Mark W. White

A stochastic model of the electrically stimulated auditory nerve: single-pulse response

Most models of neural response to electrical stimulation, such as the Hodgkin-Huxley equations, are deterministic, despite significant physiological evidence for the existence of stochastic activity. For instance, the range of discharge probabilities measured in response to single electrical pulses cannot be explained at all by deterministic models. Furthermore, there is growing evidence that the stochastic component of auditory nerve response to electrical stimulation may be fundamental to functionally significant physiological and psychophysical phenomena. In this paper authors present a simple and computationally efficient stochastic model of single-fiber response to single biphasic electrical pulses, based on a deterministic threshold model of action potential generation. Comparisons with physiological data from cat auditory nerve fibers are made, and it is shown that the stochastic model predicts discharge probabilities measured in response to single biphasic pulses more accurately than does the equivalent deterministic model. In addition, physiological data show an increase in stochastic activity with increasing pulse width of anodic/cathodic biphasic pulses, a phenomenon not present for monophasic stimuli. Those and other data from the auditory nerve are then used to develop a population model of the total auditory nerve, where each fiber is described by the single-fiber model.

A stochastic model of the electrically stimulated auditory nerve: pulse-train response

The single-pulse model of the companion paper [see ibid., vol. 46, no. 6, p. 617-29, 1999] is extended to describe responses to pulse trains by introducing a phenomenological refractory mechanism. Comparisons with physiological data from cat auditory nerve fibers are made for pulse rates between 100 and 800 pulses/s. First, it is shown that both the shape and slope of mean discharge rate curves are better predicted by the stochastic model than by the deterministic model. Second, while interpulse effects such as refractory effects do indeed increase the dynamic range at higher pulse rates, both the physiological data and the model indicate that much of the dynamic range for pulse-train stimuli is due to stochastic activity. Third, it is shown that the stochastic model is able to predict the general magnitude and behavior of variance in discharge rate as a function of pulse rate, while the deterministic model predicts no variance at all.

The effects of stochastic neural activity in a model predicting intensity perception with cochlear implants: low-rate stimulation

Most models of auditory nerve response to electrical stimulation are deterministic, despite significant physiological evidence for stochastic activity. Furthermore, psychophysical models and analyses of physiological data using deterministic descriptions do not accurately predict many psychophysical phenomena. Here, the authors investigate whether inclusion of stochastic activity in neural models improves such predictions. To avoid the complication of interpulse interactions and to enable the use of a simpler and faster auditory nerve model the authors restrict their investigation to single pulses and low-rate (<200 pulses/s) pulse trains. They apply signal detection theory to produce direct predictions of behavioral threshold, dynamic range and intensity difference limen. Specifically, the authors investigate threshold versus pulse duration (the strength-duration characteristics), threshold and uncomfortable loudness (and the corresponding dynamic range) versus phase duration, the effects of electrode configuration on dynamic range and on strength-duration, threshold versus number of pulses (the temporal-integration characteristics), intensity difference limen as a function of loudness, and the effects of neural survival on these measures. For all psychophysical measures investigated, the inclusion of stochastic activity in the auditory nerve model was found to produce more accurate predictions.

Deafness alters auditory nerve fibre responses to cochlear implant stimulation

Here we characterized the relationship between duration of sensorineural hearing loss and the response of the auditory nerve to electrical stimulus rate. Electrophysiological recordings were made from undeafened guinea pigs and those ototoxically deafened for either 5 weeks or 6 months. Auditory neuron survival decreased significantly with the duration of deafness. Extracellular recordings were made from auditory nerve fibres responding to biphasic, charge‐balanced current pulses delivered at rates of 20 and 200 pulses/s via a monopolar scala tympani stimulating electrode. The response to 20 pulses/s electrical stimulation of the deafened cochlea exhibited a decrease in spike latency, unaltered temporal jitter and unaltered dynamic range (of nerve firing rate against stimulus current), and a reduction in threshold after 6 months of deafness. The response to a 200‐pulse/s stimulus was similar except that the dynamic range was greater than with 20 pulses/s and was also greater in deafened animals than in undeafened animals. Deafness and pulse rate are related; in deaf animals spike recovery appears to be complete between successive stimulus pulses at a low rate (20 pulses/s), but incomplete between pulses at a moderate pulse rate (200 pulses/s). These results suggest that changes in the function of individual auditory nerve fibres after deafness may affect clinical responses during high‐rate stimulation such as that used in contemporary speech processing strategies, but not during lower rate stimulation such as that used to record evoked potentials.

Autonomous mobile robot global self‐localization using Kohonen and region‐feature neural networks

This article presents and compares two neural network-based approaches to global selflocalization (GSL) for autonomous mobile robots using: (1) a Kohonen neural network; and (2) a region-feature neural network (RFNN). Both approaches categorize discrete regions of space (topographical nodes) in a manner similar to optical character recognition (OCR). That is, the mapped sonar data assumes the form of a character unique to that region. Hence, it is believed that an autonomous vehicle can determine which room it is in from sensory data gathered from exploration. With a robust exploration routine, the GSL solution can be time-, translation-, and rotation-invariant. The GSL solution can also become independent of the mobile robot used to collect the sensor data. This suggests that a single robot can transfer its knowledge of various learned regions to other mobile robots. The classification rate of both approaches are comparable and, thus, worthy of presentation. The observed pros and cons of both approaches are also discussed.  1997 John Wiley & Sons, Inc.

Examining the auditory nerve fiber response to high rate cochlear implant stimulation: chronic sensorineural hearing loss and facilitation.

Neural prostheses, such as cochlear and retinal implants, induce perceptual responses by electrically stimulating sensory nerves. These devices restore sensory system function by using patterned electrical stimuli to evoke neural responses. An understanding of their function requires knowledge of the nerves responses to relevant electrical stimuli as well as the likely effects of pathology on nerve function. We describe how sensorineural hearing loss (SNHL) affects the response properties of single auditory nerve fibers (ANFs) to electrical stimuli relevant to cochlear implants. The response of 188 individual ANFs were recorded in response to trains of stimuli presented at 200, 1,000, 2,000, and 5,000 pulse/s in acutely and chronically deafened guinea pigs. The effects of stimulation rate and SNHL on ANF responses during the 0-2 ms period following stimulus onset were examined to minimize the influence of ANF adaptation. As stimulation rate increased to 5,000 pulse/s, threshold decreased, dynamic range increased and first spike latency decreased. Similar effects of stimulation rate were observed following chronic SNHL, although onset threshold and first spike latency were reduced and onset dynamic range increased compared with acutely deafened animals. Facilitation, defined as an increased nerve excitability caused by subthreshold stimulation, was observed in both acute and chronic SNHL groups, although the magnitude of its effect was diminished in the latter. These results indicate that facilitation, demonstrated here using stimuli similar to those used in cochlear implants, influences the ANF response to pulsatile electrical stimulation and may have important implications for cochlear implant signal processing strategies.

Feature extraction by genetic algorithms for neural networks in breast cancer classification

In todays world, in which computerized recognition is expanding its horizons in the field of medicine, breast cancer classification is receiving wide attention. In this application, artificial neural networks have achieved reasonable recognition rates. However, to improve performance, a technique is needed to screen the features of the input data, to extract the important ones and suppress those that are irrelevant. Although neural networks do have this capability to some extent, here it is shown that by using a hybrid genetic algorithm and neural network (GANN), the feature extraction can be performed more effectively. Another advantage of augmenting NN training with a GA is that the extracted features using GA are explicit and perceivable. Although the authors evaluated the technique using breast cancer data, the methodology is designed to handle any other kind of classification task.

Image segmentation with neurocomputers

Our ultimate application interest is the automated understanding of images, especially non-biological images such as range or synthetic aperture radar. In this report, we describe one aspect of the processing of such images, segmentation, and show that one can design a neural network to perform this computation. An important step14 in image analysis is segmentation. An image is an array of discrete sampled values called pixels. An image is understood when the objects portrayed in that image are recognized3 or at least characterized. An object is characterized in terms of the segments of the image that it subtends. An image segment1,6 is a collection of pixels that satisfies certain conditions of adjacency and similarity.

Renewal-Process Approximation of a Stochastic Threshold Model for Electrical Neural Stimulation

In a recent set of modeling studies we have developed a stochastic threshold model of auditory nerve response to single biphasic electrical pulses (Bruce et al., 1999c) and moderate rate (less than 800 pulses per second) pulse trains (Bruce et al., 1999a). In this article we derive an analytical approximation for the single-pulse model, which is then extended to describe the pulse-train model in the case of evenly timed, uniform pulses. This renewal-process description provides an accurate and computationally efficient model of electrical stimulation of single auditory nerve fibers by a cochlear implant that may be extended to other forms of electrical neural stimulation.

Self-organizing geometric certainty maps: a compact and multifunctional approach to map building, place recognition and motion planning

In this paper we show how a self-organizing Kohonen neural network can use hyperellipsoid clustering (HEC) to build maps from actual sonar data. Since the HEC algorithm uses the Mahalanobis distance, the elongated shapes (typical of sonar data) can be learned. The Mahalanobis distance metric also gives a stochastic measurement of a data points association with a node. Hence, the HEC Kohonen can be used to build topographical maps and to recognize its own topographical cites for self-localization. The number of nodes can also be regulated in a self-organizing manner by using the Kolmogorov-Smirnov (KS) test for cluster compactness. The KS test determines whether a node should be divided (mitosis) or pruned completely. By incorporating principal component analysis, the HEC Kohonen can handle problems with several dimensions (3D, time-series, etc.). The HEC Kohonen is also multifunctional in that it accommodates compact geometric motion planning and self-referencing algorithms. It can also be used to solve a host of other pattern recognition problems.