The Minimum Description Length (MDL) principle states that the optimal model for a given data set is that which compresses it best. Due to practial limitations the model can be restricted to a class such as linear regression models, which we address in this study. As in other formulations such as the LASSO and forward step-wise regression we are interested in sparsifying the feature set while preserving generalization ability. We derive a well-principled set of codes for both parameters and error residuals along with smooth approximations to lengths of these codes as to allow gradient descent optimization of description length, and go on to show that sparsification and feature selection using our approach is faster than the LASSO on several datasets from the UCI and StatLib repositories, with favorable generalization accuracy, while being fully automatic, requiring neither cross-validation nor tuning of regularization hyper-parameters, allowing even for a nonlinear expansion of the feature set followed by sparsification.

The article is about AI research in the context of a complex artistic behavior: expressive music performance. A computer program is presented that learns to play piano with 'expression' and that even won an international computer piano performance contest. A superficial analysis of an expressive performance generated by the system seems to suggest creative musical abilities. After a critical discussion of the processes underlying this behavior, we abandon the question of whether the system is really creative, and turn to the true motivation that drives this research: to use AI methods to investigate and better understand music performance as a human creative behavior. A number of recent and current results from our research are briefly presented that indicate that machines can give us interesting insights into such a complex creative behavior, even if they may not be creative themselves.

Medical Informatics and the application of modern signal processing in the assistance of the diagnostic process in medical imaging is one of the more recent and active research areas today. This thesis addresses a variety of issues related to the general problem of medical image analysis, specifically in mammography, and presents a series of algorithms and design approaches for all the intermediate levels of a modern system for computer-aided diagnosis (CAD). The diagnostic problem is analyzed with a systematic approach, first defining the imaging characteristics and features that are relevant to probable pathology in mammo-grams. Next, these features are quantified and fused into new, integrated radio-logical systems that exhibit embedded digital signal processing, in order to improve the final result and minimize the radiological dose for the patient. In a higher level, special algorithms are designed for detecting and encoding these clinically interest-ing imaging features, in order to be used as input to advanced pattern classifiers and machine learning models. Finally, these approaches are extended in multi-classifier models under the scope of Game Theory and optimum collective deci-sion, in order to produce efficient solutions for combining classifiers with minimum computational costs for advanced diagnostic systems. The material covered in this thesis is related to a total of 18 published papers, 6 in scientific journals and 12 in international conferences.

We study the estimation of $\beta$ for the nonlinear model $y = f(X\sp{\top}\beta) + \epsilon$ when $f$ is a nonlinear transformation that is known, $\beta$ has sparse nonzero coordinates, and the number of observations can be much smaller than that of parameters ($n\ll p$). We show that in order to bound the $L_2$ error of the $L_0$ regularized estimator $\hat\beta$, i.e., $\|\hat\beta - \beta\|_2$, it is sufficient to establish two conditions. Based on this, we obtain bounds of the $L_2$ error for (1) $L_0$ regularized maximum likelihood estimation (MLE) for exponential linear models and (2) $L_0$ regularized least square (LS) regression for the more general case where $f$ is analytic. For the analytic case, we rely on power series expansion of $f$, which requires taking into account the singularities of $f$.

This paper reviews the functional aspects of statistical learning theory. The main point under consideration is the nature of the hypothesis set when no prior information is available but data. Within this framework we first discuss about the hypothesis set: it is a vectorial space, it is a set of pointwise defined functions, and the evaluation functional on this set is a continuous mapping. Based on these principles an original theory is developed generalizing the notion of reproduction kernel Hilbert space to non hilbertian sets. Then it is shown that the hypothesis set of any learning machine has to be a generalized reproducing set. Therefore, thanks to a general "representer theorem", the solution of the learning problem is still a linear combination of a kernel. Furthermore, a way to design these kernels is given. To illustrate this framework some examples of such reproducing sets and kernels are given.

We present here an introduction to Brainstorming approach, that was recently proposed as a consensus meta-learning technique, and used in several practical applications in bioinformatics and chemoinformatics. The consensus learning denotes heterogeneous theoretical classification method, where one trains an ensemble of machine learning algorithms using different types of input training data representations. In the second step all solutions are gathered and the consensus is build between them. Therefore no early solution, given even by a generally low performing algorithm, is not discarder until the late phase of prediction, when the final conclusion is drawn by comparing different machine learning models. This final phase, i.e. consensus learning, is trying to balance the generality of solution and the overall performance of trained model.

The Lasso is a very well known penalized regression model, which adds an $L_{1}$ penalty with parameter $\lambda_{1}$ on the coefficients to the squared error loss function. The Fused Lasso extends this model by also putting an $L_{1}$ penalty with parameter $\lambda_{2}$ on the difference of neighboring coefficients, assuming there is a natural ordering. In this paper, we develop a fast path algorithm for solving the Fused Lasso Signal Approximator that computes the solutions for all values of $\lambda_1$ and $\lambda_2$. In the supplement, we also give an algorithm for the general Fused Lasso for the case with predictor matrix $\bX \in \mathds{R}^{n \times p}$ with $\text{rank}(\bX)=p$.

Machine learning, data mining and artificial intelligence (AI) based methods have been used to determine the relations between chemical structure and biological activity, called quantitative structure activity relationships (QSARs) for the compounds. Pre-processing of the dataset, which includes the mapping from a large number of molecular descriptors in the original high dimensional space to a small number of components in the lower dimensional space while retaining the features of the original data, is the first step in this process. A common practice is to use a mapping method for a dataset without prior analysis. This pre-analysis has been stressed in our work by applying it to two important classes of QSAR prediction problems: drug design (predicting anti-HIV-1 activity) and predictive toxicology (estimating hepatocarcinogenicity of chemicals). We apply one linear and two nonlinear mapping methods on each of the datasets. Based on this analysis, we conclude the nature of the inherent relationships between the elements of each dataset, and hence, the mapping method best suited for it. We also show that proper preprocessing can help us in choosing the right feature extraction tool as well as give an insight about the type of classifier pertinent for the given problem.

In the last few years, due to the growing ubiquity of unlabeled data, much effort has been spent by the machine learning community to develop better understanding and improve the quality of classifiers exploiting unlabeled data. Following the manifold regularization approach, Laplacian Support Vector Machines (LapSVMs) have shown the state of the art performance in semi--supervised classification. In this paper we present two strategies to solve the primal LapSVM problem, in order to overcome some issues of the original dual formulation. Whereas training a LapSVM in the dual requires two steps, using the primal form allows us to collapse training to a single step. Moreover, the computational complexity of the training algorithm is reduced from O(n^3) to O(n^2) using preconditioned conjugate gradient, where n is the combined number of labeled and unlabeled examples. We speed up training by using an early stopping strategy based on the prediction on unlabeled data or, if available, on labeled validation examples. This allows the algorithm to quickly compute approximate solutions with roughly the same classification accuracy as the optimal ones, considerably reducing the training time. Due to its simplicity, training LapSVM in the primal can be the starting point for additional enhancements of the original LapSVM formulation, such as those for dealing with large datasets. We present an extensive experimental evaluation on real world data showing the benefits of the proposed approach.

This paper proposes an original approach to cluster multi-component data sets, including an estimation of the number of clusters. From the construction of a minimal spanning tree with Prim's algorithm, and the assumption that the vertices are approximately distributed according to a Poisson distribution, the number of clusters is estimated by thresholding the Prim's trajectory. The corresponding cluster centroids are then computed in order to initialize the generalized Lloyd's algorithm, also known as $K$-means, which allows to circumvent initialization problems. Some results are derived for evaluating the false positive rate of our cluster detection algorithm, with the help of approximations relevant in Euclidean spaces. Metrics used for measuring similarity between multi-dimensional data points are based on symmetrical divergences. The use of these informational divergences together with the proposed method leads to better results, compared to other clustering methods for the problem of astrophysical data processing. Some applications of this method in the multi/hyper-spectral imagery domain to a satellite view of Paris and to an image of the Mars planet are also presented. In order to demonstrate the usefulness of divergences in our problem, the method with informational divergence as similarity measure is compared with the same method using classical metrics. In the astrophysics application, we also compare the method with the spectral clustering algorithms.