Catálogo de publicaciones - libros

For the extraction of sources with unsupervised learning techniques invariance under certain transformations, such as shifts, rotations or scaling, is often a desirable property. A straight-forward approach for accomplishing this goal is to include these transformations and its parameters into the mixing model. For the case of one-dimensional signals in presence of shifts this problem has been termed anechoic demixing, and several algorithms for the analysis of time series have been proposed. Here, we generalize this approach for sources depending on multi-dimensional arguments and apply it for learning of translation-invariant features from higher-dimensional data, such as images. A new algorithm for the solution of such high-dimensional anechoic demixing problems based on the Wigner-Ville distribution is presented. It solves the multi-dimensional problem by projection onto multiple one-dimensional problems. The feasibility of this algorithm is demonstrated by learning independent features from sets of real images.

This paper proposes several algorithms to recover the transmitted signals in systems with multiple antennas that make use of orthogonal space time block code (O-STBC) to attain full transmit diversity. We interpret the scheme proposed by Alamouti in [1] and half-rate systems presented in [2] as classic blind source separation (BSS) problems where the received signals (observations) are instantaneous mixtures of the transmitted signals (sources). In order to recover the sources, we first propose to perform an eigenvalue decomposition of matrices containing fourth-order cross-cumulants of the observations. Subsequently, we show that the performance of this approach can be improved by doing a simultaneous diagonalization of the cumulant matrices. This second approach can be interpreted as a particular case of Joint Approximate Diagonalization of Eigen-matrices (JADE) algorithm for systems where the mixing matrix is orthogonal.

This paper proposes to explore the potential of Blind Source Separation (BSS) techniques for the estimation of modal parameters, namely the resonant frequencies, vibration modes and damping ratios. The concept of , which was introduced in recent publications, allows to consider BSS as a simple way of doing output-only modal analysis. This work illustrates the proposed methodology using free and random responses of an experimental truss structure.

This paper presents a novel algorithm to build hierarchies from independent component analyzer mixtures and its application to image similarity measure. The hierarchy algorithm composes an agglomerative (bottom-up) clustering from the estimated parameters (basis vectors and bias terms) of the ICA mixture. Merging at different levels of the hierarchy is made using the Kullback-Leibler distance between clusters. The procedure is applied to merge similar patches on a natural image, to group different images of an object, and to create hierarchical levels of clustering from images of different objects. Results show suitable image hierarchies obtained by clustering from basis functions to higher-level structures.

Deriving a thematically meaningful partition of an unlabeled text corpus is a challenging task. In comparison to classic term-based document indexing, the use of document representations based on latent thematic generative models can lead to improved clustering. However, determining the optimal indexing technique is not straightforward, as it depends on the clustering problem faced and the partitioning strategy adopted. So as to overcome this indeterminacy, we propose deriving a consensus labeling upon the results of clustering processes executed on several document representations. Experiments conducted on subsets of two standard text corpora evaluate distinct clustering strategies based on latent thematic spaces and highlight the usefulness of consensus clustering to overcome the optimal document indexing indeterminacy.

Top-down and bottom-up processing are two distinct yet highly interactive modes of neuronal activity underlying normal and abnormal human cognition. Here we characterize the dynamic processes that contribute to these two modes of cognitive operation. We used a blind source separation algorithm called second-order blind identification (SOBI [1]) to extract from high-density scalp EEG (128 channels) two components that index neuronal activity in two distinct local networks: one in the occipital lobe and one in the frontal lobe. We then applied Granger causality analysis to the SOBI-recovered neuronal signals from these two local networks to characterize feed-forward and feedback influences between them. With three repeated observations made at least one week apart, we show that feed-forward influence is dominated by alpha while feedback influence is dominated by theta band activity and that this direction-selective dominance pattern is jointly modulated by situational familiarity and demand for visual processing.

Multiple factors related to scene structure, illumination, and imaging contribute to image formation. Independent Components Analysis (ICA) maximizes the statistical independence of the representational components of a training image ensemble, but it cannot distinguish between these different factors, or modes. To address this problem, we introduce a nonlinear, multifactor model that generalizes ICA. Our model of image ensembles learns the statistically independent components of each of the multiple factors. We present an associated dimensionality reduction algorithm for multifactor subspace analysis. As an application, we consider the multilinear analysis of ensembles of facial images that combine several modes, including different facial geometries (people), expressions, head poses, and lighting conditions. For the purposes of face recognition, we introduce a that simultaneously projects an unknown test image into the multiple constituent mode spaces in order to infer its mode labels. We show that multilinear ICA computes a set of factor subspaces that yield improved recognition rates.

We consider Independent Component Analysis (ICA) for the case of binary sources, where addition has the meaning of the boolean “Exclusive Or” (XOR) operation. Thus, each mixture-signal is given by the XOR of one or more of the source-signals. While such mixtures can be considered linear transformations over the finite Galois Field of order 2, they are certainly nonlinear over the field of real-valued numbers, so classical ICA principles may be inapplicable in this framework. Nevertheless, we show that if none of the independent random sources is uniform (i.e., neither one has probability 0.5 for 1/0), then any invertible mixing is identifiable (up to permutation ambiguity). We then propose a practical deflation algorithm for source separation based on entropy minimization, and present empirical performance results by simulation.

Since Independent Component Analysis was developed, it has been a hotspot in the field of signal processing, and has received increasing attention in feature extraction, data compression, and so on. In this paper, a novel ICA-based image/video processing method, called ICA transform (ICAT), is proposed. Instead of the traditional blocking, ICAT derives more than one sub-images/sub-videos from one original image/video by down-sampling, and features are obtained from these sub-images/sub-videos by using ICA. That helps ICAT extracts features with the global characteristics of the original. And the comparison between ICAT and Digital Wavelet Transform (DWT) is performed in image/video processing, which exhibits that the results obtained by using ICAT has something similar to those of DWT, even something superior. And the comparison also demonstrates that ICAT is promising in image/video processing.

This keynote talk describes a state-of-the-art method for the blind source separation (BSS) of convolutive mixtures of audio signals. Independent component analysis (ICA) is used as a major statistical tool for separating the mixtures. We provide examples to show how ICA criteria change as the number of audio sources increases. We then discuss a frequency-domain approach where simple instantaneous ICA is employed in each frequency bin. A directivity pattern analysis of the ICA solutions provides us with a physical interpretation of the ICA-based separation. It tells us the relationship between ICA-based BSS and adaptive beamforming. In order to obtain properly separated signals with the frequency-domain approach, the permutation and scaling ambiguity of the ICA solutions should be aligned appropriately. We describe two complementary methods for aligning the permutations, i.e., collecting separated frequency components originating from the same source. The first method exploits the signal envelope dependence of the same source across frequencies. The second method relies on the spatial diversity of the sources, and is closely related to source localization techniques. Finally, we describe methods for sparse source separation, which can be applied even to an underdetermined case.