Minimum description length, regularisation and multi-modal data

Relationships between clustering, description length, and regularisation are pointed out, motivating the introduction of a cost function with a description length interpretation and the unusual and useful property of having its minimum approximated by the densest mode of a distribution. A simple inverse kinematics example is used to demonstrate that this property can be used to select and learn one branch of a multi-valued mapping. This property is also used to develop a method for setting regularisation parameters according to the scale on which structure is exhibited in the training data. The regularisation technique is demonstrated on two real data sets, a classification problem and a regression problem.

Minimum description length, regularisation and multi-modal data

Conventional feed forward Neural Networks have used the sum-of-squares cost function for training. A new cost function is presented here with a description length interpretation based on Rissanen's Minimum Description Length principle. It is a heuristic that has a rough interpretation as the number of data points fit by the model. Not concerned with finding optimal descriptions, the cost function prefers to form minimum descriptions in a naive way for computational convenience. The cost function is called the Naive Description Length cost function. Finding minimum description models will be shown to be closely related to the identification of clusters in the data. As a consequence the minimum of this cost function approximates the most probable mode of the data rather than the sum-of-squares cost function that approximates the mean. The new cost function is shown to provide information about the structure of the data. This is done by inspecting the dependence of the error to the amount of regularisation. This structure provides a method of selecting regularisation parameters as an alternative or supplement to Bayesian methods. The new cost function is tested on a number of multi-valued problems such as a simple inverse kinematics problem. It is also tested on a number of classification and regression problems. The mode-seeking property of this cost function is shown to improve prediction in time series problems. Description length principles are used in a similar fashion to derive a regulariser to control network complexity.

Minimum Description Length Principle in Discriminating Marginal Distributions

2010 Mathematics Subject Classification: 94A17, 62B10, 62F03.

Categorical data clustering using a minimum message length criterion

Research on cluster analysis for categorical data continues to develop, new clustering algorithms being proposed. However, in this context, the determination of the number of clusters is rarely addressed. We propose a new approach in which clustering and the estimation of the number of clusters is done simultaneously for categorical data. We assume that the data originate from a finite mixture of multinomial distributions and use a minimum message length criterion (MML) to select the number of clusters (Wallace and Bolton, 1986). For this purpose, we implement an EM-type algorithm (Silvestre et al., 2008) based on the (Figueiredo and Jain, 2002) approach. The novelty of the approach rests on the integration of the model estimation and selection of the number of clusters in a single algorithm, rather than selecting this number based on a set of pre-estimated candidate models. The performance of our approach is compared with the use of Bayesian Information Criterion (BIC) (Schwarz, 1978) and Integrated Completed Likelihood (ICL) (Biernacki et al., 2000) using synthetic data. The obtained results illustrate the capacity of the proposed algorithm to attain the true number of cluster while outperforming BIC and ICL since it is faster, which is especially relevant when dealing with large data sets.

Automated image analysis of lung branching morphogenesis from microscopic images of fetal rat explants

Background: Regulating mechanisms of branching morphogenesis of fetal lung rat explants have been an essential tool for molecular research. This work presents a new methodology to accurately quantify the epithelial, outer contour and peripheral airway buds of lung explants during cellular development from microscopic images. Methods: The outer contour was defined using an adaptive and multi-scale threshold algorithm whose level was automatically calculated based on an entropy maximization criterion. The inner lung epithelial was defined by a clustering procedure that groups small image regions according to the minimum description length principle and local statistical properties. Finally, the number of peripheral buds were counted as the skeleton branched ends from a skeletonized image of the lung inner epithelial. Results: The time for lung branching morphometric analysis was reduced in 98% in contrast to the manual method. Best results were obtained in the first two days of cellular development, with lesser standard deviations. Non-significant differences were found between the automatic and manual results in all culture days. Conclusions: The proposed method introduces a series of advantages related to its intuitive use and accuracy, making the technique suitable to images with different lightning characteristics and allowing a reliable comparison between different researchers.

Automated image analysis of lung branching morphogenesis

Regulating mechanisms of branchingmorphogenesis of fetal lung rat explants have been an essential tool formolecular research.This work presents a new methodology to accurately quantify the epithelial, outer contour, and peripheral airway buds of lung explants during cellular development frommicroscopic images. Methods.Theouter contour was defined using an adaptive and multiscale threshold algorithm whose level was automatically calculated based on an entropy maximization criterion. The inner lung epithelium was defined by a clustering procedure that groups small image regions according to the minimum description length principle and local statistical properties. Finally, the number of peripheral buds was counted as the skeleton branched ends from a skeletonized image of the lung inner epithelia. Results. The time for lung branching morphometric analysis was reduced in 98% in contrast to themanualmethod. Best results were obtained in the first two days of cellular development, with lesser standard deviations. Nonsignificant differences were found between the automatic and manual results in all culture days. Conclusions. The proposed method introduces a series of advantages related to its intuitive use and accuracy, making the technique suitable to images with different lighting characteristics and allowing a reliable comparison between different researchers.

Preliminary clinical trial in percutaneous nephrolithotomy using a real-time navigation system for percutaneous kidney access

Background: Precise needle puncture of renal calyces is a challenging and essential step for successful percutaneous nephrolithotomy. This work tests and evaluates, through a clinical trial, a real-time navigation system to plan and guide percutaneous kidney puncture. Methods: A novel system, entitled i3DPuncture, was developed to aid surgeons in establishing the desired puncture site and the best virtual puncture trajectory, by gathering and processing data from a tracked needle with optical passive markers. In order to navigate and superimpose the needle to a preoperative volume, the patient, 3D image data and tracker system were previously registered intraoperatively using seven points that were strategically chosen based on rigid bone structures and nearby kidney area. In addition, relevant anatomical structures for surgical navigation were automatically segmented using a multi-organ segmentation algorithm that clusters volumes based on statistical properties and minimum description length criterion. For each cluster, a rendering transfer function enhanced the visualization of different organs and surrounding tissues. Results: One puncture attempt was sufficient to achieve a successful kidney puncture. The puncture took 265 seconds, and 32 seconds were necessary to plan the puncture trajectory. The virtual puncture path was followed correctively until the needle tip reached the desired kidney calyceal. Conclusions: This new solution provided spatial information regarding the needle inside the body and the possibility to visualize surrounding organs. It may offer a promising and innovative solution for percutaneous punctures.

Learning dependent sources using mixtures of Dirichlet: applications on hyperspectral unmixing

This paper is an elaboration of the DECA algorithm [1] to blindly unmix hyperspectral data. The underlying mixing model is linear, meaning that each pixel is a linear mixture of the endmembers signatures weighted by the correspondent abundance fractions. The proposed method, as DECA, is tailored to highly mixed mixtures in which the geometric based approaches fail to identify the simplex of minimum volume enclosing the observed spectral vectors. We resort then to a statitistical framework, where the abundance fractions are modeled as mixtures of Dirichlet densities, thus enforcing the constraints on abundance fractions imposed by the acquisition process, namely non-negativity and constant sum. With respect to DECA, we introduce two improvements: 1) the number of Dirichlet modes are inferred based on the minimum description length (MDL) principle; 2) The generalized expectation maximization (GEM) algorithm we adopt to infer the model parameters is improved by using alternating minimization and augmented Lagrangian methods to compute the mixing matrix. The effectiveness of the proposed algorithm is illustrated with simulated and read data.

Hyperspectral unmixing based on mixtures of Dirichlet components

This paper introduces a new unsupervised hyperspectral unmixing method conceived to linear but highly mixed hyperspectral data sets, in which the simplex of minimum volume, usually estimated by the purely geometrically based algorithms, is far way from the true simplex associated with the endmembers. The proposed method, an extension of our previous studies, resorts to the statistical framework. The abundance fraction prior is a mixture of Dirichlet densities, thus automatically enforcing the constraints on the abundance fractions imposed by the acquisition process, namely, nonnegativity and sum-to-one. A cyclic minimization algorithm is developed where the following are observed: 1) The number of Dirichlet modes is inferred based on the minimum description length principle; 2) a generalized expectation maximization algorithm is derived to infer the model parameters; and 3) a sequence of augmented Lagrangian-based optimizations is used to compute the signatures of the endmembers. Experiments on simulated and real data are presented to show the effectiveness of the proposed algorithm in unmixing problems beyond the reach of the geometrically based state-of-the-art competitors.

Unmixing hyperspectral data: independent and dependent component analysis

The development of high spatial resolution airborne and spaceborne sensors has improved the capability of ground-based data collection in the fields of agriculture, geography, geology, mineral identification, detection [2, 3], and classification [4–8]. The signal read by the sensor from a given spatial element of resolution and at a given spectral band is a mixing of components originated by the constituent substances, termed endmembers, located at that element of resolution. This chapter addresses hyperspectral unmixing, which is the decomposition of the pixel spectra into a collection of constituent spectra, or spectral signatures, and their corresponding fractional abundances indicating the proportion of each endmember present in the pixel [9, 10]. Depending on the mixing scales at each pixel, the observed mixture is either linear or nonlinear [11, 12]. The linear mixing model holds when the mixing scale is macroscopic [13]. The nonlinear model holds when the mixing scale is microscopic (i.e., intimate mixtures) [14, 15]. The linear model assumes negligible interaction among distinct endmembers [16, 17]. The nonlinear model assumes that incident solar radiation is scattered by the scene through multiple bounces involving several endmembers [18]. Under the linear mixing model and assuming that the number of endmembers and their spectral signatures are known, hyperspectral unmixing is a linear problem, which can be addressed, for example, under the maximum likelihood setup [19], the constrained least-squares approach [20], the spectral signature matching [21], the spectral angle mapper [22], and the subspace projection methods [20, 23, 24]. Orthogonal subspace projection [23] reduces the data dimensionality, suppresses undesired spectral signatures, and detects the presence of a spectral signature of interest. The basic concept is to project each pixel onto a subspace that is orthogonal to the undesired signatures. As shown in Settle [19], the orthogonal subspace projection technique is equivalent to the maximum likelihood estimator. This projection technique was extended by three unconstrained least-squares approaches [24] (signature space orthogonal projection, oblique subspace projection, target signature space orthogonal projection). Other works using maximum a posteriori probability (MAP) framework [25] and projection pursuit [26, 27] have also been applied to hyperspectral data. In most cases the number of endmembers and their signatures are not known. Independent component analysis (ICA) is an unsupervised source separation process that has been applied with success to blind source separation, to feature extraction, and to unsupervised recognition [28, 29]. ICA consists in finding a linear decomposition of observed data yielding statistically independent components. Given that hyperspectral data are, in given circumstances, linear mixtures, ICA comes to mind as a possible tool to unmix this class of data. In fact, the application of ICA to hyperspectral data has been proposed in reference 30, where endmember signatures are treated as sources and the mixing matrix is composed by the abundance fractions, and in references 9, 25, and 31–38, where sources are the abundance fractions of each endmember. In the first approach, we face two problems: (1) The number of samples are limited to the number of channels and (2) the process of pixel selection, playing the role of mixed sources, is not straightforward. In the second approach, ICA is based on the assumption of mutually independent sources, which is not the case of hyperspectral data, since the sum of the abundance fractions is constant, implying dependence among abundances. This dependence compromises ICA applicability to hyperspectral images. In addition, hyperspectral data are immersed in noise, which degrades the ICA performance. IFA [39] was introduced as a method for recovering independent hidden sources from their observed noisy mixtures. IFA implements two steps. First, source densities and noise covariance are estimated from the observed data by maximum likelihood. Second, sources are reconstructed by an optimal nonlinear estimator. Although IFA is a well-suited technique to unmix independent sources under noisy observations, the dependence among abundance fractions in hyperspectral imagery compromises, as in the ICA case, the IFA performance. Considering the linear mixing model, hyperspectral observations are in a simplex whose vertices correspond to the endmembers. Several approaches [40–43] have exploited this geometric feature of hyperspectral mixtures [42]. Minimum volume transform (MVT) algorithm [43] determines the simplex of minimum volume containing the data. The MVT-type approaches are complex from the computational point of view. Usually, these algorithms first find the convex hull defined by the observed data and then fit a minimum volume simplex to it. Aiming at a lower computational complexity, some algorithms such as the vertex component analysis (VCA) [44], the pixel purity index (PPI) [42], and the N-FINDR [45] still find the minimum volume simplex containing the data cloud, but they assume the presence in the data of at least one pure pixel of each endmember. This is a strong requisite that may not hold in some data sets. In any case, these algorithms find the set of most pure pixels in the data. Hyperspectral sensors collects spatial images over many narrow contiguous bands, yielding large amounts of data. For this reason, very often, the processing of hyperspectral data, included unmixing, is preceded by a dimensionality reduction step to reduce computational complexity and to improve the signal-to-noise ratio (SNR). Principal component analysis (PCA) [46], maximum noise fraction (MNF) [47], and singular value decomposition (SVD) [48] are three well-known projection techniques widely used in remote sensing in general and in unmixing in particular. The newly introduced method [49] exploits the structure of hyperspectral mixtures, namely the fact that spectral vectors are nonnegative. The computational complexity associated with these techniques is an obstacle to real-time implementations. To overcome this problem, band selection [50] and non-statistical [51] algorithms have been introduced. This chapter addresses hyperspectral data source dependence and its impact on ICA and IFA performances. The study consider simulated and real data and is based on mutual information minimization. Hyperspectral observations are described by a generative model. This model takes into account the degradation mechanisms normally found in hyperspectral applications—namely, signature variability [52–54], abundance constraints, topography modulation, and system noise. The computation of mutual information is based on fitting mixtures of Gaussians (MOG) to data. The MOG parameters (number of components, means, covariances, and weights) are inferred using the minimum description length (MDL) based algorithm [55]. We study the behavior of the mutual information as a function of the unmixing matrix. The conclusion is that the unmixing matrix minimizing the mutual information might be very far from the true one. Nevertheless, some abundance fractions might be well separated, mainly in the presence of strong signature variability, a large number of endmembers, and high SNR. We end this chapter by sketching a new methodology to blindly unmix hyperspectral data, where abundance fractions are modeled as a mixture of Dirichlet sources. This model enforces positivity and constant sum sources (full additivity) constraints. The mixing matrix is inferred by an expectation-maximization (EM)-type algorithm. This approach is in the vein of references 39 and 56, replacing independent sources represented by MOG with mixture of Dirichlet sources. Compared with the geometric-based approaches, the advantage of this model is that there is no need to have pure pixels in the observations. The chapter is organized as follows. Section 6.2 presents a spectral radiance model and formulates the spectral unmixing as a linear problem accounting for abundance constraints, signature variability, topography modulation, and system noise. Section 6.3 presents a brief resume of ICA and IFA algorithms. Section 6.4 illustrates the performance of IFA and of some well-known ICA algorithms with experimental data. Section 6.5 studies the ICA and IFA limitations in unmixing hyperspectral data. Section 6.6 presents results of ICA based on real data. Section 6.7 describes the new blind unmixing scheme and some illustrative examples. Section 6.8 concludes with some remarks.

Hyperspectral imagery framework for unmixing and dimensionality estimation

In hyperspectral imagery a pixel typically consists mixture of spectral signatures of reference substances, also called endmembers. Linear spectral mixture analysis, or linear unmixing, aims at estimating the number of endmembers, their spectral signatures, and their abundance fractions. This paper proposes a framework for hyperpsectral unmixing. A blind method (SISAL) is used for the estimation of the unknown endmember signature and their abundance fractions. This method solve a non-convex problem by a sequence of augmented Lagrangian optimizations, where the positivity constraints, forcing the spectral vectors to belong to the convex hull of the endmember signatures, are replaced by soft constraints. The proposed framework simultaneously estimates the number of endmembers present in the hyperspectral image by an algorithm based on the minimum description length (MDL) principle. Experimental results on both synthetic and real hyperspectral data demonstrate the effectiveness of the proposed algorithm.

Spectrum sensing through software defined radio

Dissertação apresentada para obtenção do Grau de Mestre em Engenharia Electrotécnica e de Computadores, pela Universidade Nova de Lisboa, Faculdade de Ciências e Tecnologia

Adaptive model selection using empirical complexities

Given $n$ independent replicates of a jointly distributed pair $(X,Y)\in {\cal R}^d \times {\cal R}$, we wish to select from a fixed sequence of model classes ${\cal F}_1, {\cal F}_2, \ldots$ a deterministic prediction rule $f: {\cal R}^d \to {\cal R}$ whose risk is small. We investigate the possibility of empirically assessingthe {\em complexity} of each model class, that is, the actual difficulty of the estimation problem within each class. The estimated complexities are in turn used to define an adaptive model selection procedure, which is based on complexity penalized empirical risk.The available data are divided into two parts. The first is used to form an empirical cover of each model class, and the second is used to select a candidate rule from each cover based on empirical risk. The covering radii are determined empirically to optimize a tight upper bound on the estimation error. An estimate is chosen from the list of candidates in order to minimize the sum of class complexity and empirical risk. A distinguishing feature of the approach is that the complexity of each model class is assessed empirically, based on the size of its empirical cover.Finite sample performance bounds are established for the estimates, and these bounds are applied to several non-parametric estimation problems. The estimates are shown to achieve a favorable tradeoff between approximation and estimation error, and to perform as well as if the distribution-dependent complexities of the model classes were known beforehand. In addition, it is shown that the estimate can be consistent,and even possess near optimal rates of convergence, when each model class has an infinite VC or pseudo dimension.For regression estimation with squared loss we modify our estimate to achieve a faster rate of convergence.

Apprentissage automatique de la morphologie : le cas des structures racine-schème

RÉSUMÉ Cette thèse porte sur le développement de méthodes algorithmiques pour découvrir automatiquement la structure morphologique des mots d'un corpus. On considère en particulier le cas des langues s'approchant du type introflexionnel, comme l'arabe ou l'hébreu. La tradition linguistique décrit la morphologie de ces langues en termes d'unités discontinues : les racines consonantiques et les schèmes vocaliques. Ce genre de structure constitue un défi pour les systèmes actuels d'apprentissage automatique, qui opèrent généralement avec des unités continues. La stratégie adoptée ici consiste à traiter le problème comme une séquence de deux sous-problèmes. Le premier est d'ordre phonologique : il s'agit de diviser les symboles (phonèmes, lettres) du corpus en deux groupes correspondant autant que possible aux consonnes et voyelles phonétiques. Le second est de nature morphologique et repose sur les résultats du premier : il s'agit d'établir l'inventaire des racines et schèmes du corpus et de déterminer leurs règles de combinaison. On examine la portée et les limites d'une approche basée sur deux hypothèses : (i) la distinction entre consonnes et voyelles peut être inférée sur la base de leur tendance à alterner dans la chaîne parlée; (ii) les racines et les schèmes peuvent être identifiés respectivement aux séquences de consonnes et voyelles découvertes précédemment. L'algorithme proposé utilise une méthode purement distributionnelle pour partitionner les symboles du corpus. Puis il applique des principes analogiques pour identifier un ensemble de candidats sérieux au titre de racine ou de schème, et pour élargir progressivement cet ensemble. Cette extension est soumise à une procédure d'évaluation basée sur le principe de la longueur de description minimale, dans- l'esprit de LINGUISTICA (Goldsmith, 2001). L'algorithme est implémenté sous la forme d'un programme informatique nommé ARABICA, et évalué sur un corpus de noms arabes, du point de vue de sa capacité à décrire le système du pluriel. Cette étude montre que des structures linguistiques complexes peuvent être découvertes en ne faisant qu'un minimum d'hypothèses a priori sur les phénomènes considérés. Elle illustre la synergie possible entre des mécanismes d'apprentissage portant sur des niveaux de description linguistique distincts, et cherche à déterminer quand et pourquoi cette coopération échoue. Elle conclut que la tension entre l'universalité de la distinction consonnes-voyelles et la spécificité de la structuration racine-schème est cruciale pour expliquer les forces et les faiblesses d'une telle approche. ABSTRACT This dissertation is concerned with the development of algorithmic methods for the unsupervised learning of natural language morphology, using a symbolically transcribed wordlist. It focuses on the case of languages approaching the introflectional type, such as Arabic or Hebrew. The morphology of such languages is traditionally described in terms of discontinuous units: consonantal roots and vocalic patterns. Inferring this kind of structure is a challenging task for current unsupervised learning systems, which generally operate with continuous units. In this study, the problem of learning root-and-pattern morphology is divided into a phonological and a morphological subproblem. The phonological component of the analysis seeks to partition the symbols of a corpus (phonemes, letters) into two subsets that correspond well with the phonetic definition of consonants and vowels; building around this result, the morphological component attempts to establish the list of roots and patterns in the corpus, and to infer the rules that govern their combinations. We assess the extent to which this can be done on the basis of two hypotheses: (i) the distinction between consonants and vowels can be learned by observing their tendency to alternate in speech; (ii) roots and patterns can be identified as sequences of the previously discovered consonants and vowels respectively. The proposed algorithm uses a purely distributional method for partitioning symbols. Then it applies analogical principles to identify a preliminary set of reliable roots and patterns, and gradually enlarge it. This extension process is guided by an evaluation procedure based on the minimum description length principle, in line with the approach to morphological learning embodied in LINGUISTICA (Goldsmith, 2001). The algorithm is implemented as a computer program named ARABICA; it is evaluated with regard to its ability to account for the system of plural formation in a corpus of Arabic nouns. This thesis shows that complex linguistic structures can be discovered without recourse to a rich set of a priori hypotheses about the phenomena under consideration. It illustrates the possible synergy between learning mechanisms operating at distinct levels of linguistic description, and attempts to determine where and why such a cooperation fails. It concludes that the tension between the universality of the consonant-vowel distinction and the specificity of root-and-pattern structure is crucial for understanding the advantages and weaknesses of this approach.

Hyperspectral unmixing with simultaneous dimensionality estimation

This paper is an elaboration of the simplex identification via split augmented Lagrangian (SISAL) algorithm (Bioucas-Dias, 2009) to blindly unmix hyperspectral data. SISAL is a linear hyperspectral unmixing method of the minimum volume class. This method solve a non-convex problem by a sequence of augmented Lagrangian optimizations, where the positivity constraints, forcing the spectral vectors to belong to the convex hull of the endmember signatures, are replaced by soft constraints. With respect to SISAL, we introduce a dimensionality estimation method based on the minimum description length (MDL) principle. The effectiveness of the proposed algorithm is illustrated with simulated and real data.