Line Katrine Harder Clemmensen

Sparse Discriminant Analysis

We consider the problem of performing interpretable classification in the high-dimensional setting, in which the number of features is very large and the number of observations is limited. This setting has been studied extensively in the chemometrics literature, and more recently has become commonplace in biological and medical applications. In this setting, a traditional approach involves performing feature selection before classification. We propose sparse discriminant analysis, a method for performing linear discriminant analysis with a sparseness criterion imposed such that classification and feature selection are performed simultaneously. Sparse discriminant analysis is based on the optimal scoring interpretation of linear discriminant analysis, and can be extended to perform sparse discrimination via mixtures of Gaussians if boundaries between classes are nonlinear or if subgroups are present within each class. Our proposal also provides low-dimensional views of the discriminative directions.

Adverse effects on sexual development in rat offspring after low dose exposure to a mixture of endocrine disrupting pesticides.

The present study investigated whether a mixture of low doses of five environmentally relevant endocrine disrupting pesticides, epoxiconazole, mancozeb, prochloraz, tebuconazole and procymidone, would cause adverse developmental toxicity effects in rats. In rat dams, a significant increase in gestation length was seen, while in male offspring increased nipple retention and increased incidence and severity of genital malformations were observed. Severe mixture effects on gestation length, nipple retention and genital malformations were seen at dose levels where the individual pesticides caused no or smaller effects when given alone. Generally, the mixture effect predictions based on dose-additivity were in good agreement with the observed effects. The results indicate that there is a need for modification of risk assessment procedures for pesticides, in order to take account of the mixture effects and cumulative intake, because of the potentially serious impact of mixed exposure on development and reproduction in humans.

Effects of network resolution on topological properties of human neocortex

Graph theoretical analyses applied to neuroimaging datasets have provided valuable insights into the large-scale anatomical organization of the human neocortex. Most of these studies were performed with different cortical scales leading to cortical networks with different levels of small-world organization. The present study investigates how resolution of thickness-based cortical scales impacts on topological properties of human anatomical cortical networks. To this end, we designed a novel approach aimed at determining the best trade-off between small-world attributes of anatomical cortical networks and the number of cortical regions included in the scale. Results revealed that schemes comprising 540-599 regions (surface areas spanning between 250 and 275 mm(2)) at sparsities below 10% showed a superior balance between small-world organization and the size of the cortical scale employed. Furthermore, we found that the cortical scale representing the best trade-off (599 regions) was more resilient to targeted attacks than atlas-based schemes (Desikan-Killiany atlas, 66 regions) and, most importantly, it did not differ that much from the finest cortical scale tested in the present study (1494 regions). In summary, our study confirms that topological organization of anatomical cortical networks varies with both sparsity and resolution of cortical scale, and it further provides a novel methodological framework aimed at identifying cortical schemes that maximize small-worldness with the lowest scale resolution possible.

Precise acquisition and unsupervised segmentation of multi-spectral images

In this work, an integrated imaging system to obtain accurate and reproducible multi-spectral images and a novel multi-spectral image segmentation algorithm are proposed. The system collects up to 20 different spectral bands within a range that vary from 395nm to 970nm. The system is designed to acquire geometrically and chromatically corrected images in homogeneous and diffuse illumination, so images can be compared over time. The proposed segmentation algorithm combines the information provided by all the spectral bands to segment the different regions of interest. Three experiments are conducted to show the ability of the system to acquire highly precise, reproducible and standardized multi-spectral images and to show its applicabilities in different situations.

Supervised feature selection for linear and non-linear regression of L*a*b* color from multispectral images of meat

In food quality monitoring, color is an important indicator factor of quality. The CIELab (L*a*b*) color space as a device independent color space is an appropriate means in this case. The commonly used colorimeter instruments can neither measure the L*a*b color in a wide area over the target surface nor in a contact-less mode. However, developing algorithms for conversion of food items images into L*a*b color space can solve both of these issues. This paper addresses the problem of L*a*b color prediction from multispectral images of different types of raw meat. The efficiency of using multispectral images instead of the standard RGB is investigated. In addition, it is demonstrated that due to the fiber structure and transparency of raw meat, the prediction models built on the standard color patches do not work for raw meat test samples. As a result, multispectral images of different types of meat samples (430-970nm) were used for training and testing of the L*a*b prediction models. Finding a sparse solution or the use of a minimum number of bands is of particular interest to make an industrial vision set-up simpler and cost effective. In this paper, a wide range of linear, non-linear, kernel-based regression and sparse regression methods are compared. In order to improve the prediction results of these models, we propose a supervised feature selection strategy which is compared with the Principal component analysis (PCA) as a pre-processing step. The results showed that the proposed feature selection method outperforms the PCA for both linear and non-linear methods. The highest performance was obtained by linear ridge regression applied on the selected features from the proposed Elastic net (EN) -based feature selection strategy. All the best models use a reduced number of wavelengths for each of the L*a*b components.

A comparison of dimension reduction methods with application to multi-spectral images of sand used in concrete

This paper presents a comparison of dimension reduction methods based on a novel machine vision application for estimating moisture content in sand used to make concrete. For the application in question it is very important to know the moisture content of the sand so as to ensure good-quality concrete. In order to achieve a continuous in-line approach for the concrete mixing, digital image analysis is used. Multi-spectral images, consisting of nine spectral bands in the visible and near infrared (NIR) range, were acquired. Each image consists of approximately 9 million pixels. Five different sand types were examined with 20–60 images for each type. To reduce the amount of data, features were extracted from the multi-spectral images; the features were summary statistics on single bands and pairs of bands as well as morphological summaries. The number of features (2,016) is high in relation to the number of observations and, therefore, dimension reductive methods are needed. Furthermore, speed, which is an important consideration, is aided by the use of a small number of variables. On top of that, fewer dimensions tend to give more robust results. Two traditional statistical methods for dimension reduction (forward selection and principal components) combined with ordinary least squares and one sophisticated chemometrics algorithm (genetic algorithm-partial least squares) are compared to the recently proposed least angle regression-elastic net (LARS-EN) model selection method.

Regularized generalized eigen-decomposition with applications to sparse supervised feature extraction and sparse discriminant analysis

We propose a general technique for obtaining sparse solutions to generalized eigenvalue problems, and call it Regularized Generalized Eigen-Decomposition (RGED). For decades, Fishers discriminant criterion has been applied in supervised feature extraction and discriminant analysis, and it is formulated as a generalized eigenvalue problem. Thus RGED can be applied to effectively extract sparse features and calculate sparse discriminant directions for all variants of Fisher discriminant criterion based models. Particularly, RGED can be applied to matrix-based and even tensor-based discriminant techniques, for instance, 2D-Linear Discriminant Analysis (2D-LDA). Furthermore, an iterative algorithm based on the alternating direction method of multipliers is developed. The algorithm approximately solves RGED with monotonically decreasing convergence and at an acceptable speed for results of modest accuracy. Numerical experiments based on four data sets of different types of images show that RGED has competitive classification performance with existing multidimensional and sparse techniques of discriminant analysis. HighlightsWe propose a new technique called Regularized Generalized Eigen Decomposition (RGED).RGED solves generalized eigenvalue problems and obtains sparse solutions.It is easy and straightforward applying RGED to sparse discriminant analysis and feature extraction.An algorithm is developed to solve it with monotonically decreasing convergence.RGED has competitive classification performance comparing with other methods.

Pig herd monitoring and undesirable tripping and stepping prevention

Optical flow is used to monitor pig herd movement.Modified angular histograms are used to summarize the OF vectors.The abnormal pig movement is detected while unloading from a truck.The detection is performed using support vector machines.Moving pig measures are used to correct lens and foreshortening distortions. Humane handling and slaughter of livestock are of major concern in modern societies. Monitoring animal wellbeing in slaughterhouses is critical in preventing unnecessary stress and physical damage to livestock, which can also affect the meat quality. The goal of this study is to monitor pig herds at the slaughterhouse and identify undesirable events such as pigs tripping or stepping on each other. In this paper, we monitor pig behavior in color videos recorded during unloading from transportation trucks. We monitor the movement of a pig herd where the pigs enter and leave a surveyed area. The method is based on optical flow, which is not well explored for monitoring all types of animals, but is the method of choice for human crowd monitoring. We recommend using modified angular histograms to summarize the optical flow vectors. We show that the classification rate based on support vector machines is 93% of all frames. The sensitivity of the model is 93.5% with 90% specificity and 6.5% false alarm rate. The radial lens distortion and camera position required for convenient surveillance make the recordings highly distorted. Therefore, we also propose a new approach to correct lens and foreshortening distortions by using moving reference points. The method can be applied real-time during the actual unloading operations of pigs. In addition, we present a method for identification of the causes leading to undesirable events, which currently only runs off-line. The comparative analysis of three drivers, which performed the unloading of the pigs from the trucks in the available datasets, indicates that the drivers perform significantly differently. Driver 1 has 2.95 times higher odds to have pigs tripping and stepping on each other than the two others, and Driver 2 has 1.11 times higher odds than Driver 3.

Multiplicative Updates for the Lasso

Multiplicative updates have proven useful for non-negativity constrained optimization. Presently, we demonstrate how multiplicative updates also can be used for unconstrained optimization. This is for instance useful when estimating the least absolute shrinkage and selection operator (LASSO) i.e. least squares minimization with L1-norm regularization, since the multiplicative updates (MU) can efficiently exploit the structure of the problem traditionally solved using quadratic programming (QP). We derive two algorithms based on MU for the LASSO and compare the performance to Matlabs standard QP solver as well as the basis pursuit denoising algorithm (BP) which can be obtained from www.sparselab.stanford.edu. The algorithms were tested on three benchmark bio-informatic datasets: A small scale data set where the number of observations is larger than the number of variables estimated (M < J) and two large scale microarray data sets (M Gt J). For small scale data the two MU algorithms, QP and BP give identical results while the time used is more or less of the same order. However, for large scale problems QP is unstable and slow. Both algorithms based on MU on the other hand are stable and faster but not as efficient as the BP algorithm and converge slowly for small regularizations. The benefit of the present MU algorithms is that they are easy to implement, they bridge multiplicative updates to unconstrained optimization and the updates derived monotonically decrease the cost-function thus does not need any objective function evaluation. Finally, both MU are potentially useful for a wide range of other models such as the elastic net or the fused LASSO. The Matlab implementations of the LASSO based on MU can be downloaded.

In silico modelling of permeation enhancement potency in Caco-2 monolayers based on molecular descriptors and random forest

Structural traits of permeation enhancers are important determinants of their capacity to promote enhanced drug absorption. Therefore, in order to obtain a better understanding of structure-activity relationships for permeation enhancers, a Quantitative Structural Activity Relationship (QSAR) model has been developed. The random forest-QSAR model was based upon Caco-2 data for 41 surfactant-like permeation enhancers from Whitehead et al. (2008) and molecular descriptors calculated from their structure. The QSAR model was validated by two test-sets: (i) an eleven compound experimental set with Caco-2 data and (ii) nine compounds with Caco-2 data from literature. Feature contributions, a recent developed diagnostic tool, was applied to elucidate the contribution of individual molecular descriptors to the predicted potency. Feature contributions provided easy interpretable suggestions of important structural properties for potent permeation enhancers such as segregation of hydrophilic and lipophilic domains. Focusing on surfactant-like properties, it is possible to model the potency of the complex pharmaceutical excipients, permeation enhancers. For the first time, a QSAR model has been developed for permeation enhancement. The model is a valuable in silico approach for both screening of new permeation enhancers and physicochemical optimisation of surfactant enhancer systems.