This paper addresses the problem of streaming packetized media

This paper proposes a new method for evaluating the quality of retrieval functions. Unlike traditional methods that require relevance judgments by experts or explicit user feedback, it is based entirely on clickthrough data. This is a key advantage, since clickthrough data can be collected at very low cost and without overhead for the user. Taking an approach from experiment design, the paper proposes an experiment setup that generates unbiased feedback about the relative quality of two search results without explicit user feedback. A theoretical analysis shows that the method gives the same results as evaluation with traditional relevance judgments under mild assumptions. An empirical analysis verifies that the assumptions are indeed justified and that the new method leads to conclusive results in a WWW retrieval study. 1

We propose novel algorithms for solving the so-called Support Vector Multiple Kernel Learning problem and show how they can be used to understand the resulting support vector decision function. While classical kernel-based algorithms (such as SVMs) are based on a single kernel, in Multiple Kernel Learning a quadraticallyconstraint quadratic program is solved in order to find a sparse convex combination of a set of support vector kernels. We show how this problem can be cast into a semi-infinite linear optimization problem which can in turn be solved efficiently using a boosting-like iterative method in combination with standard SVM optimization algorithms. The proposed method is able to deal with thousands of examples while combining hundreds of kernels within reasonable time. In the second part we show how this technique can be used to understand the obtained decision function in order to extract biologically relevant knowledge about the sequence analysis problem at hand. We consider the problem of splice site identification and combine string kernels at different sequence positions and with various substring (oligomer) lengths. The proposed algorithm computes a sparse weighting over the length and the substring, highlighting which substrings are important for discrimination. Finally, we propose a bootstrap scheme in order to reliably identify a few statistically significant positions, which can then be used for further analysis such as consensus finding.

In the present study, we introduce a simple iterative procedure that allows to correct the outputs of a classifier with respect to the new a priori probabilities of a new data set to be scored, even when these new a priori probabilities are unknown in advance. We also show that a significant increase in classification accuracy can be observed when using this procedure properly. More specifically, by applying the correcting procedure to the outputs of a simple logistic regression model, we observe an increase of 5.8% of classification rate on a di#cult real-world multi-class problem -- the automatic labeling of geographical maps based on remote sensing information.

roduction-based knowledge into the recognition framework. By using an explicit time-domain articulatory model of the mechanisms of co-articulation, it is hoped to obtain a more accurate model of contextual effects in the acoustic signal, while using fewer parameters than traditional acoustically-driven approaches. Separate articulatory and acoustic models are provided, and in each case the parameters of the models are automatically optimised over a training data set. A predictive statistically-based model of co-articulation is described, and found to yield improved articulatory modelling accuracy compared with X-ray articulatory traces. Parameterised acoustic vectors are synthesised by a set of artificial neural networks, and the resulting acoustic representations are used to re-score N-best recognition hypothesis lists produced by an HMM-based recogniser. The system is evaluated on two test databases, one including speaker-specific X-ray training data and the other aco

We look at a single point in the feature space, two classes, and L classiers estimating the posterior probability for class ! 1 . Assuming that the estimates are independent and identically distributed (normal or uniform), we give formulas for the classication error for the following fusion methods: average, minimum, maximum, median, majority vote and oracle. Keywords Classier combination, theoretical error, expert fusion, order statistics, majority vote, independent classiers. I. Introduction Classier combination has received considerable attention in the past decade and is now an established pattern recognition ospring. Recently, the focus has been shifting from practical heuristic solution of the combination problem towards explaining why combination methods and strategies work so well and in what cases some methods are better than others. Let D = fD 1 ; : : : ; DL g be a set (pool/committee/ensemble) of classiers, also regarded as \experts". By combining the individual outpu...

Internet-based telepresence and teleaction systems require packet-based transmission of haptic data and typically generate high packet rates between operator and teleoperator. This leads to the necessity of packet rate reduction techniques. The so-called deadband approach presented earlier by the authors uses a psychophysically motivated scheme based on Weber’s difference threshold (just noticable difference- JND) where force sample values are only transmitted if the change exceeds this threshold. This approach has been extended to three dimensions resulting in an additional perceptual domain- namely force direction. An experimental evaluation with human subjects was conducted in order to examine the change of the JND in 3D when force magnitude and force direction are combined. Our results show that the extension into dimensions yields to an increased JND in certain cases. Thus, higher compression ratios of haptic data and reduction in number of packets sent over the network can be reached. 1.

Abstract: This paper reports an investigation of some methods for isolating, or segmenting, characters during the reading of machineprinted text by optical character recognition systems. Two new segmentation algorithms using feature extraction techniques are presented; both are intended for use in the recognition of machine-printed lines of lo-, 11- and 12-pitch serif-type multifont characters. One of the methods, called quasi-topological segmentation, bases the decision to “section ” a character on a combination of featureextraction and character-width measurements. The other method, topological segmentation, involves feature extraction alone. The algorithms have been tested with an evaluation method that is independent of any particular recognition system. Test results are based on application of the algorithm to upper-case alphanumeric characters gathered from print sources that represent the existing world of machine printing. The topological approach demonstrated better performance on the test data than did the quasitopological approach.

Automated tests can play a key role in ensuring system quality in software development. However, significant problems arise in automating tests of stochastic algorithms. Normally, developers write tests that simply check whether the actual result is equal to the expected result (perhaps within some tolerance). But for stochastic algorithms, restricting ourselves in this way severely limits the kinds of tests we can write: either to trivial tests, or to fragile and hard-tounderstand tests that rely on a particular seed for a random number generator. A richer and more powerful set of tests is possible if we accommodate tests of statistical properties of the results of running an algorithm many times. The work described in this paper has been done in the context of a real-world application, a large-scale simulation of urban development designed to inform major decisions about land use and transportation. We describe our earlier experience with using automated testing for this system, in which we took a conventional approach, and the resulting difficulties. We then present a statistically based approach for testing stochastic algorithms based on hypothesis testing. Three different ways of constructing such tests are given, which cover the most commonly used distributions. We evaluate these tests in terms of frequency of failing when they should and when they should not, and conclude with guidelines and practical suggestions for implementing such unit tests for other stochastic applications. Categories and Subject Descriptors:

Whether or not an edge is significant for a particular application or situation depends on how the edge is defined and the environment that the edge is in. If an edge is used as the boundary between a defective region and normal region for visual surface inspection purposes, usually, it has a clear definition because of the explicitly stated requirements or specifications from the inspection tasks. However, meaningful examination of edges for the purpose of visual inspection requires a criterion that can be used to relate the characteristics of the edges in the image with the physical properties given in the inspection specifications. This paper proposes a method with which a candidate edge can be characterized and examined against a given condition. This criterion is arrived at by examining the distribution of grey level values of the pixels located in every individual edge neighborhood and the edge properties such as edge strength, edge length etc. In consideration of computational s...