Showing papers by "Helsinki University of Technology published in 1997"

A fast fixed-point algorithm for independent component analysis

Abstract: We introduce a novel fast algorithm for independent component analysis, which can be used for blind source separation and feature extraction. We show how a neural network learning rule can be transformed into a fixedpoint iteration, which provides an algorithm that is very simple, does not depend on any user-defined parameters, and is fast to converge to the most accurate solution allowed by the data. The algorithm finds, one at a time, all nongaussian independent components, regardless of their probability distributions. The computations can be performed in either batch mode or a semiadaptive manner. The convergence of the algorithm is rigorously proved, and the convergence speed is shown to be cubic. Some comparisons to gradient-based algorithms are made, showing that the new algorithm is usually 10 to 100 times faster, sometimes giving the solution in just a few iterations.

Language-specific phoneme representations revealed by electric and magnetic brain responses

Abstract: There is considerable debate about whether the early processing of sounds depends on whether they form part of speech. Proponents of such speech specificity postulate the existence of language-dependent memory traces, which are activated in the processing of speech1–3 but not when equally complex, acoustic non-speech stimuli are processed. Here we report the existence of these traces in the human brain. We presented to Finnish subjects the Finnish phoneme prototype /e/ as the frequent stimulus, and other Finnish phoneme prototypes or a non-prototype (the Estonian prototype /o/) as the infrequent stimulus. We found that the brain's automatic change-detection response, reflected electrically as the mismatch negativity (MMN)4–10, was enhanced when the infrequent, deviant stimulus was a prototype (the Finnish /o/) relative to when it was a non-prototype (the Estonian /o/). These phonemic traces, revealed by MMN, are language-specific, as /o/ caused enhancement of MMN in Estonians. Whole-head magnetic recordings11,12 located the source of this native-language, phoneme-related response enhancement, and thus the language-specific memory traces, in the auditory cortex of the left hemisphere.

Signal-space projection method for separating MEG or EEG into components.

Abstract: CURRENTS INSIDE a conducting body can be estimated by measuring the magnetic and/or the electric field at multiple locations outside and then constructing a solution to the inverse problem, i.e. determining a current configuration that could have produced the measured field. Unfortunately, there is no unique solution to this problem (HELMHOLTZ, 1853) unless restricting assumptions are made. The minimum-norm estimate (HAM/~.L,~INEN and ILMONIEMI, 1994) provides a solution with the smallest expected overall error when minimum a priori information about the source distribution is available. Other methods to estimate a continuous current distribution producing the measured signals have been studied (PASCUAL-MARQUI et al., 1994; WANG et aL, 1995; GORODNITSKY, et al., 1995). A different approach is to divide the brain activity into discrete components such as current dipoles (ScHERG, 1990; MOSHER et al., 1992). Here we widen this approach into arbitrary current configurations. In our signal-space projection (SSP) method, the signals measured by d sensors are considered to form a time-varying vector in a d-dimensional signal space. The component vectors,, i.e. the signals caused by the different neuronal sources, have different and fixed orientations in the signal space. In other words, each source has a distinct and stable field pattern. All the current eonfi~marations producing the same measured field pattern are indistinguishable on the basis of the field: they have the same vector direction in the signal space and thus belong to the same equivalence class of current configurations (TESCHE et al., 1995a). The angle in the signal space between vectors representing different equivalence classes, e.g. between component vectors, is a measure of similarity of the equivalence classes in signal space and a way to characterise the separability of sources. The cosine of this angle has previously been used as a numerical charaeterisation of the difference between topographical distributions (DESMEDT and CHALK[.IN, 1989). If the direction of at least one of the component vectors forming the measured multi-channel signal can be determined from the data, or is known otherwise, SSP can be used to simplify subsequent analysis. For example, if an early deflection in an evoked response is produced by one source, and the rest of the response is a mixture of signals from this and other sources, SSP can separate the data into two parts so that the early source contributes only to one part. In general, the signals are divided into two orthogonal parts: s~, including the time-varying contribution from sources with known signalspace directions; and s~_, including the rest of the signals. Both sl~ and s j_ can then be analysed separately in more detail. By analysing s t , we can detect activity originally masked by s~. On the other hand, the sources included in stl are seen with an enhanced signal-to-noise ratio. By forward modelling of sources in selected patches of cortex, it is possible to form a spatial filter that selectively passes only the signals that may have been generated by currents in the given patches. If the subspace defined by artefacts can be determined, the artefactflee S L can be analysed. In SSP, in contrast to PCA (HARRIS, 1975; MAIER et al., 1987) and other analysis methods (GRUMMICH et al., 1991; KOLES et aL, 1990; KOLES, 1991; SOONG and KOLES, 1995; BESA*), the source decomposition does not depend on the orthogonality of source components or the availability of source or conductivity models. No conductivity or source models are needed if the component vectors are estimated directly from the measured signals. This is useful when no source estimation is needed, e.g. when artefacts or somatomotor activity in a cogrritive study must be filtered out. The angles between the components provide an easy and illustrative way to characterise the linear dependence between the components and thus the separability of sources. The concept of signal space in MEG was introduced previously ([LMONIEMI, 1981; [LMONIEMI and WILLIAMSON,

Cortical Control of Human Motoneuron Firing During Isometric Contraction

Abstract: Salenius, Stephan, Karin Portin, Matti Kajola, Riitta Salmelin, and Riitta Hari. Cortical control of human motoneuron firing during isometric contraction. J. Neurophysiol. 77: 3401–3405, 1997. We recorded whole scalp magnetoencephalographic (MEG) signals simultaneously with the surface electromyogram from upper and lower limb muscles of six healthy right-handed adults during voluntary isometric contraction. The 15- to 33-Hz MEG signals, originating from the anterior bank of the central sulcus, i.e., the primary motor cortex, were coherent with motor unit firing in all subjects and for all muscles. The coherent cortical rhythms originated in the hand motor area for upper limb muscles (1st dorsal interosseus, extensor indicis proprius, and biceps brachii) and close to the foot area for lower limb muscles (flexor hallucis brevis). The sites of origin corresponding to different upper limb muscles did not differ significantly. The cortical signals preceded motor unit firing by 12–53 ms. The lags were shortest ...

Smodels — an implementation of the stable model and well-founded semantics for normal logic programs

Abstract: The Smodels system is a C++ implementation of the well-founded and stable model semantics for range-restricted function-free normal programs. The system includes two modules: (i) smodels which implements the two semantics for ground programs and (ii) parse which computes a grounded version of a range-restricted function-free normal program. The latter module does not produce the whole set of ground instances of the program but a subset that is sufficient in the sense that no stable models are lost. The implementation of the stable model semantics for ground programs is based on bottom-up backtracking search where a powerful pruning method is employed. The pruning method exploits an approximation technique for stable models which is closely related to the well-founded semantics. One of the advantages of this novel technique is that it can be implemented to work in linear space. This makes it possible to apply the stable model semantics also in areas where resulting programs are highly non-stratified and can possess a large number of stable models. The implementation has been tested extensively and compared with a state of the art implementation of the stable model semantics, the SLG system. In tests involving ground programs it clearly outperforms SLG.

A class of neural networks for independent component analysis

Abstract: Independent component analysis (ICA) is a recently developed, useful extension of standard principal component analysis (PCA). The ICA model is utilized mainly in blind separation of unknown source signals from their linear mixtures. In this application only the source signals which correspond to the coefficients of the ICA expansion are of interest. In this paper, we propose neural structures related to multilayer feedforward networks for performing complete ICA. The basic ICA network consists of whitening, separation, and basis vector estimation layers. It can be used for both blind source separation and estimation of the basis vectors of ICA. We consider learning algorithms for each layer, and modify our previous nonlinear PCA type algorithms so that their separation capabilities are greatly improved. The proposed class of networks yields good results in test examples with both artificial and real-world data.

New Approximations of Differential Entropy for Independent Component Analysis and Projection Pursuit

Abstract: We derive a first-order approximation of the density of maximum entropy for a continuous 1-D random variable, given a number of simple constraints. This results in a density expansion which is somewhat similar to the classical polynomial density expansions by Gram-Charlier and Edgeworth. Using this approximation of density, an approximation of 1-D differential entropy is derived. The approximation of entropy is both more exact and more robust against outliers than the classical approximation based on the polynomial density expansions, without being computationally more expensive. The approximation has applications, for example, in independent component analysis and projection pursuit.

Point-defect complexes and broadband luminescence in GaN and AlN

Abstract: We have employed the plane-wave pseudopotential method to study point defect complexes in GaN and AlN. The results reveal that defect complexes consisting of dominant donors bound to cation vacancies are likely to be formed in both materials. The position of the electronic levels in the band gap due to these defect complexes is shown to correlate well with the experimentally commonly observed broadband luminescence both in GaN and in AlN. The origin of the large bandwidth of the luminescence spectrum is discussed.

On the measurement of preferences in the analytic hierarchy process

Abstract: In this paper we apply multiattribute value theory as a framework for examining the use of pairwise comparisons in the analytic hierarchy process (AHP). On one hand our analysis indicates that pairwise comparisons should be understood in terms of preference differences between pairs of alternatives. On the other hand it points out undesirable effects caused by the upper bound and the discretization of any given ratio scale. Both these observations apply equally well to the SMART procedure which also uses estimates of weight ratios. Furthermore, we demonstrate that the AHP can be modified so as to produce results similar to those of multiattribute value measurement; we also propose new balanced scales to improve the sensitivity of the AHP ratio scales. Finally we show that the so-called supermatrix technique does not eliminate the rank reversal phenomenon which can be attributed to the normalizations in the AHP. © 1997 John Wiley & Sons, Ltd.

Heart sound segmentation algorithm based on heart sound envelogram

Abstract: Desribes the development of a segmentation algorithm which separates the heart sound signal into four parts: the first heart sound, the systole, the second heart sound and the diastole. The segmentation of phonocardiogram (PCG) signals is the first step of analysis and the most important procedure in the automatic diagnosis of heart sounds. This algorithm is based on the normalized average Shannon energy of a PCG signal. The performance of the algorithm has been evaluated using 515 periods of PCG signals recording from 37 objects including normal and abnormal. The algorithm has achieved a 93 percent correct ratio.

A novel channel estimation method for OFDM mobile communication systems based on pilot signals and transform-domain processing

Abstract: A method of estimating the channel transfer function is presented for orthogonal frequency division multiplexing (OFDM) mobile communication systems working under time-variant radio channel conditions. The proposed method employs lowpass filtering in a transform domain so that intercarrier interference and additive white Gaussian noise components in the received pilot signals are significantly reduced. The cutoff frequency of the transform-domain filter is dynamically selected by tracking the received pilot signals. The channel transfer function for all the subcarriers is obtained by a high-resolution interpolation realized by zero-padding and DFT/IDFT. The proposed method is applicable for all linear modulation OFDM systems. It is demonstrated with a 16QAM-OFDM system which includes both amplitude and phase modulations.

Independent Component Analysis for Identification of Artifacts in Magnetoencephalographic Recordings

Abstract: We have studied the application of an independent component analysis (ICA) approach to the identification and possible removal of artifacts from a magnetoencephalographic (MEG) recording. This statistical technique separates components according to the kurtosis of their amplitude distributions over time, thus distinguishing between strictly periodical signals, and regularly and irregularly occurring signals. Many artifacts belong to the last category. In order to assess the effectiveness of the method, controlled artifacts were produced, which included saccadic eye movements and blinks, increased muscular tension due to biting and the presence of a digital watch inside the magnetically shielded room. The results demonstrate the capability of the method to identify and clearly isolate the produced artifacts.

Exploration of very large databases by self-organizing maps

Abstract: This paper describes a data organization system and genuine content-addressable memory called the WEBSOM. It is a two-layer self-organizing map (SOM) architecture where documents become mapped as points on the upper map, in a geometric order that describes the similarity of their contents. By standard browsing tools one can select from the map subsets of documents that are most similar mutually. It is also possible to submit free-form queries about the wanted documents whereby the WEBSOM locates the best-matching documents. The document map exemplified in this paper has over 100000 map nodes, with 315 inputs at each, and over 1000000 documents have been organized by it. The system has been implemented by software on a general-purpose computer.

A family of fixed-point algorithms for independent component analysis

Abstract: Independent component analysis (ICA) is a statistical signal processing technique whose main applications are blind source separation, blind deconvolution, and feature extraction. Estimation of ICA is usually performed by optimizing a 'contrast' function based on higher-order cumulants. It is shown how almost any error function can be used to construct a contrast function to perform the ICA estimation. In particular, this means that one can use contrast functions that are robust against outliers. As a practical method for finding the relevant extrema of such contrast functions, a fixed-point iteration scheme is then introduced. The resulting algorithms are quite simple and converge fast and reliably. These algorithms also enable estimation of the independent components one-by-one, using a simple deflation scheme.

Frequency and periodicity are represented in orthogonal maps in the human auditory cortex: evidence from magnetoencephalography

Abstract: Timbre and pitch are two independent perceptual qualities of sounds closely related to the spectral envelope and to the fundamental frequency of periodic temporal envelope fluctuations, respectively. To a first approximation, the spectral and temporal tuning properties of neurons in the auditory midbrain of various animals are independent, with layouts of these tuning properties in approximately orthogonal tonotopic and periodotopic maps. For the first time we demonstrate by means of magnetoencephalography a periodotopic organization of the human auditory cortex and analyse its spatial relationship to the tonotopic organization by using a range of stimuli with different temporal envelope fluctuations and spectra and a magnetometer providing high spatial resolution. We demonstrate an orthogonal arrangement of tonotopic and periodotopic gradients. Our results are in line with the organization of such maps in animals and closely match the perceptual orthogonality of timbre and pitch in humans.

Calculation of solidification-related thermophysical properties for steels

Abstract: Special algorithms have been developed to calculate important solidification-related thermophysical properties: enthalpy and enthalpy-related data (i.e., specific and latent heat), density, and thermal conductivity for low-alloyed and stainless steels. The algorithms are heavily based on the use of earlier developed phase transformation models, an interdendritic solidification model (IDS), and an austenite decomposition model (ADC), which solve, as a function of temperature, the phase fractions and compositions needed in these calculations. As a result, the thermophysical properties can be calculated at any temperature, from 1600 °C to 25 °C, taking into account the discontinuities caused by special phase transformations (i.e., ferritic, austenitic and peritectic solidification, ferrite/austenite transformation, and austenite decomposition to various structures) influenced by the steel grade and the cooling conditions.

Modulation of the Parieto-Occipital Alpha Rhythm during Object Detection

Abstract: Changes in the human neuromagnetic alpha rhythm were monitored during an object detection task to study the effects of visual shape processing on the parieto-occipital activity. Pictures of coherent meaningful objects, which the observers had to detect, and of disorganized meaningless non-objects were presented briefly between masks. The non-objects were systematically followed by a higher level of alpha than the objects, the difference emerging on average 400 msec after the stimulus, with a median delay of 130 msec after evoked response onsets in the occipital, temporal, and parietal cortices. Without attention to visual shape, the alpha levels did not differ between objects and non-objects. The alpha level was higher after non-objects than missed objects, and higher after missed than correctly detected objects, suggesting that the alpha level is inversely related to saliency or familiarity of the object and does not directly reflect visual awareness. The reactive alpha rhythm was generated in the parieto-occipital sulcus, which in several primate species includes areas belonging to the dorsal visual pathway. According to current views, the parietal cortex produces attentional signals that filter out irrelevant information in the ventral visual stream. Our results reinforce the idea of bidirectional interaction: information derived from visual shape can rapidly modify activity in the parieto-occipital region. The synchronized alpha oscillations may reflect attenuation of occipito-parietal information transfer and disengagement of parietal cortex from object selection.