In previous work [1], we presented several implementations of acoustic anomaly detection by training a model on purely normal data and estimating the divergence between it and other input. Here, we reformulate the problem in an unsupervised framework and allow for anomalous contamination of the training data. We focus exclusively on methods employing Gaussian mixture models (GMMs) since they are often used in speech processing systems. After analyzing what caused the Kullback-Leibler (KL) divergence between GMMs to break down in the face of training contamination, we came up with a promising solution. By trimming one quarter of the most divergent Gaussians from the mixture model, we significantly outperformed the untrimmed approximation for contamination levels of 10% and above, reducing the equal error rate from 33.8 % to 6.4 % at 33 % contamination. The performance of the trimmed KL divergence showed no significant dependence on the investigated contamination levels. Index Terms: anomaly detection, speech activity detection, unsupervised learning, Kullback-Leibler divergence.

When it becomes necessary to reduce the complexity of a classifier, dimensionality reduction can be an effective way to address classifier complexity. Linear discriminant analysis (LDA) is one approach to dimensionality reduction that makes use of a linear transformation matrix. The widely used Fisher's LDA is "sub-optimal" when the sample class covariance matrices are unequal, meaning that another linear transformation exists that produces lower loss in discrimination power. We introduce a geometric approach to linear discriminant analysis (GLDA) that can reduce the number of dimensions from n to m for any number of classes. GLDA is able to compute a better linear transformation matrix than Fisher's LDA for unequal sample class covariance matrices and is equivalent to Fisher's LDA when those matrices are equal or proportional. The classification problems we present demonstrate and strongly suggest that geometric LDA can generate the, "optimal" classifier in a lower dimension.

To extract the potential promised by superscalar processors, algorithm designers must streamline memory references and allow for efficient data reuse throughout the memory hierarchy. Two parameterized Householder QR factorization algorithms are presented that take into account the caches and registers typical of such processors. Guidelines are developed for choosing parameter values that obtain near-optimal cache and register utilization. The new algorithms are implemented and performance-tuned on an Intel Pentium Pro system, a single thin POWER2 node of the IBM Scalable Parallel system 2 (SP2), and a single R8000 processor of a Silicon Graphs POWER Challenge XL.

Rapid computation of the QR factorization of a matrix is fundamental to many scientific and engineering problems. The paper presents a family of algorithms parameterized by the number of processors available P, arithmetic grain aggregation parameters g1, g2, …, gP, and communication grain aggregation parameter h, which computer the QR factorization of a matrix A ∈ Cm × n with minimal latency. The approach is particularly well suited for dedicated distributed memory architectures such as linear arrays of INMOS Transputers, Texas Instruments C40s or Analog Devices 21060s.

One aspect of the PMC system level fault model, the properties of the implied faulty sets, is considered. For tau -diagnosable systems that have at most tau faulty units, lower bounds on the cardinality of the maximal implied faulty sets are given. Then it is shown that these bounds are greatest lower bounds, and it is indicated how these results can be used in diagnosis algorithms.>

In this paper we present a parallel procedure for the solution of first-order linear recurrence systems of size N when the number of processors p is small in relation to N. We show that when 1 < p2 ⩽ N, a first-order linear recurrence system of size N can be solved in 5(N − 1)(p + 1) steps on a p processor SIMD machine and at most 5(N − 12)/(p + 32) steps on a p processor MIMD machine. As a special case, we further show that our approach precisely achieves the lower bound 2(N − 1)(p + 1) for solving the parallel prefix problem on a p processor machine.

A τdiagnosable system is a system in which all faults may be identified from the test results, provided that the number of faults does not exceed τ. In this correspondence we present an algorithm that may be used for the diagnosis of the system level BGM fault model proposed by Barsi, Grandoni, and Maestrini, whenever the system is τ-diagnosable and the number of faults is at most τ.

Pertains to the analysis of an algorithm for the automatic fault diagnosis of asymmetric modular networks.

A new diagnosis algorithm for determining the existing fault situation in a symmetric multiple processor architecture is given. The algorithm assumes that there are n processors, each of which is tested by at least t other processors, and at most t of which are faulty. The existing fault situation is always diagnosed if n ≥ 2t + 1 and, in some cases, can still be diagnosed if n < 2t + 1. The implementation of the algorithm is straightforward and suitable for microprocessor applications.