (inproceedings)

Abstract
This paper presents a semi-supervised graph-based method for the classification of hyperspectral images. The method is designed to exploit the spatial/contextual information in the images through composite kernels. The proposed method produces smoother classifications with respect to the intrinsic structure collectively revealed by known labeled and unlabeled points. Good accuracy in high dimensional spaces and low number of labeled samples (ill-posed situations) are produced as compared to standard inductive support vector machines.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Supervised Probabilistic Principal Component Analysis

Yu, S., Yu, K., Tresp, V., Kriegel, H., Wu, M.

In KDD 2006, pages: 464-473, (Editors: Ungar, L. ), ACM Press, New York, NY, USA, 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, August 2006 (inproceedings)

Abstract
Principal component analysis (PCA) has been extensively applied in data mining, pattern recognition and information retrieval for unsupervised dimensionality reduction. When labels of data are available, e.g.,~in a classification or regression task, PCA is however not able to use this information. The problem is more interesting if only part of the input data are labeled, i.e.,~in a semi-supervised setting. In this paper we propose a supervised PCA model called SPPCA and a semi-supervised PCA model called S$^2$PPCA, both of which are extensions of a probabilistic PCA model. The proposed models are able to incorporate the label information into the projection phase, and can naturally handle multiple outputs (i.e.,~in multi-task learning problems). We derive an efficient EM learning algorithm for both models, and also provide theoretical justifications of the model behaviors. SPPCA and S$^2$PPCA are compared with other supervised projection methods on various learning tasks, and show not only promising performance but also good scalability.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Broad-Coverage Sense Disambiguation and Information Extraction with a Supersense Sequence Tagger

Ciaramita, M., Altun, Y.

In pages: 594-602, (Editors: Jurafsky, D. , E. Gaussier), Association for Computational Linguistics, Stroudsburg, PA, USA, 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP), July 2006 (inproceedings)

Abstract
In this paper we approach word sense disambiguation and information extraction as a unified tagging problem. The task consists of annotating text with the tagset defined by the 41 Wordnet supersense classes for nouns and verbs. Since the tagset is directly related to Wordnet synsets, the tagger returns partial word sense disambiguation. Furthermore, since the noun tags include the standard named entity detection classes – person, location, organization, time, etc. – the tagger, as a by-product, returns extended named entity information. We cast the problem of supersense tagging as a sequential labeling task and investigate it empirically with a discriminatively-trained Hidden Markov Model. Experimental evaluation on the main sense-annotated datasets available, i.e., Semcor and Senseval, shows considerable improvements over the best known “first-sense” baseline.

Web [BibTex]

Web [BibTex]


no image
A Continuation Method for Semi-Supervised SVMs

Chapelle, O., Chi, M., Zien, A.

In ICML 2006, pages: 185-192, (Editors: Cohen, W. W., A. Moore), ACM Press, New York, NY, USA, 23rd International Conference on Machine Learning, June 2006 (inproceedings)

Abstract
Semi-Supervised Support Vector Machines (S3VMs) are an appealing method for using unlabeled data in classification: their objective function favors decision boundaries which do not cut clusters. However their main problem is that the optimization problem is non-convex and has many local minima, which often results in suboptimal performances. In this paper we propose to use a global optimization technique known as continuation to alleviate this problem. Compared to other algorithms minimizing the same objective function, our continuation method often leads to lower test errors.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Classification of natural scenes: Critical features revisited

Drewes, J., Wichmann, F., Gegenfurtner, K.

Journal of Vision, 6(6):561, 6th Annual Meeting of the Vision Sciences Society (VSS), June 2006 (poster)

Abstract
Human observers are capable of detecting animals within novel natural scenes with remarkable speed and accuracy. Despite the seeming complexity of such decisions it has been hypothesized that a simple global image feature, the relative abundance of high spatial frequencies at certain orientations, could underly such fast image classification (A. Torralba & A. Oliva, Network: Comput. Neural Syst., 2003). We successfully used linear discriminant analysis to classify a set of 11.000 images into “animal” and “non-animal” images based on their individual amplitude spectra only (Drewes, Wichmann, Gegenfurtner VSS 2005). We proceeded to sort the images based on the performance of our classifier, retaining only the best and worst classified 400 images (“best animals”, “best distractors” and “worst animals”, “worst distractors”). We used a Go/No-go paradigm to evaluate human performance on this subset of our images. Both reaction time and proportion of correctly classified images showed a significant effect of classification difficulty. Images more easily classified by our algorithm were also classified faster and better by humans, as predicted by the Torralba & Oliva hypothesis. We then equated the amplitude spectra of the 400 images, which, by design, reduced algorithmic performance to chance whereas human performance was only slightly reduced (cf. Wichmann, Rosas, Gegenfurtner, VSS 2005). Most importantly, the same images as before were still classified better and faster, suggesting that even in the original condition features other than specifics of the amplitude spectrum made particular images easy to classify, clearly at odds with the Torralba & Oliva hypothesis.

Web DOI [BibTex]

Web DOI [BibTex]


no image
Unifying Divergence Minimization and Statistical Inference Via Convex Duality

Altun, Y., Smola, A.

In Learning Theory, pages: 139-153, (Editors: Lugosi, G. , H.-U. Simon), Springer, Berlin, Germany, 19th Annual Conference on Learning Theory (COLT), June 2006 (inproceedings)

Abstract
In this paper we unify divergence minimization and statistical inference by means of convex duality. In the process of doing so, we prove that the dual of approximate maximum entropy estimation is maximum a posteriori estimation as a special case. Moreover, our treatment leads to stability and convergence bounds for many statistical learning problems. Finally, we show how an algorithm by Zhang can be used to solve this class of optimization problems efficiently.

Web DOI [BibTex]

Web DOI [BibTex]


no image
Trading Convexity for Scalability

Collobert, R., Sinz, F., Weston, J., Bottou, L.

In ICML 2006, pages: 201-208, (Editors: Cohen, W. W., A. Moore), ACM Press, New York, NY, USA, 23rd International Conference on Machine Learning, June 2006 (inproceedings)

Abstract
Convex learning algorithms, such as Support Vector Machines (SVMs), are often seen as highly desirable because they offer strong practical properties and are amenable to theoretical analysis. However, in this work we show how non-convexity can provide scalability advantages over convexity. We show how concave-convex programming can be applied to produce (i) faster SVMs where training errors are no longer support vectors, and (ii) much faster Transductive SVMs.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Personalized handwriting recognition via biased regularization

Kienzle, W., Chellapilla, K.

In ICML 2006, pages: 457-464, (Editors: Cohen, W. W., A. Moore), ACM Press, New York, NY, USA, 23rd International Conference on Machine Learning, June 2006 (inproceedings)

Abstract
We present a new approach to personalized handwriting recognition. The problem, also known as writer adaptation, consists of converting a generic (user-independent) recognizer into a personalized (user-dependent) one, which has an improved recognition rate for a particular user. The adaptation step usually involves user-specific samples, which leads to the fundamental question of how to fuse this new information with that captured by the generic recognizer. We propose adapting the recognizer by minimizing a regularized risk functional (a modified SVM) where the prior knowledge from the generic recognizer enters through a modified regularization term. The result is a simple personalization framework with very good practical properties. Experiments on a 100 class real-world data set show that the number of errors can be reduced by over 40% with as few as five user samples per character.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Deterministic annealing for semi-supervised kernel machines

Sindhwani, V., Keerthi, S., Chapelle, O.

In ICML 2006, pages: 841-848, (Editors: Cohen, W. W., A. Moore), ACM Press, New York, NY, USA, 23rd International Conference on Machine Learning, June 2006 (inproceedings)

Abstract
An intuitive approach to utilizing unlabeled data in kernel-based classification algorithms is to simply treat the unknown labels as additional optimization variables. For margin-based loss functions, one can view this approach as attempting to learn low-density separators. However, this is a hard optimization problem to solve in typical semi-supervised settings where unlabeled data is abundant. The popular Transductive SVM algorithm is a label-switching-retraining procedure that is known to be susceptible to local minima. In this paper, we present a global optimization framework for semi-supervised Kernel machines where an easier problem is parametrically deformed to the original hard problem and minimizers are smoothly tracked. Our approach is motivated from deterministic annealing techniques and involves a sequence of convex optimization problems that are exactly and efficiently solved. We present empirical results on several synthetic and real world datasets that demonstrate the effectiveness of our approach.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Clustering Graphs by Weighted Substructure Mining

Tsuda, K., Kudo, T.

In ICML 2006, pages: 953-960, (Editors: Cohen, W. W., A. Moore), ACM Press, New York, NY, USA, 23rd International Conference on Machine Learning, June 2006 (inproceedings)

Abstract
Graph data is getting increasingly popular in, e.g., bioinformatics and text processing. A main difficulty of graph data processing lies in the intrinsic high dimensionality of graphs, namely, when a graph is represented as a binary feature vector of indicators of all possible subgraphs, the dimensionality gets too large for usual statistical methods. We propose an efficient method for learning a binomial mixture model in this feature space. Combining the $ell_1$ regularizer and the data structure called DFS code tree, the MAP estimate of non-zero parameters are computed efficiently by means of the EM algorithm. Our method is applied to the clustering of RNA graphs, and is compared favorably with graph kernels and the spectral graph distance.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
A Choice Model with Infinitely Many Latent Features

Görür, D., Jäkel, F., Rasmussen, C.

In ICML 2006, pages: 361-368, (Editors: Cohen, W. W., A. Moore), ACM Press, New York, NY, USA, 23rd International Conference on Machine Learning, June 2006 (inproceedings)

Abstract
Elimination by aspects (EBA) is a probabilistic choice model describing how humans decide between several options. The options from which the choice is made are characterized by binary features and associated weights. For instance, when choosing which mobile phone to buy the features to consider may be: long lasting battery, color screen, etc. Existing methods for inferring the parameters of the model assume pre-specified features. However, the features that lead to the observed choices are not always known. Here, we present a non-parametric Bayesian model to infer the features of the options and the corresponding weights from choice data. We use the Indian buffet process (IBP) as a prior over the features. Inference using Markov chain Monte Carlo (MCMC) in conjugate IBP models has been previously described. The main contribution of this paper is an MCMC algorithm for the EBA model that can also be used in inference for other non-conjugate IBP models---this may broaden the use of IBP priors considerably.

PostScript PDF Web DOI [BibTex]

PostScript PDF Web DOI [BibTex]


no image
Learning High-Order MRF Priors of Color Images

McAuley, J., Caetano, T., Smola, A., Franz, MO.

In ICML 2006, pages: 617-624, (Editors: Cohen, W. W., A. Moore), ACM Press, New York, NY, USA, 23rd International Conference on Machine Learning, June 2006 (inproceedings)

Abstract
In this paper, we use large neighborhood Markov random fields to learn rich prior models of color images. Our approach extends the monochromatic Fields of Experts model (Roth and Blackwell, 2005) to color images. In the Fields of Experts model, the curse of dimensionality due to very large clique sizes is circumvented by parameterizing the potential functions according to a product of experts. We introduce several simplifications of the original approach by Roth and Black which allow us to cope with the increased clique size (typically 3x3x3 or 5x5x3 pixels) of color images. Experimental results are presented for image denoising which evidence improvements over state-of-the-art monochromatic image priors.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
Inference with the Universum

Weston, J., Collobert, R., Sinz, F., Bottou, L., Vapnik, V.

In ICML 2006, pages: 1009-1016, (Editors: Cohen, W. W., A. Moore), ACM Press, New York, NY, USA, 23rd International Conference on Machine Learning, June 2006 (inproceedings)

Abstract
WIn this paper we study a new framework introduced by Vapnik (1998) and Vapnik (2006) that is an alternative capacity concept to the large margin approach. In the particular case of binary classification, we are given a set of labeled examples, and a collection of "non-examples" that do not belong to either class of interest. This collection, called the Universum, allows one to encode prior knowledge by representing meaningful concepts in the same domain as the problem at hand. We describe an algorithm to leverage the Universum by maximizing the number of observed contradictions, and show experimentally that this approach delivers accuracy improvements over using labeled data alone.

PDF Web DOI [BibTex]

PDF Web DOI [BibTex]


no image
The pedestal effect is caused by off-frequency looking, not nonlinear transduction or contrast gain-control

Wichmann, F., Henning, B.

Journal of Vision, 6(6):194, 6th Annual Meeting of the Vision Sciences Society (VSS), June 2006 (poster)

Abstract
The pedestal or dipper effect is the large improvement in the detectabilty of a sinusoidal grating observed when the signal is added to a pedestal or masking grating having the signal‘s spatial frequency, orientation, and phase. The effect is largest with pedestal contrasts just above the ‘threshold‘ in the absence of a pedestal. We measured the pedestal effect in both broadband and notched masking noise---noise from which a 1.5- octave band centered on the signal and pedestal frequency had been removed. The pedestal effect persists in broadband noise, but almost disappears with notched noise. The spatial-frequency components of the notched noise that lie above and below the spatial frequency of the signal and pedestal prevent the use of information about changes in contrast carried in channels tuned to spatial frequencies that are very much different from that of the signal and pedestal. We conclude that the pedestal effect in the absence of notched noise results principally from the use of information derived from channels with peak sensitivities at spatial frequencies that are different from that of the signal and pedestal. Thus the pedestal or dipper effect is not a characteristic of individual spatial-frequency tuned channels.

Web DOI [BibTex]

Web DOI [BibTex]


no image
Statistical Convergence of Kernel CCA

Fukumizu, K., Bach, F., Gretton, A.

In Advances in neural information processing systems 18, pages: 387-394, (Editors: Weiss, Y. , B. Schölkopf, J. Platt), MIT Press, Cambridge, MA, USA, Nineteenth Annual Conference on Neural Information Processing Systems (NIPS), May 2006 (inproceedings)

Abstract
While kernel canonical correlation analysis (kernel CCA) has been applied in many problems, the asymptotic convergence of the functions estimated from a finite sample to the true functions has not yet been established. This paper gives a rigorous proof of the statistical convergence of kernel CCA and a related method (NOCCO), which provides a theoretical justification for these methods. The result also gives a sufficient condition on the decay of the regularization coefficient in the methods to ensure convergence.

PDF Web [BibTex]

PDF Web [BibTex]


no image
Maximum Margin Semi-Supervised Learning for Structured Variables

Altun, Y., McAllester, D., Belkin, M.

In Advances in neural information processing systems 18, pages: 33-40, (Editors: Weiss, Y. , B. Schölkopf, J. Platt), MIT Press, Cambridge, MA, USA, Nineteenth Annual Conference on Neural Information Processing Systems (NIPS), May 2006 (inproceedings)

Abstract
Many real-world classification problems involve the prediction of multiple inter-dependent variables forming some structural dependency. Recent progress in machine learning has mainly focused on supervised classification of such structured variables. In this paper, we investigate structured classification in a semi-supervised setting. We present a discriminative approach that utilizes the intrinsic geometry of input patterns revealed by unlabeled data points and we derive a maximum-margin formulation of semi-supervised learning for structured variables. Unlike transductive algorithms, our formulation naturally extends to new test points.

PDF Web [BibTex]

PDF Web [BibTex]


no image
Generalized Nonnegative Matrix Approximations with Bregman Divergences

Dhillon, I., Sra, S.

In Advances in neural information processing systems 18, pages: 283-290, (Editors: Weiss, Y. , B. Schölkopf, J. Platt), MIT Press, Cambridge, MA, USA, Nineteenth Annual Conference on Neural Information Processing Systems (NIPS), May 2006 (inproceedings)

Abstract
Nonnegative matrix approximation (NNMA) is a recent technique for dimensionality reduction and data analysis that yields a parts based, sparse nonnegative representation for nonnegative input data. NNMA has found a wide variety of applications, including text analysis, document clustering, face/image recognition, language modeling, speech processing and many others. Despite these numerous applications, the algorithmic development for computing the NNMA factors has been relatively efficient. This paper makes algorithmic progress by modeling and solving (using multiplicative updates) new generalized NNMA problems that minimize Bregman divergences between the input matrix and its lowrank approximation. The multiplicative update formulae in the pioneering work by Lee and Seung [11] arise as a special case of our algorithms. In addition, the paper shows how to use penalty functions for incorporating constraints other than nonnegativity into the problem. Further, some interesting extensions to the use of "link" functions for modeling nonlinear relationships are also discussed.

PDF Web [BibTex]

PDF Web [BibTex]


no image
Fast Gaussian Process Regression using KD-Trees

Shen, Y., Ng, A., Seeger, M.

In Advances in neural information processing systems 18, pages: 1225-1232, (Editors: Weiss, Y. , B. Schölkopf, J. Platt), MIT Press, Cambridge, MA, USA, Nineteenth Annual Conference on Neural Information Processing Systems (NIPS), May 2006 (inproceedings)

Abstract
The computation required for Gaussian process regression with n training examples is about O(n3) during training and O(n) for each prediction. This makes Gaussian process regression too slow for large datasets. In this paper, we present a fast approximation method, based on kd-trees, that significantly reduces both the prediction and the training times of Gaussian process regression.

PDF Web [BibTex]

PDF Web [BibTex]