Predicting socioeconomic indicators using transfer learning on imagery data: an application in Brazil.

Censuses and other surveys responsible for gathering socioeconomic data are expensive and time consuming. For this reason, in poor and developing countries there often is a long gap between these surveys, which hinders the appropriate formulation of public policies as well as the development of researches. One possible approach to overcome this challenge for some socioeconomic indicators is to use satellite imagery to estimate these variables, although it is not possible to replace demographic census surveys completely due to its territorial coverage, level of disaggregation of information and large set of information. Even though using orbital images properly requires, at least, a basic remote sensing knowledge level, these images have the advantage of being commonly free and easy to access. In this paper, we use daytime and nighttime satellite imagery and apply a transfer learning technique to estimate average income, GDP per capita and a constructed water index at the city level in two Brazilian states, Bahia and Rio Grande do Sul. The transfer learning approach could explain up to 64% of the variation in city-level variables depending on the state and variable. Although data from different countries may be considerably different, results are consistent with the literature and encouraging as it is a first analysis of its kind for Brazil.

A Fully Natural Gradient Scheme for Improving Inference of the Heterogeneous Multioutput Gaussian Process Model.

A recent novel extension of multioutput Gaussian processes (GPs) handles heterogeneous outputs, assuming that each output has its own likelihood function. It uses a vector-valued GP prior to jointly model all likelihoods' parameters as latent functions drawn from a GP with a linear model of coregionalization (LMC) covariance. By means of an inducing points' framework, the model is able to obtain tractable variational bounds amenable to stochastic variational inference (SVI). Nonetheless, the strong conditioning between the variational parameters and the hyperparameters burdens the adaptive gradient optimization methods used in the original approach. To overcome this issue, we borrow ideas from variational optimization introducing an exploratory distribution over the hyperparameters, allowing inference together with the posterior's variational parameters through a fully natural gradient (NG) optimization scheme. Furthermore, in this work, we introduce an extension of the heterogeneous multioutput model, where its latent functions are drawn from convolution processes. We show that our optimization scheme can achieve better local optima solutions with higher test performance rates than adaptive gradient methods for both the LMC and the convolution process model. We also show how to make the convolutional model scalable by means of SVI and how to optimize it through a fully NG scheme. We compare the performance of the different methods over the toy and real databases.

Bayesian inversion of a diffusion model with application to biology.

A common task in experimental sciences is to fit mathematical models to real-world measurements to improve understanding of natural phenomenon (reverse-engineering or inverse modelling). When complex dynamical systems are considered, such as partial differential equations, this task may become challenging or ill-posed. In this work, a linear parabolic equation is considered as a model for protein transcription from MRNA. The objective is to estimate jointly the differential operator coefficients, namely the rates of diffusion and self-regulation, as well as a functional source. The recent Bayesian methodology for infinite dimensional inverse problems is applied, providing a unique posterior distribution on the parameter space continuous in the data. This posterior is then summarized using a Maximum a Posteriori estimator. Finally, the theoretical solution is illustrated using a state-of-the-art MCMC algorithm adapted to this non-Gaussian setting.

Physically-Inspired Gaussian Process Models for Post-Transcriptional Regulation in Drosophila.

The regulatory process of Drosophila is thoroughly studied for understanding a great variety of biological principles. While pattern-forming gene networks are analyzed in the transcription step, post-transcriptional events (e.g., translation, protein processing) play an important role in establishing protein expression patterns and levels. Since the post-transcriptional regulation of Drosophila depends on spatiotemporal interactions between mRNAs and gap proteins, proper physically-inspired stochastic models are required to study the link between both quantities. Previous research attempts have shown that using Gaussian processes (GPs) and differential equations lead to promising predictions when analyzing regulatory networks. Here, we aim at further investigating two types of physically-inspired GP models based on a reaction-diffusion equation where the main difference lies in where the prior is placed. While one of them has been studied previously using protein data only, the other is novel and yields a simple approach requiring only the differentiation of kernel functions. In contrast to other stochastic frameworks, discretizing the spatial space is not required here. Both GP models are tested under different conditions depending on the availability of gap gene mRNA expression data. Finally, their performances are assessed on a high-resolution dataset describing the blastoderm stage of the early embryo of Drosophila melanogaster.

Switched Latent Force Models for Reverse-Engineering Transcriptional Regulation in Gene Expression Data.

To survive environmental conditions, cells transcribe their response activities into encoded mRNA sequences in order to produce certain amounts of protein concentrations. The external conditions are mapped into the cell through the activation of special proteins called transcription factors (TFs). Due to the difficult task to measure experimentally TF behaviors, and the challenges to capture their quick-time dynamics, different types of models based on differential equations have been proposed. However, those approaches usually incur in costly procedures, and they present problems to describe sudden changes in TF regulators. In this paper, we present a switched dynamical latent force model for reverse-engineering transcriptional regulation in gene expression data which allows the exact inference over latent TF activities driving some observed gene expressions through a linear differential equation. To deal with discontinuities in the dynamics, we introduce an approach that switches between different TF activities and different dynamical systems. This creates a versatile representation of transcription networks that can capture discrete changes and non-linearities. We evaluate our model on both simulated data and real data (e.g., microaerobic shift in E. coli, yeast respiration), concluding that our framework allows for the fitting of the expression data while being able to infer continuous-time TF profiles.

Gaussian process dynamical models for multimodal affect recognition.

Affective computing systems has a great potential in applications for biofeedback systems and cognitive conductual therapies. Here, by analyzing the physiological behavior of a given subject, we can infer the affective state of an emotional process. Since, emotions can be modeled as dynamic manifestations of these signals, a continuous analysis in the valence/arousal space, brings more information of the affective state related to an emotional process. In this paper we propose a method for dynamic affect recognition from multimodal physiological signals. Our model is based on learning a latent space using Gaussian process latent variable models (GP-LVM), which maps high dimensional data (multimodal physiological signals) in a low dimensional latent space. We incorporate the dynamics to the model by learning the latent representation, with associated dynamics. Finally, a support vector classifier is implemented to evaluate the relevance of the latent space features in the affective recognition process. The results show that the proposed method can efficiently model a physiological time-series and recognize with high accuracy an affective process.

Multi-output Gaussian processes for enhancing resolution of diffusion tensor fields.

Second order diffusion tensor (DT) fields are widely used in several clinical applications: brain fibers connections, diagnosis of neuro-degenerative diseases, image registration, brain conductivity models, etc. However, due to current acquisition protocols and hardware limitations in MRI machines, the diffusion magnetic resonance imaging (dMRI) data is obtained with low spatial resolution (1 or 2 mm3 for each voxel). This issue can be significant, because tissue fibers are much smaller than voxel size. Interpolation has become in a successful methodology for enhancing spatial resolution of DT fields. In this work, we present a feature-based interpolation approach through multi-output Gaussian processes (MOGP). First, we extract the logarithm of eigenvalues (direction) and the Euler angles (orientation) from diffusion tensors and we consider each feature as a separated but related output. Then, we interpolate the features along the whole DT field. In this case, the independent variables are the space coordinates (x, y, z). For this purpose, we assume that all features follow a multi-output Gaussian process with a common covariance matrix. Finally, we reconstruct new tensors from the interpolated eigenvalues and Euler angles. Accuracy of our methodology is better compared to approaches in the state of the art for performing DT interpolation, and it achieves a performance similar to the recently introduced method based on Generalized Wishart processes for interpolation of positive semidefinite matrices. We also show that MOGP preserves important properties of diffusion tensors such as fractional anisotropy.

A probabilistic framework based on SLIC-superpixel and Gaussian processes for segmenting nerves in ultrasound images.

We deal with an important problem in the field of anesthesiology known as automatic segmentation of nerve structures depicted in ultrasound images. This is important to aid the experts in anesthesiology, in order to carry out Peripheral Nerve Blocking (PNB). Ultrasound imaging has gained recent interest for performing PNB procedures since it offers a non-invasive visualization of the nerve and the anatomical structures around it. However, the location of these nerves in ultrasound images is a difficult task for the specialist due to the artifacts (i.e. speckle noise) that affect the intelligibility of a given image. In this paper, we present a probabilistic approach based on Simple Linear Iterative Clustering (SLIC-superpixels) and Gaussian processes for automatic segmentation of peripheral nerves. First, we use Graph cuts segmentation to define a region of interest (ROI). Such a ROI is divided into several correlated regions using SLIC-superpixels, then, a nonlinear Wavelet transform is applied as feature extraction stage. Finally, we use a classification scheme based on Gaussian Processes in order to predict the label of each parametrized superpixel (the label can be "nerve" or "background"). The accuracy of the proposed method is measured in terms of the Dice coefficient. Results obtained show performances with a Dice coefficient of 0.6524±0.0085 which brings accurate performances in nerve segmentation processes.

Kernel temporal enhancement approach for LORETA source reconstruction using EEG data.

Reconstruction of brain sources from magnetoencephalography and electroencephalography (M/EEG) data is a well known problem in the neuroengineering field. A inverse problem should be solved and several methods have been proposed. Low Resolution Electromagnetic Tomography (LORETA) and the different variations proposed as standardized LORETA (sLORETA) and the standardized weighted LORETA (swLORETA) have solved the inverse problem following a non-parametric approach, that is by setting dipoles in the whole brain domain in order to estimate the dipole positions from the M/EEG data and assuming some spatial priors. Errors in the reconstruction of sources are presented due the low spatial resolution of the LORETA framework and the influence of noise in the observable data. In this work a kernel temporal enhancement (kTE) is proposed in order to build a preprocessing stage of the data that allows in combination with the swLORETA method a improvement in the source reconstruction. The results are quantified in terms of three dipole error localization metrics and the strategy of swLORETA + kTE obtained the best results across different signal to noise ratio (SNR) in random dipoles simulation from synthetic EEG data.

Automatic segmentation of nerve structures in ultrasound images using Graph Cuts and Gaussian processes.

Peripheral Nerve Blocking (PNB), is a procedure used for performing regional anesthesia, that comprises the administration of anesthetic in the proximity of a nerve. Several techniques have been used with the purpose of locating nerve structures when the PNB procedure is performed: anatomical surface landmarks, elicitation of paresthesia, nerve stimulation and ultrasound imaging. Among those, ultrasound imaging has gained great attention because it is not invasive and offers an accurate location of the nerve and the structures around it. However, the segmentation of nerve structures in ultrasound images is a difficult task for the specialist, since such images are affected by echo perturbations and speckle noise. The development of systems for the automatic segmentation of nerve structures can aid the specialist for locating nerve structures accurately. In this paper we present a methodology for the automatic segmentation of nerve structures in ultrasound images. An initial step is carried out using Graph Cut segmentation in order to generate regions of interest; we then use machine learning techniques with the aim of segmenting the nerve structure; here, a specific non-linear Wavelet transform is used for the feature extraction stage, and Gaussian processes for the classification step. The methodology performance is measured in terms of accuracy and the dice coefficient. Results show that the implemented methodology can be used for automatically segmenting nerve structures.

Peripheral nerve segmentation using Nonparametric Bayesian Hierarchical Clustering.

Several cases related to chronic pain, due to accidents, illness or surgical interventions, depend on anesthesiology procedures. These procedures are assisted with ultrasound images. Although, the ultrasound images are a useful instrument in order to guide the specialist in anesthesiology, the lack of intelligibility due to speckle noise, makes the clinical intervention a difficult task. In a similar manner, some artifacts are introduced in the image capturing process, challenging the expertise of anesthesiologists for not confusing the true nerve structures. Accordingly, an assistance methodology using image processing can improve the accuracy in the anesthesia practice. This paper proposes a peripheral nerve segmentation method in medical ultrasound images, based on Nonparametric Bayesian Hierarchical Clustering. The experimental results show segmentation performances with a Mean Squared Error performance of 1.026 ± 0.379 pixels for ulnar nerve, 0.704 ± 0.233 pixels for median nerve and 1.698 ± 0.564 pixels for peroneal nerve. Likewise, the model allows to emphasize other soft structures like muscles and aqueous tissues, that might be useful for an anesthesiologist.

Automatic assessment of voice quality in the context of multiple annotations.

Approaches to evaluate voice quality include perceptual analysis, and acoustical analysis. Perceptual analysis is subjective and depends mostly on the ability of a specialist to assess a pathology, whereas acoustical analysis is objective, but highly relies on the quality of the so called annotations that the specialist assigns to the voice signal. The quality of the annotations for acoustical analysis depends heavily on the expertise and knowledge of the specialist. We face a scenario where we have annotations performed by several specialists with different levels of expertise and knowledge. Traditional pattern recognition methods employed in acoustical analysis are no longer applicable, since these methods are designed for scenarios where a "ground-truth" label is assigned by the specialist. In this paper, we apply recent developments in machine learning for taking into account multiple annotators for acoustical analysis of voice signals. For the classification step we compare two techniques, one of them based on Gaussian Processes for regression with multiple annotators, and the other is a multi-class Logistic Regression model that measures the annotator performance in terms of sensitivity and specificity. The performance of classifiers is assessed in terms of Cohen's Kappa index. Results show that the multi-annotator classification schemes have better performance when compared to techniques based on a traditional classifier where the true label is estimated from the multiple annotations available using majority voting.

Latent force models for describing transcriptional regulation processes in the embryo development problem for the Drosophila melanogaster.

In the embryo development problem for the Drosophila melanogaster, a set of molecules known as mor-phogens are responsible for the embryo segmentation. These morphogens are encoded by different genes, including the GAP genes, maternal coordination genes and pair-rule genes. One of the maternal coordination genes encodes the Bicoid morphogen, which is the responsible for the development of the Drosophila embryo at the anterior part and for the control and regulation of the GAP genes in segmentation of the early development of the Drosophila melanogaster. The work presented in this document, reports a methodology that tends to integrate mechanistic and data driven based models, aiming at making inference over the mRNA Bicoid from gene expression data at the protein level for the Bicoid morphogen. The fundamental contribution of this work is the description of the concentration gradient of the Bicoid morphogen in the continuous spatio-temporal domain as well as the output regression (gene expression at protein level) using a Gaussian process described by a mechanistically inspired covariance function. Regression results and metrics computed for the Bicoid protein expression both in the temporal and spatial domains, showed outstanding performance with respect to reported experiments from previous studies. In this paper, a correlation coefficient of r = 0.9758 against a correlation coefficient of r = 0.9086 is being reported, as well as a SMSE of 0.0303±0.1512 against a SMSE of 0.1106±0.5090 and finally reporting a MSLL of -1.7036 ± 1.3472 against -1.0151±1.7669.

Sparse representation of MER signals for localizing the Subthalamic Nucleus in Parkinson's disease surgery.

Deep brain stimulation (DBS) of Subthalamic Nucleus (STN) is the best method for treating advanced Parkinson's disease (PD), leading to striking improvements in motor function and quality of life of PD patients. During DBS, online analysis of microelectrode recording (MER) signals is a powerful tool to locate the STN. Therapeutic outcomes depend of a precise positioning of a stimulator device in the target area. In this paper, we show how a sparse representation of MER signals allows to extract discriminant features, improving the accuracy in identification of STN. We apply three techniques for over-complete representation of signals: Method of Frames (MOF), Best Orthogonal Basis (BOB) and Basis Pursuit (BP). All the techniques are compared to classical methods for signal processing like Wavelet Transform (WT), and a more sophisticated method known as adaptive Wavelet with lifting schemes (AW-LS). We apply each processing method in two real databases and we evaluate its performance with simple supervised classifiers. Classification outcomes for MOF, BOB and BP clearly outperform WT and AW-LF in all classifiers for both databases, reaching accuracy values over 98%.

Multiple-output support vector machine regression with feature selection for arousal/valence space emotion assessment.

Human emotion recognition (HER) allows the assessment of an affective state of a subject. Until recently, such emotional states were described in terms of discrete emotions, like happiness or contempt. In order to cover a high range of emotions, researchers in the field have introduced different dimensional spaces for emotion description that allow the characterization of affective states in terms of several variables or dimensions that measure distinct aspects of the emotion. One of the most common of such dimensional spaces is the bidimensional Arousal/Valence space. To the best of our knowledge, all HER systems so far have modelled independently, the dimensions in these dimensional spaces. In this paper, we study the effect of modelling the output dimensions simultaneously and show experimentally the advantages in modeling them in this way. We consider a multimodal approach by including features from the Electroencephalogram and a few physiological signals. For modelling the multiple outputs, we employ a multiple output regressor based on support vector machines. We also include an stage of feature selection that is developed within an embedded approach known as Recursive Feature Elimination (RFE), proposed initially for SVM. The results show that several features can be eliminated using the multiple output support vector regressor with RFE without affecting the performance of the regressor. From the analysis of the features selected in smaller subsets via RFE, it can be observed that the signals that are more informative into the arousal and valence space discrimination are the EEG, Electrooculogram/Electromiogram (EOG/EMG) and the Galvanic Skin Response (GSR).

A latent force model for describing electric propagation in deep brain stimulation: a simulation study.

Deep brain stimulation (DBS) is a neurosurgical method used to treat symptoms of movement disorders by implanting electrodes in deep brain areas. Often, the DBS modeling approaches found in the literature assume a quasi-static approximation, and discard any dynamic behavior. Nevertheless, in a real DBS system the stimulus corresponds to a wave that changes as a function of time. It is clear that DBS demands an approach that takes into account the time-varying behavior of the input stimulus. In this work, we present a novel latent force model for describing the dynamic electric propagation occurred during DBS. The performance of the proposed model was studied by simulations under different conditions. The results show that our approach is able to take into account the time variations of the source and the produced field. Moreover, by restricting our model it is possible to obtain solutions for electrostatic formulations, here experimental results were compared with the finite element method. Additionally, our approach allows a solution to the inverse problem, which is a valuable clinical application allowing the appropriate tuning of the DBS device by the expert physician.

Dynamic physiological signal analysis based on Fisher kernels for emotion recognition.

Emotional behavior is an active area of study in the fields of neuroscience and affective computing. This field has the fundamental role of emotion recognition in the maintenance of physical and mental health. Valence/Arousal levels are two orthogonal, independent dimensions of any emotional stimulus and allows an analysis framework in affective research. In this paper we present our framework for emotional regression based on machine learning techniques. Autoregressive coefficients and hidden markov models on physiological signals, based on Fisher Kernels characterization are presented for mapping variable length sequences to new dimension feature vector space. Then, support vector regression is performed over the Fisher Scores for emotional recognition. Also quantitatively we evaluated the accuracy of the proposed model by acomplishing a hold-out cross validation over the dataset. The experimental results show that the proposed model can effectively perform the regression in comparison with static characterization methods.

Unsupervised learning applied in MER and ECG signals through Gaussians mixtures with the Expectation-Maximization algorithm and Variational Bayesian Inference.

Automatic identification of biosignals is one of the more studied fields in biomedical engineering. In this paper, we present an approach for the unsupervised recognition of biomedical signals: Microelectrode Recordings (MER) and Electrocardiography signals (ECG). The unsupervised learning is based in classic and bayesian estimation theory. We employ gaussian mixtures models with two estimation methods. The first is derived from the frequentist estimation theory, known as Expectation-Maximization (EM) algorithm. The second is obtained from bayesian probabilistic estimation and it is called variational inference. In this framework, both methods are used for parameters estimation of Gaussian mixtures. The mixtures models are used for unsupervised pattern classification, through the responsibility matrix. The algorithms are applied in two real databases acquired in Parkinson's disease surgeries and electrocardiograms. The results show an accuracy over 85% in MER and 90% in ECG for identification of two classes. These results are statistically equal or even better than parametric (Naive Bayes) and nonparametric classifiers (K-nearest neighbor).

Feature selection for multimodal emotion recognition in the arousal-valence space.

Emotion recognition is a challenging research problem with a significant scientific interest. Most of the emotion assessment studies have focused on the analysis of facial expressions. Recently, it has been shown that the simultaneous use of several biosignals taken from the patient may improve the classification accuracy. An open problem in this area is to identify which biosignals are more relevant for emotion recognition. In this paper, we perform Recursive Feature Elimination (RFE) to select a subset of features that allows emotion classification. Experiments are carried out over a multimodal database with arousal and valence annotations, and a diverse range of features extracted from physiological, neurophysiological, and video signals. Results show that several features can be eliminated while still preserving classification accuracy in setups of 2 and 3 classes. Using a small subset of the features, it is possible to reach 70% accuracy for arousal and 60% accuracy for valence in some experiments. Experimentally, it is shown that the Galvanic Skin Response (GSR) is relevant for arousal classification, while the electroencephalogram (EEG) is relevant for valence.

Linear latent force models using Gaussian processes.

Purely data-driven approaches for machine learning present difficulties when data are scarce relative to the complexity of the model or when the model is forced to extrapolate. On the other hand, purely mechanistic approaches need to identify and specify all the interactions in the problem at hand (which may not be feasible) and still leave the issue of how to parameterize the system. In this paper, we present a hybrid approach using Gaussian processes and differential equations to combine data-driven modeling with a physical model of the system. We show how different, physically inspired, kernel functions can be developed through sensible, simple, mechanistic assumptions about the underlying system. The versatility of our approach is illustrated with three case studies from motion capture, computational biology, and geostatistics.