Nilanjan Dasgupta

Genetic algorithm wavelet design for signal classification

Biorthogonal wavelets are applied to parse multiaspect transient scattering data in the context of signal classification. A language-based genetic algorithm is used to design wavelet filters that enhance classification performance. The biorthogonal wavelets are implemented via the lifting procedure and the optimization is carried out using a classification-based cost function. Example results are presented for target classification using measured scattering data.

Dual hidden Markov model for characterizing wavelet coefficients from multi-aspect scattering data

Angle-dependent scattering (electromagnetic or acoustic) is considered from a general target, for which the scattered signal is a non-stationary function of the target-sensor orientation. A statistical model is presented for the wavelet coefficients of such a signal, in which the angular non-stationarity is characterized by an “outer” hidden Markov model (HMMo). The statistics of the wavelet coefficients, within a state of the outer HMM, are characterized by a second, “inner” HMMi, exploiting the tree structure of the wavelet decomposition. This dual-HMM construct is demonstrated by considering multi-aspect target identification using measured acoustic scattering data.

Class-based target identification with multiaspect scattering data

In underwater sensing applications, it is often difficult to train a classifier in advance for all targets that may be seen during testing, due to the large number of targets that may be encountered. We therefore partition the training data into target classes, with each class characteristic of multiple targets that share similar scattering physics. In some cases, one may have a priori insight into which targets should constitute a given class, while in other cases this segmentation must be done autonomously based on the scattering data. For the latter case, we constitute the classes using an information-theoretic mapping criterion. Having defined the target classes, the second phase of our identification procedure involves determining those features that enhance the similarity between the targets in a given class. This is achieved by using a genetic algorithm (GA)-based feature-selection algorithm with a Kullback-Leibler (KL) cost function. The classifier employed is appropriate for multiaspect scattering data and is based on a hidden Markov model (HMM). The performance of the class-based classification algorithm is examined using both measured and computed acoustic scattering data from submerged elastic targets.

A Bayesian approach to unsupervised feature selection and density estimation using expectation propagation

We propose an approximate Bayesian approach for unsupervised feature selection and density estimation, where the importance of the features for clustering is used as the measure for feature selection. Traditional maximum-likelihood (ML) model-parameter optimization schemes estimate the feature saliencies for a fixed model structure (i.e., a fixed number of clusters). In practice, the number of clusters present in the data for mixture-based modeling is unknown. In an ML framework, the number of clusters typically needs to be ascertained prior to estimating the feature saliencies. We propose a density estimation scheme that addresses model complexity (number of clusters present) and model-parameter estimation (feature saliencies) in a single optimization framework. The approximate Bayesian approach presented here, based on the expectation propagation method, obtains a full posterior distribution on the saliency of the features, along with full posterior distribution of other model parameters (including the number of clusters) that represent the underlying statistics of the data. The performance of the algorithm, is analyzed based on its ability to identify the features salient for clustering the multivariate data.

Time-reversal imaging for classification of submerged elastic targets via Gibbs sampling and the Relevance Vector Machine

Time-reversal imaging (TRI) is analogous to matched-field processing, although TRI is typically very wideband and is appropriate for subsequent target classification (in addition to localization). Time-reversal techniques, as applied to acoustic target classification, are highly sensitive to channel mismatch. Hence, it is crucial to estimate the channel parameters before time-reversal imaging is performed. The channel-parameter statistics are estimated here by applying a geoacoustic inversion technique based on Gibbs sampling. The maximum a posteriori (MAP) estimate of the channel parameters are then used to perform time-reversal imaging. Time-reversal implementation requires a fast forward model, implemented here by a normal-mode framework. In addition to imaging, extraction of features from the time-reversed images is explored, with these applied to subsequent target classification. The classification of time-reversed signatures is performed by the relevance vector machine (RVM). The efficacy of the technique is analyzed on simulated in-channel data generated by a free-field finite element method (FEM) code, in conjunction with a channel propagation model, wherein the final classification performance is demonstrated to be relatively insensitive to the associated channel parameters. The underlying theory of Gibbs sampling and TRI are presented along with the feature extraction and target classification via the RVM.

Texture analysis with variational hidden Markov trees

A variational Bayes formulation of the hidden Markov tree (HMT) model is proposed for texture analysis, utilizing a multilevel wavelet decomposition of imagery. The variational method yields an approximation to the full posterior of the HMT parameters. Texture classification is based on the posterior predictive distribution or marginalized evidence, with example results presented.

ICA and PLS modeling for functional analysis and drug sensitivity for DNA microarray signals

The DNA microarray technique offers an ability to analyze the expression profile of a genome. The complex correlation between the large number of genes present in the genome undermines straightforward understanding of their functionality. In this paper, we have proposed a pair of modeling schemes to recognize the functional identities of the known genes. In Independent Component Analysis (ICA), each of the microarray signals is modeled as a linear combination of some underlying independent components having specific biological interpretation. The second algorithm, Partial Least Squares (PLS) is proposed to identify the latent functional units contributing to drug sensitivity from the microarray data. Applications of this research include prediction of drug responses based on gene expressions, and also to identify the function(s) of a new gene. We consider the Rosetta compendium data set with yeast gene profiles, and the NCI-60 data set of human gene expressions as a function of drug type (cancer drugs are considered).

On the extended-Born technique for scattering from buried dielectric targets

The extended-Born technique is an approximate nonlinear method for analyzing scattering from a weak discontinuity. Moreover, when applied to the low-frequency (electromagnetic induction) applications for which it was developed originally, the extended-Born method has accurately modeled scattering from inhomogeneities considerably larger than those appropriate for the standard linear Born technique. We examine the extended-Born technique at radar frequencies considering three-dimensional (3-D) scattering from a dielectric target buried in a lossy half space.

Time-reversal imaging for wideband underwater target classification

Time-reversal imaging is addressed for sensing an elastic target situated in an acoustic waveguide. It is demonstrated that the channel parameters associated with a given measurement may be determined via a genetic-algorithm (GA) search in parameter space. Target classification based on time-reversal imagery is considered, with this implemented via a relevance-vector machine.

Modeling Pharmacogenomics of the NCI-60 Anticancer Data Set: Utilizing Kernel Pls to Correlate the Microarray Data to Therapeutic Responses

Modeling the relationship between genomic features and therapeutic response is of central interest in pharmacogenomics [Musumarra et al., 2001]. The NCI-60 cancer data set with both gene expression and drug activity measurements provides an excellent opportunity for this modeling exercise. To correlate the gene expression profile with the drug activity pattern, we utilized a soft modeling technique called Partial Least Squares (PLS) [Tobias, 2000]. Soft modeling requires less stringent assumptions about the data than other modeling techniques [Falk et al., 1992], A high level of collinearity in multidimensional gene expression profiles motivates us to undertake the PLS approach, which not only trims data redundancy but also exposes the underlying hidden functional units as latent features. It is believed that these functional gene groups play a key role in determining the efficacy of the cancer drugs to different cell lines (types of cancer). We have shown the efficacy of PLS in identifying drug resistant and drug sensitive genes. We have also investigated techniques to exploit the non-linear dependence between individual gene expressions in order to explain variations in the drug activity pattern. This is facilitated by a kernel function that implicitly carries out the regression in a higher-dimensional space where the data is linear [Christiannini et al., 2000]. The kernel-based non-linear approach is shown to be more effective in defining the correlation between the drug response and the gene expressions. The PLS approach, as implemented here, could be used to differentiate cancer cell lines between renal cancer and melanoma, for example, or different drug groups like Alkylating agents and Tubulin-active anti-mitotic agents.