Customizing kernel functions for SVM-based hyperspectral image classification.

Previous research applying kernel methods such as support vector machines (SVMs) to hyperspectral image classification has achieved performance competitive with the best available algorithms. However, few efforts have been made to extend SVMs to cover the specific requirements of hyperspectral image classification, for example, by building tailor-made kernels. Observation of real-life spectral imagery from the AVIRIS hyperspectral sensor shows that the useful information for classification is not equally distributed across bands, which provides potential to enhance the SVM's performance through exploring different kernel functions. Spectrally weighted kernels are, therefore, proposed, and a set of particular weights is chosen by either optimizing an estimate of generalization error or evaluating each band's utility level. To assess the effectiveness of the proposed method, experiments are carried out on the publicly available 92AV3C dataset collected from the 220-dimensional AVIRIS hyperspectral sensor. Results indicate that the method is generally effective in improving performance: spectral weighting based on learning weights by gradient descent is found to be slightly better than an alternative method based on estimating "relevance" between band information and ground truth.

Neural network models of categorical perception.

Studies of the categorical perception (CP) of sensory continua have a long and rich history in psychophysics. In 1977, Macmillan, Kaplan, and Creelman introduced the use of signal detection theory to CP studies. Anderson and colleagues simultaneously proposed the first neural model for CP, yet this line of research has been less well explored. In this paper, we assess the ability of neural-network models of CP to predict the psychophysical performance of real observers with speech sounds and artificial/novel stimuli. We show that a variety of neural mechanisms are capable of generating the characteristics of CP. Hence, CP may not be a special model of perception but an emergent property of any sufficiently powerful general learning system.

Parity still isn't a generalisation problem.

Clark & Thornton take issue with my claim that parity is not a generalisation problem, and that nothing can be inferred about back-propagation in particular, or learning in general, from failures of parity generalisation. They advance arguments to support their contention that generalisation is a relevant issue. In this continuing commentary, I examine generalisation more closely in order to refute these arguments. Different learning algorithms will have different patterns of failure: back-propagation has no special status in this respect. This is not to deny that a particular algorithm might fortuitously happen to produce the "intended" function in an (oxymoronic) parity-generalisation task.

Determining and improving the fault tolerance of multilayer perceptrons in a pattern-recognition application.

We investigate empirically the performance under damage conditions of single- and multilayer perceptrons (MLP's), with various numbers of hidden units, in a representative pattern-recognition task. While some degree of graceful degradation was observed, the single-layer perceptron was considerably less fault tolerant than any of the multilayer perceptrons, including one with fewer adjustable weights. Our initial hypothesis that fault tolerance would be significantly improved for multilayer nets with larger numbers of hidden units proved incorrect. Indeed, there appeared to be a liability to having excess hidden units. A simple technique (called augmentation) is described, which was successful in translating excess hidden units into improved fault tolerance. Finally, our results were supported by applying singular value decomposition (SVD) analysis to the MLP's internal representations.

A computational model of afferent neural activity from the cochlea to the dorsal acoustic stria.

The first comprehensive computational model of the precortical mammalian auditory system to include afferent neural processing up to the level of the dorsal acoustic stria (DAS) is described. The model consists of two scissile stages simulating (1) the cochlea and auditory nerve (AN) and (2) the dorsal cochlear nucleus (DCN). The model derives its input from a 128-channel cochlear filterbank. Cochlear transduction, rectification, logarithmic compression, and two-tone suppression functions are performed at the first stage of the simulation. The 512 artificial neurons employed model the cell at the level of transmembrane potential and have interconnections that follow closely those reported in recent anatomical and physiological studies of the cat AN and DCN. The responses of the model to pure-tone stimuli (at various sound-pressure levels) and noise stimuli (at various levels and bandwidths) are reported in detail and compare well with published results. The model is being used to investigate the representation of initial English stop consonants (differing in voice-onset time) in the DAS; this work is briefly described.

Hand-held text-to-speech device for the non-vocal disabled.

A hand-held, battery-powered synthetic speech aid for the non-vocally disabled has been constructed. The device accepts as its input, largely unrestricted text keyed by the user. This is converted by text-to-speech software, based on 349 letter-to-sound rules and some simple rules of continuity, intonation and stress, to appropriate control signals which drive a single-chip (series formant) speech synthesizer. A number of implementation constraints are imposed by portability; the system has, as far as possible, been designed using CMOS components. To extend the time for which the system will operate between battery charges, power saving facilities are incorporated. Hand-held use implies the need for a one-handed keyboard: a unique integral keyboard is used, designed to minimize the visual search time to locate a letter key. Considerable attention has been paid to rule-search strategies, the handling of 'exceptions' which violate the letter-to-sound principle and the resolution of conflicts when more than one rule might apply. The quality and intelligibility of speech from a rule-based system is typically poor, and every effort has been made to improve it. Limits on possible improvement are, however, set by the use of a proprietary single chip synthesizer and by the minimal nature of a portable system. To facilitate the task of composing messages, a two-line liquid crystal display is provided together with a range of editing functions. The display can also be shown to the message receiver should he/she be deaf, or used for silent communication as an analogue to 'whispering'.(ABSTRACT TRUNCATED AT 250 WORDS)

Television and video technology in the education of deaf children.

This paper describes a survey into the use of television and related visual media in the education of hearing-impaired children in the UK. The main purpose of the survey was to furnish background information for the development of a low-cost video subtitling system for use in special schools and in units for hearing-impaired children in ordinary schools. An effective method of reducing cost is to base the system on equipment that schools and units already have. The survey indicates that domestic-type video recorders are becoming almost universal. Suitable microcomputers for running subtitling software are also widespread. Other results of the survey reveal: (i) widespread and fairly heavy usage of schools television programmes; (ii) the need for various visual aids and for considerable effort on the part of teachers in making the programmes accessible to hearing-impaired children; (iii) a significant demand for subtitling equipment; (iv) higher incidence of more sophisticated television equipment (video cameras, teletext receivers) and of human technical support in special schools than in hearing-impaired units.

Text composition by the physically disabled: a rate prediction model for scanning input.

Keyboard 'bypass' techniques allow physically disabled people, otherwise unable to use conventional means of text composition, to do so. Ability to compose text is extremely important to the disabled, offering potential for non-speech communication, computer access, creative writing, etc. Consequently, these techniques have received a good deal of attention and many diverse systems have evolved. They all suffer, however, from the drawback of inherently slow input, quite apart from any disability on the part of the user. For this reason, text composition rate (or communication rate) is the major figure of merit. Since many diverse systems and approaches exist, quantitative methods of comparison are required to guide prescription and development of such aids. Only recently have attempts to produce models which predict communication rate been made. This paper extends the earlier model of Rosen and Gooenough-Trepagnier to encompass scanning-input systems. Scanning input is of considerable interest since it can be used by very severely disabled people. The model developed is applied to the comparison of two very different systems: row-column scanning and the 'scanning Microwriter'. According to the model, row-scanning is very much faster than the scanning Microwriter when a letter-frequency arrangement of the character selections is used. The relation of the model to classical information theory, treating the disabled user as an information source, is also explored.

Statistical clustering procedures applied to low-cost speech recognition.

A new generation of low-cost speech recognition devices are appearing, which offer much promise for useful applications in biomedical engineering. These devices are statistical pattern recognizers. Input utterances are classified by comparison with a set of templates derived during 'speaker training'. For useful application of these devices, recognition accuracy must be high and speaker training must not be unacceptably complicated or tedious. This paper investigates techniques which consider the statistical nature of the input utterances, used to improve recognition accuracy. Word classification based on the Mahalonobis distance metric, and using templates derived from cluster analysis of the training inputs, was found to give results superior to the other strategies studied. This classifier was unsuitable for implementation in a real-time, low-cost system but the principle of clustering was successfully applied to produce an adaptive system which tracked changes in the user's voice. This allowed training to be drastically simplified by updating templates during normal operation. The adaptive system achieved 98.8% recognition accuracy on a 32 word vocabulary compared to 94.8% without adaptation.

Tracking of complex binaural images.

Subjects were presented with an initially stationary binaural image formed by the fusion of two identical pulses, without interaural time difference (ITD), at the two ears. The image was then made to traverse the subject's auditory perceptual space by introducing ITD, varying linearly with time, under computer control. The direction of movement, i.e. towards the right or left ear, could be reversed by the listener, by pressing a button. Subjects were requested to keep the image central, by pressing the button when they judged that deviation from subjective centre had occurred. Experiments of this type can be considered as analogous to Békésy audiometry, where the subject automatically traces his threshold of hearing, in that here the listener traces out his auditory perceptual centre as it varies with time. Hence, equivalent analyses to those employed for Békésy audiometry are possible. Subsequent to the initial part of the experiment, an additional pulse was added to one channel, preceding the original pulse, to form a pulse pair. The monaural masking of the original pulse by this additional pulse thus acts to shift the pre-existing binaural image. The effect of varying the amplitude and onset time of the masking pulse, relative to the original pulse, on the Békésy-type trace was examined.