Search Results

This chapter examines the most relevant aspects concerning the use of independent component analysis (ICA) for the analysis of functional magnetic resonance imaging (fMRI) data. In particular, after illustrating the fMRI-ICA model (“Problem formulation and application to fMRI”), the chapter compares the most commonly used ICA algorithms in the context of fMRI data analysis. The problems of choosing the dimensionality of the ICA decomposition, and of selecting the “meaningful” components, are considered. Optimizations of the ICA algorithms for dealing with the specific spatiotemporal properties of the fMRI data, and extensions of the ICA to multisubject fMRI studies, are described. For each of these aspects, different approaches from various groups are briefly reviewed.

Independent component analysis (ICA) is increasingly utilized as a tool for evaluating the hidden spatiotemporal structure contained within brain imaging data. This chapter first provides a brief overview of ICA and how ICA is applied to functional magnetic resonance imaging (fMRI) data. It then discusses group ICA and the application of group ICA for data fusion, with an emphasis on the methods developed within our group. It also discusses, within a larger context, the many alternative approaches that are feasible and currently in use.

This chapter introduces and applies the concept of parallel spatial and temporal unmixing with group independent component analysis (ICA) for concurrent electroencephalography-functional magnetic resonance imaging (EEG-fMRI). Hemodynamic response function (HRF) deconvolution and single-trial estimation in the fMRI data were employed, and the single-trial weights were used as predictors for the amplitude modulation in the EEG. For illustration, data from a previously published performance-monitoring experiment were analyzed, in order to identify error-preceding activity in the EEG modality. EEG components that displayed such slow trends, and which were coupled to the corresponding fMRI components, are described. Parallel ICA for analysis of concurrent EEG-fMRI on a trial-by-trial basis is a very useful addition to the toolbelt of researchers interested in multimodal integration.

Independent component analysis (ICA) is a linear decomposition technique that aims to reveal the underlying statistical sources of mixed signals. The EEG signal consists of a mixture of various brain and non-brain contributions. Accordingly, a valid and powerful unmixing tool promises a better, more accessible representation of the statistical sources contributing to the mixed recorded signal. ICA, being potentially such a tool, may help in the detection of signal sources that cannot be identified on the raw data level alone using other, more conventional techniques. The application of ICA to EEG signals has become popular, as it provides two key features: it is a powerful way to remove artifacts from EEG data, and it helps to disentangle otherwise mixed brain signals. This chapter is concerned with evaluating and optimizing EEG decompositions by means of ICA. First, it discusses typical ICA results with reference to artifact- and brain-related components. Then, it elaborates on different EEG pre-processing steps, considered in light of the statistical assumptions underlying ICA. As such, the motivation for the chapter is to provide some practical guidelines for those researchers who wish to successfully decompose multi-channel EEG recordings.

This chapter gives an overview of data integration methods for simultaneous EEG-fMRI, in which EEG features are extracted and used to parametrically model the fMRI data. Up to now, variants of EEG-informed fMRI analysis have been most widely and successfully applied. After a brief discussion of the rationale of this approach, its variants for ongoing and event-related EEG phenomena are explained. Studies applying EEG-informed fMRI are reviewed. The advantage of denoising methods such as independent component analysis allowing single-trial quantifications of the EEG phenomena of interest is discussed. To allow clear interpretations of covariations between electrophysiological and hemodynamic measures, further dependent variables such as behavioral data should be taken into account. The chapter closes with an outlook on future questions and ongoing methodological developments.

Languages are traditionally subdivided into geographically distinct dialects, although any such division is just a coarse approximation of a more fine-grained variation. This underlying variation is usually visualised in the form of maps, where the distribution of various features is shown as isoglosses. Component models such as factor analysis can be used to analyse spatial distributions of a large number of different features — such as the isogloss data in a dialect atlas or the distributions of ethnological or archaeological phenomena — with the goal of finding dialects or similar cultural aggregates. However, there are several such methods, and it is not obvious how their differences affect their usability for computational dialectology. This chapter addresses this question by comparing five such methods (factor analysis, non-negative matrix factorisation, aspect Bernoulli, independent component analysis, and principal components analysis) with two data sets describing Finnish dialectal variation. There are some fundamental differences between these methods, and some of these have implications that affect the dialectological interpretation of the results.

To understand cognition and emotion in the real world, it is critical to investigate the phenomena of interest within the rich context of moment-to-moment variations in the real world, which we assume is at least in part encoded in the high-dimensional state of the brain. Here the chapter reviews empirical evidence from a series of novel validation studies that demonstrate the technical capabilities of one blind source separation (BSS) algorithm—second-order blind identification (SOBI)—in enabling neuronscientists and clinicians to investigate human brain functions, cognition, and behavior using the electroencephalography (EEG). The chapter concludes that by shifting from an EEG-sensor-based to a neuronal-source-based characterization of brain states, one may better capture the rich context of moment-to-moment variations in the real world.

This chapter presents an approach for developing a fully automatic Facial Action Coding System (FACS). The approach uses state-of-the-art machine learning techniques that can be applied to recognition of any facial action. The results of Study I provided guidance as to which image representations, or feature extraction methods, are most effective for facial action recognition. Gabor wavelets and Independent Component Analysis gave best performance. Study II found that machine learning techniques applied directly to the warped images is a promising approach for automatic coding of spontaneous facial expressions. Generally, the data employed hand-labeled feature points for the head pose tracking step. Furthermore, three of the issues are discussed in detail: (1) collection of a database of spontaneous facial expressions, (2) fully automatic face detection and tracking, and (3) fully automatic 3D head pose estimation.