Adaptive Sparse Gaussian Process.

Adaptive learning is necessary for nonstationary environments where the learning machine needs to forget past data distribution. Efficient algorithms require a compact model update to not grow in computational burden with the incoming data and with the lowest possible computational cost for online parameter updating. Existing solutions only partially cover these needs. Here, we propose the first adaptive sparse Gaussian process (GP) able to address all these issues. We first reformulate a variational sparse GP (VSGP) algorithm to make it adaptive through a forgetting factor. Next, to make the model inference as simple as possible, we propose updating a single inducing point of the SGP model together with the remaining model parameters every time a new sample arrives. As a result, the algorithm presents a fast convergence of the inference process, which allows an efficient model update (with a single inference iteration) even in highly nonstationary environments. Experimental results demonstrate the capabilities of the proposed algorithm and its good performance in modeling the predictive posterior in mean and confidence interval estimation compared to state-of-the-art approaches.

A semi-empirical approach to calibrate simulation models for semiconductor devices.

Semiconductor device optimization using computer-based prototyping techniques like simulation or machine learning digital twins can be time and resource efficient compared to the conventional strategy of iterating over device design variations by fabricating the actual device. Ideally, simulation models require perfect calibration of material parameters for the model to represent a particular semiconductor device. This calibration process itself can require characterization information of the device and its precursors and extensive expert knowledge of non characterizable parameters and their tuning. We propose a hybrid method to calibrate multiple simulation models for a device using minimal characterization data and machine learning-based prediction models. A photovoltaic device is chosen as the example for this technique where optical and electrical simulation models of an industrially manufactured silicon solar cell are calibrated and the simulated device performance is compared with the measurement data from the physical device.

Automated malarial retinopathy detection using transfer learning and multi-camera retinal images.

Cerebral malaria (CM) is a fatal syndrome found commonly in children less than 5 years old in Sub-saharan Africa and Asia. The retinal signs associated with CM are known as malarial retinopathy (MR), and they include highly specific retinal lesions such as whitening and hemorrhages. Detecting these lesions allows the detection of CM with high specificity. Up to 23% of CM, patients are over-diagnosed due to the presence of clinical symptoms also related to pneumonia, meningitis, or others. Therefore, patients go untreated for these pathologies, resulting in death or neurological disability. It is essential to have a low-cost and high-specificity diagnostic technique for CM detection, for which We developed a method based on transfer learning (TL). Models pre-trained with TL select the good quality retinal images, which are fed into another TL model to detect CM. This approach shows a 96% specificity with low-cost retinal cameras.

Girasol, a sky imaging and global solar irradiance dataset.

The energy available in a microgrid that is powered by solar energy is tightly related to the weather conditions at the moment of generation. A very short-term forecast of solar irradiance provides the microgrid with the capability of automatically controlling the dispatch of energy. We propose a dataset to forecast Global Solar Irradiance (GSI) using a data acquisition system (DAQ) that simultaneously records sky imaging and GSI measurements, with the objective of extracting features from clouds and use them to forecast the power produced by a Photovoltaic (PV) system. The DAQ system is nicknamed the Girasol Machine (Girasol means Sunflower in Spanish). The sky imaging system consists of a longwave infrared (IR) camera and a visible (VI) light camera with a fisheye lens attached to it. The cameras are installed inside a weatherproof enclosure that it is mounted on a solar tracker. The tracker updates its pan and tilt every second using a solar position algorithm to maintain the Sun in the center of the IR and VI images. A pyranometer is situated on a horizontal mount next to the DAQ system to measure GSI. The dataset, composed of IR images, VI images, GSI measurements, and the Sun's positions, has been tagged with timestamps.

Discovering brain regions relevant to obsessive-compulsive disorder identification through bagging and transduction.

In the present study we applied a multivariate feature selection method based on the analysis of the sign consistency of voxel weights across bagged linear Support Vector Machines (SVMs) with the aim of detecting brain regions relevant for the discrimination of subjects with obsessive-compulsive disorder (OCD, n=86) from healthy controls (n=86). Each participant underwent a structural magnetic resonance imaging (sMRI) examination that was pre-processed in Statistical Parametric Mapping (SPM8) using the standard pipeline of voxel-based morphometry (VBM) studies. Subsequently, we applied our multivariate feature selection algorithm, which also included an L2 norm regularization to account for the clustering nature of MRI data, and a transduction-based refinement to further control overfitting. Our approach proved to be superior to two state-of-the-art feature selection methods (i.e., mass-univariate t-Test selection and recursive feature elimination), since, following the application of transductive refinement, we obtained a lower test error rate of the final classifier. Importantly, the regions identified by our method have been previously reported to be altered in OCD patients in studies using traditional brain morphometry methods. By contrast, the discrimination patterns obtained with the t-Test and the recursive feature elimination approaches extended across fewer brain regions and included fewer voxels per cluster. These findings suggest that the feature selection method presented here provides a more comprehensive characterization of the disorder, thus yielding not only a superior identification of OCD patients on the basis of their brain anatomy, but also a discrimination map that incorporates most of the alterations previously described to be associated with the disorder.

A multiple kernel learning approach to perform classification of groups from complex-valued fMRI data analysis: application to schizophrenia.

FMRI data are acquired as complex-valued spatiotemporal images. Despite the fact that several studies have identified the presence of novel information in the phase images, they are usually discarded due to their noisy nature. Several approaches have been devised to incorporate magnitude and phase data, but none of them has performed between-group inference or classification. Multiple kernel learning (MKL) is a powerful field of machine learning that finds an automatic combination of kernel functions that can be applied to multiple data sources. By analyzing this combination of kernels, the most informative data sources can be found, hence providing a better understanding of the analyzed learning task. This paper presents a methodology based on a new MKL algorithm (ν-MKL) capable of achieving a tunable sparse selection of features' sets (brain regions' patterns) that improves the classification accuracy rate of healthy controls and schizophrenia patients by 5% when phase data is included. In addition, the proposed method achieves accuracy rates that are equivalent to those obtained by the state of the art lp-norm MKL algorithm on the schizophrenia dataset and we argue that it better identifies the brain regions that show discriminative activation between groups. This claim is supported by the more accurate detection achieved by ν-MKL of the degree of information present on regions of spatial maps extracted from a simulated fMRI dataset. In summary, we present an MKL-based methodology that improves schizophrenia characterization by using both magnitude and phase fMRI data and is also capable of detecting the brain regions that convey most of the discriminative information between patients and controls.

Identification of patterns of gray matter abnormalities in schizophrenia using source-based morphometry and bagging.

Despite its reliable diagnosis, schizophrenia lacks an objective diagnostic test or a validated biomarker, which prevents a better understanding of this disorder. Structural magnetic resonance imaging (sMRI) has been vastly explored to find consistent abnormality patterns of gray matter concentration (GMC) in schizophrenia, yet we are far from having reached conclusive evidence. This paper presents a machine learning approach based on resampling techniques to find brain regions with consistent patterns of GMC differences between healthy controls and schizophrenia patients, these regions being detected by means of source-based morphometry. This work uses multi-site data from the Mind Clinical Imaging Consortium, which is composed of sMRI data from 124 controls and 110 patients. Our method achieves a better classification rate than other algorithms and detects regions with GMC differences between both groups that are consistent with several findings on the literature. In addition, the results obtained on data from multiple sites suggest that it may be possible to replicate these results on other datasets.

Spatially aggregated multiclass pattern classification in functional MRI using optimally selected functional brain areas.

In previous works, boosting aggregation of classifier outputs from discrete brain areas has been demonstrated to reduce dimensionality and improve the robustness and accuracy of functional magnetic resonance imaging (fMRI) classification. However, dimensionality reduction and classification of mixed activation patterns of multiple classes remain challenging. In the present study, the goals were (a) to reduce dimensionality by combining feature reduction at the voxel level and backward elimination of optimally aggregated classifiers at the region level, (b) to compare region selection for spatially aggregated classification using boosting and partial least squares regression methods and (c) to resolve mixed activation patterns using probabilistic prediction of individual tasks. Brain activation maps from interleaved visual, motor, auditory and cognitive tasks were segmented into 144 functional regions. Feature selection reduced the number of feature voxels by more than 50%, leaving 95 regions. The two aggregation approaches further reduced the number of regions to 30, resulting in more than 75% reduction of classification time and misclassification rates of less than 3%. Boosting and partial least squares (PLS) were compared to select the most discriminative and the most task correlated regions, respectively. Successful task prediction in mixed activation patterns was feasible within the first block of task activation in real-time fMRI experiments. This methodology is suitable for sparsifying activation patterns in real-time fMRI and for neurofeedback from distributed networks of brain activation.

Analysis of fMRI time series with mutual information.

Neuroimaging plays a fundamental role in the study of human cognitive neuroscience. Functional magnetic resonance imaging (fMRI), based on the Blood Oxygenation Level Dependent signal, is currently considered as a standard technique for a system level understanding of the human brain. The problem of identifying regionally specific effects in neuroimaging data is usually solved by applying Statistical Parametric Mapping (SPM). Here, a mutual information (MI) criterion is used to identify regionally specific effects produced by a task. In particular, two MI estimators are presented for its use in fMRI data. The first one uses a Parzen probability density estimation, and the second one is based on a K Nearest Neighbours (KNN) estimation. Additionally, a statistical measure has been introduced to automatically detect the voxels which are relevant to the fMRI task. Experiments demonstrate the advantages of MI estimators over SPM maps; firstly, providing more significant differences between relevant and irrelevant voxels; secondly, presenting more focalized activation; and, thirdly, detecting small areas related to the task. These findings, and the improved performance of KNN MI estimator in multisubject and multistimuli studies, make the proposed methods a good alternative to SPM.

Characterization of groups using composite kernels and multi-source fMRI analysis data: application to schizophrenia.

Pattern classification of brain imaging data can enable the automatic detection of differences in cognitive processes of specific groups of interest. Furthermore, it can also give neuroanatomical information related to the regions of the brain that are most relevant to detect these differences by means of feature selection procedures, which are also well-suited to deal with the high dimensionality of brain imaging data. This work proposes the application of recursive feature elimination using a machine learning algorithm based on composite kernels to the classification of healthy controls and patients with schizophrenia. This framework, which evaluates nonlinear relationships between voxels, analyzes whole-brain fMRI data from an auditory task experiment that is segmented into anatomical regions and recursively eliminates the uninformative ones based on their relevance estimates, thus yielding the set of most discriminative brain areas for group classification. The collected data was processed using two analysis methods: the general linear model (GLM) and independent component analysis (ICA). GLM spatial maps as well as ICA temporal lobe and default mode component maps were then input to the classifier. A mean classification accuracy of up to 95% estimated with a leave-two-out cross-validation procedure was achieved by doing multi-source data classification. In addition, it is shown that the classification accuracy rate obtained by using multi-source data surpasses that reached by using single-source data, hence showing that this algorithm takes advantage of the complimentary nature of GLM and ICA.

Support vector machines with constraints for sparsity in the primal parameters.

This paper introduces a new support vector machine (SVM) formulation to obtain sparse solutions in the primal SVM parameters, providing a new method for feature selection based on SVMs. This new approach includes additional constraints to the classical ones that drop the weights associated to those features that are likely to be irrelevant. A ν-SVM formulation has been used, where ν indicates the fraction of features to be considered. This paper presents two versions of the proposed sparse classifier, a 2-norm SVM and a 1-norm SVM, the latter having a reduced computational burden with respect to the first one. Additionally, an explanation is provided about how the presented approach can be readily extended to multiclass classification or to problems where groups of features, rather than isolated features, need to be selected. The algorithms have been tested in a variety of synthetic and real data sets and they have been compared against other state of the art SVM-based linear feature selection methods, such as 1-norm SVM and doubly regularized SVM. The results show the good feature selection ability of the approaches.

Atlas-based automated positioning of outer volume suppression slices in short-echo time 3D MR spectroscopic imaging of the human brain.

Spatial suppression of peripheral lipid-containing regions in volumetric MR spectroscopic imaging of the human brain requires placing large numbers of outer volume suppression (OVS) slices, which is time-consuming, prone to operator error and may introduce subject-dependent variability in volume coverage. We developed a novel, computationally efficient atlas-based approach for automated positioning of up to 16 OVS slices and the MR spectroscopic imaging slab. Standardized positions in Montreal Neurological Institute atlas space were established offline using a recently developed iterative optimization procedure. During the scanning session, positions in subject space were computed using affine transformation of standardized positions in Montreal Neurological Institute space. Offline analysis using magnetization prepared rapid gradient echo scans from 11 subjects demonstrated reliable OVS placement, comparable with but faster than iterative placement in subject space. This atlas-based method was further validated in 14 subjects using 3D short-echo time proton-echo-planar-spectroscopic-imaging at 3 T. Comparison of manual and automatic placement using 8 OVS slices demonstrated consistent MR spectroscopic imaging volume selection and comparable spectral quality with similar degree of lipid suppression and number of usable voxels. Automated positioning of 16 OVS slices enabled larger volume coverage, while maintaining similar spectral quality and lipid suppression. Atlas-based automatic prescription of short echo time MR spectroscopic imaging is expected to be advantageous for longitudinal and cross-sectional studies.

Automatic placement of outer volume suppression slices in MR spectroscopic imaging of the human brain.

Spatial suppression of peripheral regions (outer volume suppression) is used in MR spectroscopic imaging to reduce contamination from strong lipid and water signals. The manual placement of outer volume suppression slices requires significant operator interaction, which is time consuming and introduces variability in volume coverage. Placing a large number of outer volume saturation bands for volumetric MR spectroscopic imaging studies is particularly challenging and time consuming and becomes unmanageable as the number of suppression bands increases. In this study, a method is presented that automatically segments a high-resolution MR image in order to identify the peripheral lipid-containing regions. This method computes an optimized placement of suppression bands in three dimensions and is based on the maximization of a criterion function. This criterion function maximizes coverage of peripheral lipid-containing areas and minimizes suppression of cortical brain regions and regions outside of the head. Computer simulation demonstrates automatic placement of 16 suppression slices to form a convex hull that covers peripheral lipid-containing regions above the base of the brain. In vivo metabolite mapping obtained with short echo time proton-echo-planar spectroscopic imaging shows that the automatic method yields a placement of suppression slices that is very similar to that of a skilled human operator in terms of lipid suppression and usable brain voxels.

Upport vector machines for nonlinear kernel ARMA system identification.

Nonlinear system identification based on support vector machines (SVM) has been usually addressed by means of the standard SVM regression (SVR), which can be seen as an implicit nonlinear autoregressive and moving average (ARMA) model in some reproducing kernel Hilbert space (RKHS). The proposal of this letter is twofold. First, the explicit consideration of an ARMA model in an RKHS (SVM-ARMA2K) is proposed. We show that stating the ARMA equations in an RKHS leads to solving the regularized normal equations in that RKHS, in terms of the autocorrelation and cross correlation of the (nonlinearly) transformed input and output discrete time processes. Second, a general class of SVM-based system identification nonlinear models is presented, based on the use of composite Mercer's kernels. This general class can improve model flexibility by emphasizing the input-output cross information (SVM-ARMA4K), which leads to straightforward and natural combinations of implicit and explicit ARMA models (SVR-ARMA2K and SVR-ARMA4K). Capabilities of these different SVM-based system identification schemes are illustrated with two benchmark problems.

fMRI pattern classification using neuroanatomically constrained boosting.

Pattern classification in functional MRI (fMRI) is a novel methodology to automatically identify differences in distributed neural substrates resulting from cognitive tasks. Reliable pattern classification is challenging due to the high dimensionality of fMRI data, the small number of available data sets, interindividual differences, and dependence on the acquisition methodology. Thus, most previous fMRI classification methods were applied in individual subjects. In this study, we developed a novel approach to improve multiclass classification across groups of subjects, field strengths, and fMRI methods. Spatially normalized activation maps were segmented into functional areas using a neuroanatomical atlas and each map was classified separately using local classifiers. A single multiclass output was applied using a weighted aggregation of the classifier's outputs. An Adaboost technique was applied, modified to find the optimal aggregation of a set of spatially distributed classifiers. This Adaboost combined the region-specific classifiers to achieve improved classification accuracy with respect to conventional techniques. Multiclass classification accuracy was assessed in an fMRI group study with interleaved motor, visual, auditory, and cognitive task design. Data were acquired across 18 subjects at different field strengths (1.5 T, 4 T), with different pulse sequence parameters (voxel size and readout bandwidth). Misclassification rates of the boosted classifier were between 3.5% and 10%, whereas for the single classifier, these were between 15% and 23%, suggesting that the boosted classifier provides a better generalization ability together with better robustness. The high computational speed of boosting classification makes it attractive for real-time fMRI to facilitate online interpretation of dynamically changing activation patterns.