Abstract

Abstract
In this paper we propose a new method to perform incremental discretization. The basic idea is to perform the task in two layers. The first layer receives the sequence of input data and keeps some, statistics on the data using many more intervals than required. Based on the statistics stored by the first layer, the second layer creates the final discretization. The proposed architecture processes streaming examples in a single scan, in constant time and space even for infinite sequences of examples. We experimentally demonstrate that incremental discretization is able to maintain the performance of learning algorithms in comparison to a batch discretization. The proposed method is much more appropriate in incremental learning, and in problems where data flows continuously, as in most of the recent data mining applications. Copyright 2006 ACM.

Abstract
Methods for learning sparse classification are among the state-of-the-art in supervised learning. Sparsity, essential to achieve good generalization capabilities, can be enforced by using heavy tailed priors/regularizers on the weights of the linear combination of functions. These priors/regularizers favour a few large weights and many to exactly zero. The Sparse Multinomial Logistic Regression algorithm [1] is one of such methods, that adopts a Laplacian prior to enforce sparseness. Its applicability to large datasets is still a delicate task from the computational point of view, sometimes even impossible to perform. This work implements an iterative procedure to calculate the weights of the decision function that is O(m(2)) faster than the original method introduced in [1] (m is the number of classes). The benchmark dataset Indian Pines is used to test this modification. Results over subsets of this dataset are presented and compared with others computed with support vector machines.

Abstract
Document classification has already been widely studied. In fact, some studies compared feature selection techniques or feature space transformation whereas some others compared the performance of different algorithms. Recently, following the rising interest towards the Support Vector Machine, various studies showed that SVM outperforms other classification algorithms. So should we just not bother about other classification algorithms and opt always for SVM ? We have decided to investigate this issue and compared SVM to kNN and naive Bayes on binary classification tasks. An important issue is to compare optimized versions of these algorithms, which is what we have done. Our results show all the classifiers achieved comparable performance on most problems. One surprising result is that SVM was not a clear winner, despite quite good overall performance. If a suitable preprocessing is used with kNN, this algorithm continues to achieve very good results and scales up well with the number of documents, which is not the case for SVM. As for naive Bayes, it also achieved good performance.

Abstract
In this paper we present a management model to deal with the problem of tracking a large number of features during long image sequences. Some usual difficulties are related to this problem: features may be temporarily occluded or might even have disappeared definitively; the computational cost involved should always be reduced to the strictly necessary. The proposed Net Present Value (NPV) model, based on the economic Theory of Capital, considers the tracking of each missing feature as an investment. Thus, using the NPV criterion, with adequate receipt and outlay functions, each occluded feature may be kept on tracking or it may be excluded of the tracking process depending on its historical behavior. This methodology may be applied to any tracking system as long as the tracking results may be evaluated in each temporal step. Experimental results, both on synthetic and real image sequences, which validate our model, will be also presented.

Abstract
Long-period gratings were written using arc discharges in pristine nitrogen doped fibres. Comparison of the resulting gratings spectra show that the resonant positions of gratings in fibres pre-annealed at 1050 degrees C for 30 min are shifted towards shorter wavelengths and their coupling strength is considerably higher. The refractive index and residual stress profiles were measured before and after annealing, as well as the two-dimensional stress profiles inside the grating region. Arc discharges induce periodic and strong asymmetric stresses in the fibre core of the pre-annealed nitrogen doped fibre.