Pulse-stream VLSI neural networks mixing analog and digitaltechniques

The pulse-stream technique, which represents neural states as sequences of pulses, is reviewed. Several general issues are raised, and generic methods appraised, for pulsed encoding, arithmetic, and intercommunication schemes. Two contrasting synapse designs are presented and compared. The first is based on a fully analog computational form in which the only digital component is the signaling mechanism itself-asynchronous, pulse-rate encoded digital voltage pulses. In this circuit, multiplication occurs in the voltage/current domain. The second design uses more conventional digital memory for weight storage, with synapse circuits based on pulse stretching. Integrated circuits implementing up to 15000 analog, fully programmable synaptic connections are described. A demonstrator project is described in which a small robot localization network is implemented using asynchronous, analog, pulse-stream devices.     

A dynamical model for generating synthetic electrocardiogram signals

A dynamical model based on three coupled ordinary differential equations is introduced which is capable of generating realistic synthetic electrocardiogram (ECG) signals. The operator can specify the mean and standard deviation of the heart rate, the morphology of the PQRST cycle, and the power spectrum of the RR tachogram. In particular, both respiratory sinus arrhythmia at the high frequencies (HFs) and Mayer waves at the low frequencies (LFs) together with the LF/HF ratio are incorporated in the model. Much of the beat-to-beat variation in morphology and timing of the human ECG, including QT dispersion and R-peak amplitude modulation are shown to result. This model may be employed to assess biomedical signal processing techniques which are used to compute clinical statistics from the ECG.     

Detection of Ectopic Beats in the Electrocardiogram Using an Auto-Associative Neural Network

Abnormal rhythms of the heart are often preceded by the occurrence of ectopic beats. These are difficult to detect as their shape is not very different from that of a normal QRS complex, the main feature in the electrocardiogram. We show how an auto-asociative multi-layer perceptron can be trained to detect normal beats only, so that the subtle abnormalities in shape of ectopic beats become clearly identifiable. This is a generic detector of abnormal beats (i.e. beats whose morphology is different from that of a normal beat) and we use ventricular ectopic beats to illustrate the performance of the algorithm. We also propose a new parameter, the variance ratio, to monitor the progress of learning in an auto-associative network.     

The relationship between symptoms and blood pressure during maintenance hemodialysis

Intradialytic hypotension (IDH) is a detrimental complication of maintenance hemodialysis, but how it is defined and reported varies widely in the literature. European Best Practice Guideline and Kidney Disease Outcomes Quality Initiative guidelines require symptoms and a mitigating intervention to fulfill the diagnosis, but morbidity and mortality outcomes are largely based on blood pressure alone. Furthermore, little is known about the incidence of asymptomatic hypotension, which may be an important cause of hypoperfusion injury and impaired outcome. Seventy‐seven patients were studied over 456 dialysis sessions. Blood pressure was measured at 15‐minute intervals throughout the session and compared with post‐dialysis symptom questionnaire results using mixed modeling to adjust for repeated measures in the same patient. The frequency of asymptomatic hypotension was estimated by logistic regression using a variety of commonly cited blood pressure metrics that describe IDH. In 113 sessions (25%) where symptoms were recorded on the questionnaire, these appear not to have been reported to dialysis staff. When symptoms were reported (293 sessions [64%]), an intervention invariably followed. Dizziness and cramp were strongly associated with changes in systolic blood pressure (SBP), but not diastolic blood pressure. Nausea occurred more frequently in younger patients but was not associated with falls in blood pressure. Thresholds that maximized the probability of an intervention rather than a session remaining asymptomatic were SBP <100 mmHg or a 20% reduction in SBP from baseline. The probability of SBP falling to <100 mmHg in an asymptomatic session was 0.23. Symptoms are frequently not reported by patients who are hypotensive during hemodialysis, which leads to an underestimation of IDH if symptom‐based definitions are used. A revised definition of IDH excluding patient‐reported symptoms would be in line with literature reporting morbidity and mortality outcomes and include sessions in which potentially detrimental asymptomatic hypotension occurs.     

Multi-sensor fusion for robust computation of breathing rate

Respiratory information can be obtained from the changes in electrical impedance across the chest, the electrocardiogram or the changes in light absorption across the finger. Running estimates of the breathing rate from each of these can be obtained from independent Kalman filters. A robust estimate of breathing rate is derived by fusing the outputs of these independent Kalman filters using a weighting calculated from the squared differences between each prediction from the filters and the corresponding measurements.     

Reflectance pulse oximetry measurements from the retinal fundus

Conventional transmission pulse oximetry is a noninvasive technique for the continuous monitoring of arterial oxygen saturation (SaO₂) from peripheral vascular beds such as the finger tip or earlobe. It is proposed to exploit the unique transparency of the ocular media to make reflectance pulse oximetry measurements on the retinal fundus. This technique potentially offers significant advantages over conventional pulse oximetry, primarily in the ability to monitor cerebral, as opposed to peripheral, oxygen saturation. An in vitro system has been developed to simulate the retinal circulation and ocular optics. This system consists of a flexible cuvette located in a model eye and an extracorporeal blood circuit to simulate arterial blood flow. The system was used to investigate the relationship between SaO₂ and the R/IR ratio in reflectance pulse oximetry. To enable in vivo measurements to be made, a standard haptic contact lens was modified to hold the pulse oximeter probe in front of the pupil. In a preliminary study, the lens was fitted to an awake volunteer and cardiac-synchronous signals were detected by the retinal pulse oximeter     

Imaging spatial distributions of resistivity?an alternative approach

We propose a method to obtain images of electrical resistivity within body segments, which is directly applicable to cranial imaging in the newborn. The method is based on sensitivity measurements and uses a weighted back-projection algorithm. Preliminary results from a laboratory phantom are presented.     

Static and dynamic novelty detection methods for jet engine health monitoring

Novelty detection requires models of normality to be learnt from training data known to be normal. The first model considered in this paper is a static model trained to detect novel events associated with changes in the vibration spectra recorded from a jet engine. We describe how the distribution of energy across the harmonics of a rotating shaft can be learnt by a support vector machine model of normality. The second model is a dynamic model partially learnt from data using an expectation-maximization-based method. This model uses a Kalman filter to fuse performance data in order to characterize normal engine behaviour. Deviations from normal operation are detected using the normalized innovations squared from the Kalman filter.     

Neural Networks for Mobile Robot Localisation using Infra-Red Range Sensing

For a robot to be fully autonomous whilst mobile, it is necessary for it to be able to determine its position in its environment. Most of the work on this problem has concentrated on using geometrical techniques which are typically implemented as part of a Kalman filter cycle. This paper examines the possibility of using a neural network to assist in the task of estimating the position of the robot. This is beneficial because it does not require beacons to be placed in the environment or the use of an explicit map of the environment. It does not require knowledge of the previous estimate of the robot’s position. In this paper, Radial Basis Function networks and Multi-Layer Perceptrons are trained to estimate the functional relationship between preprocessed range sensor data and the position of the robot. This approach is assessed using both simulated and real range data.     

Neural networks-extraordinary variation

There is an extraordinary variation in the implementations of neural networks, ranging from electronic circuits, digital or analog, to novel devices and optical implementations. To provide a wide perspective of different developments, we selected four projects to summarize. These projects were originally presented at Microneuro 1994. The first two projects are optical and optoelectronic implementations. While the majority of neural nets are built in silicon, an active research community works on optical networks. The high connectivity of neural nets makes communication among the processors one of the main difficulties in implementation. With optical beams, a large number of elements can be addressed in parallel, making optics an attractive alternative to silicon. D.C. Burns and coauthors describe a combination of optics and electronics. The authors have built an optical input plane for a neural net so that whole images with tens of thousands of pixels can be entered into a network in parallel. S.R. Skinner, J.E. Steck, and E.C. Behrman present an all optical network, where optically nonlinear materials perform not only the communication but also the calculations. Learning remains a tricky problem for a hardware implementation, in particular for analog circuits. Popular learning algorithms, such as backpropagation, require a high computational resolution. How to implement learning techniques in low-resolution electronics is the subject of intense research. G. Cairns and L. Tarassenko address this issue by comparing the required precision for different learning schemes. The final project is a digital circuit implementing a self-organizing feature map, an unsupervised learning technique. In this example, computational resolution is also a major problem. Before building a chip, researchers experimented extensively to determine the minimum resolution required for good results