Support vector machines (SVMs) have been successfully applied to classification problems. The difficulty in selecting the most effective error penalty has been partly resolved with /spl nu/-SVM. However, the use of uneven training class sizes, which occurs frequently with target detection problems, results in machines with biases towards the class with the larger training set. We propose an extended /spl nu/-SVM to counter the effects of the unbalanced training class sizes. The resulting dual /spl nu/-SVM provides the facility to counter these effects, as well as to adjust the error penalties of each class separately. The parameter /spl nu/ of each class provides a lower bound to the fraction of support vector of that class, and the upper bound to the fraction of bounded support vector of that class. These bounds allow the control on the error rates allowed for each class, and enable the training of machines with specific error rate requirements.

We describe an approach to an image interpretation problem which is based on a combination of modelling and machine learning. It involves inferring properties of the image formation process from the structures present in the image. Due to the possible complexity of the structures and the ill-posed nature of the problem, modelling does not always adequately explain the data. We use support vector machines to supervise the model fitting process by training them to identify those situations where modelling will fail.

We describe a parsimonious model for the direction-dependent transfer function of the pinna. The model describes the transfer function with reference to resonators located in particular physical positions relative to the ear canal. The purpose of the work is to provide a parametric model that permits identification with moderate data-gathering, and filter specification for any direction without the need for interpolation of responses.

An hierarchical neural network structure for clustering problems is presented and a statistical analysis of its performance is conducted. This neural network architecture aims to find, through competition and cooperation, maximally related objects in a scene. The architecture was first introduced by Maren and Ali (1983), and was named the hierarchical scene structure (HSS). We propose an enhancement of the original HSS and demonstrate that this leads to an improved performance. It is also shown that further improvement in performance can be achieved by cascading two enhanced HSS networks.>

With the increased availability of coherent wideband radars, there has been a renewed interest in radar target recognition. A large bandwidth gives high resolution in range which means target recognition may be possible. We examine some of the problems of classifying high resolution range profiles (HRRP), and investigate simple preprocessing techniques which may aid subsequent target classification. We apply these techniques to HRRP data acquired at a local airport using the Microwave Radar Division (MRD) mobile radar facility. We find that we can reliably distinguish between Boeing 727 and Boeing 737 aircraft over a range of aspect angles.>

Fully autonomous mobile robots often rely on an array of sensors to provide them with an adequate picture of their environment. Furthermore, these systems tend to have strict limitations in terms of available processing capability. Hence the so-called "smart sensing" approach is particularly appropriate as it combines a small size with a reduced requirement for interpretation of sensory input. This paper describes how a visual micro-sensor implemented in VLSI can be used for obstacle avoidance as well as localised path planning or navigation.>

The pattern search predictor (PSP) predicts samples of a signal by inspecting the past for patterns of (about ten) samples that match the most recent set. The sample subsequent to the found pattern is used to make the required estimate. PSP has been tested in a codec algorithm based on the CCITT 32-kb adaptive differential pulse-code modulation standard, using its adaptive quantizer. Study of spectrograms has shown that the error is substantially white, as expected, and that perturbations of the signal spectrograms are substantially undetectable. PSP has some promise for filling in lost data.>

In this paper a new sample association approach to be known as the Hilbert Warping (HW) algorithm and associated modules are described. This algorithm is chosen from the observation that signals of similar form but with different time scales appear as similar trajectories when represented by suitable two dimensional plots in the X-Y plane, and overcomes difficulties such as identification of signal endpoints and assumptions about the smooth nature of warping that are permissible, associated with dynamic programming algorithms. The HW algorithm can be applied to both single dimensional and multi-dimensional signals as in dynamic programming algorithms.

In some pattern recognition tasks multiple observations of an observation vector Y = {Y1, Y2, ..., YM} are available for each object and the covariances of the Yi are characteristic of the object. With the assistance of a model of the generating process for Y a theoretical basis for the comparison of the covariance matrices is developed. Measurements based on synthetic data support the theory, and an application to some speaker verification is given as an example.