André van Schaik

Neuromorphic silicon neuron circuits

Hardware implementations of spiking neurons can be extremely useful for a large variety of applications, ranging from high-speed modeling of large-scale neural systems to real-time behaving systems, to bidirectional brain–machine interfaces. The specific circuit solutions used to implement silicon neurons depend on the application requirements. In this paper we describe the most common building blocks and techniques used to implement these circuits, and present an overview of a wide range of neuromorphic silicon neurons, which implement different computational models, ranging from biophysically realistic and conductance-based Hodgkin–Huxley models to bi-dimensional generalized adaptive integrate and fire models. We compare the different design methodologies used for each silicon neuron design described, and demonstrate their features with experimental results, measured from a wide range of fabricated VLSI chips.

A new EEG recording system for passive dry electrodes

OBJECTIVE We present a new, low power EEG recording system with an ultra-high input impedance that enables the use of long-lasting, passive dry electrodes. It incorporates Bluetooth wireless connectivity and is designed to be suitable for long-term monitoring during daily activities. METHODS The new EEG system is compared to a standard and clinically available reference EEG system using wet electrodes in three separate sets of experiments. In the first two experiments, each dry electrode was surrounded by four standard wet electrodes and the alpha and mu-rhythms were recorded. In the third experiment, serial monopolar (referred to the left ear) recordings of flash visual evoked potential were performed using the new EEG system and a reference system. RESULTS These experiments showed that the signal recorded using the new EEG system is almost identical to that recorded with standard clinical EEG equipment; our measurements showed that the correlation coefficient between the dry electrode recordings and the average of the four standard electrodes surrounding each dry electrode is greater than 0.85. CONCLUSION We conclude that the new EEG system performs similarly to reference EEG systems, while providing the advantages of portability, ease of application and minimal scalp preparation. SIGNIFICANCE The proposed system using passive dry electrodes suitable for single use while performing as good as standard EEG equipment provides ease of application and minimal scalp preparation.

The role of high frequencies in speech localization

This study measured the accuracy with which human listeners can localize spoken words. A broadband (300 Hz-16 kHz) corpus of monosyllabic words was created and presented tolisteners using a virtual auditory environment. Localization was examined for 76 locations ona sphere surrounding the listener. Experiment 1 showed that low-pass filtering the speech sounds at 8 kHz degraded performance, causing an increase in polar angle errors associated with the cone of confusion. In experiment 2 it was found that performance in fact varied systematically with the level of the signal above 8 kHz. Although the lower frequencies (below 8 kHz) are known to be sufficient for accurate speech recognition in most situations, these results demonstrate that natural speech contains information between 8 and 16 kHz that is essential for accurate localization.

Event-based 64-channel binaural silicon cochlea with Q enhancement mechanisms

This paper describes an event-based binaural silicon cochlea aimed at spatial audition and auditory scene analysis. The chip has a matched pair of 64-stage cascaded analog second-order filter banks with 512 pulse-frequency modulated (PFM) address-event representation (AER) outputs. The spectral selectivity is sharpened through 2 different on-chip methods: an on-chip local Q DAC and an on-chip spatial sharpening through nearest neighbour lateral inhibition. The fabricated chip in a 4-metal 2-poly 0.35um CMOS process consumes peak 25mW power for the digital circuits and 33mW for the analog core. Dynamic range to produce PFM output is 36dB (25mVpp to 1500mVpp at microphone preamp output). Event timing jitter is 2us for 250mVpp input. The peak output bandwidth is 10M events per second (eps) but typical speech scenarios show rates of 20keps.

An FPGA Implementation of a Polychronous Spiking Neural Network with Delay Adaptation

We present an FPGA implementation of a re-configurable, polychronous spiking neural network with a large capacity for spatial-temporal patterns. The proposed neural network generates delay paths de novo, so that only connections that actually appear in the training patterns will be created. This allows the proposed network to use all the axons (variables) to store information. Spike Timing Dependent Delay Plasticity is used to fine-tune and add dynamics to the network. We use a time multiplexing approach allowing us to achieve 4096 (4k) neurons and up to 1.15 million programmable delay axons on a Virtex 6 FPGA. Test results show that the proposed neural network is capable of successfully recalling more than 95% of all spikes for 96% of the stored patterns. The tests also show that the neural network is robust to noise from random input spikes.

Benefit from spatial separation of multiple talkers in bilateral hearing-aid users: Effects of hearing loss, age, and cognition

Abstract To study the spatial hearing abilities of bilateral hearing-aid users in multi-talker situations, 20 subjects received fittings configured to preserve acoustic cues salient for spatial hearing. Following acclimatization, speech reception thresholds (SRTs) were measured for three competing talkers that were either co-located or spatially separated along the front-back or left-right dimension. In addition, the subjects’ working memory and attentional abilities were measured. Left-right SRTs varied over more than 14 dB, while front-back SRTs varied over more than 8 dB. Furthermore, significant correlations were observed between left-right SRTs, age, and low-frequency hearing loss, and also between front-back SRTs, age, and high-frequency aided thresholds. Concerning cognitive effects, left-right performance was most strongly related to attentional abilities, while front-back performance showed a relation to working memory abilities. Altogether, these results suggest that, due to raised hearing thresholds and aging, hearing-aid users have reduced access to interaural and monaural spatial cues as well as a diminished ability to ‘enhance’ a target signal by means of top-down processing. These deficits, in turn, lead to impaired functioning in complex listening environments. Sumario Para estudiar las habilidades auditivas espaciales de usuarios de audífonos bilaterales en situaciones de hablantes múltiples, 20 sujetos recibieron auxiliares configurados para preservar las más destacadas claves acústicas de la audición espacial. Después de un período de adaptación, se midieron los umbrales de recepción del habla (SRTs) con tres hablantes en competencia que fueron colocados cerca o espacialmente separados en las dimensiones frente-atrás o izquierda-derecha. Además, se midieron las habilidades de memoria activa y atención. Los SRTs izquierda-derecha variaron más de 14 dB mientras que los SRTs frente-atrás, variaron más de 8 dB. Más aún, se observaron correlaciones significativas entre SRTs frente-atrás, edad y pérdida auditiva en las frecuencias graves y también entre SRTs frente-atrás, edad y umbrales con auxiliares auditivos en las frecuencias agudas. En relación con los efectos cognitivos, el rendimiento izquierda-derecha se relacionó más firmemente con habilidades de atención, mientras que el rendimiento frente-atrás, mostró relación con habilidades de memoria activa. En general, estos resultados sugieren que, debido a la elevación de los umbrales auditivos y el envejecimiento, los usuarios de auxiliares auditivos tienen un acceso reducido a las claves espaciales interaurales y monoaurales así como una habilidad disminuida para “mejorar” una señal blanco por medio del procesamiento arriba-abajo. Estos déficits, a su vez, conducen a un funcionamiento disminuido en ambientes de escucha complejos.

A log-domain implementation of the Izhikevich neuron model

We present an implementation of the Izhikevich neuron model which uses two first-order log-domain low-pass filters and two translinear multipliers. The neuron consists of a leaky-integrate-and-fire core, a slow adaptive state variable and quadratic positive feedback. Simulation results show that this neuron can emulate different spiking behaviours observed in biological neurons.

Separation of concurrent broadband sound sources by human listeners.

The effect of spatial separation on the ability of human listeners to resolve a pair of concurrent broadband sounds was examined. Stimuli were presented in a virtual auditory environment using individualized outer ear filter functions. Subjects were presented with two simultaneous noise bursts that were either spatially coincident or separated (horizontally or vertically), and responded as to whether they perceived one or two source locations. Testing was carried out at five reference locations on the audiovisual horizon (0 degrees, 22.5 degrees, 45 degrees, 67.5 degrees, and 90 degrees azimuth). Results from experiment 1 showed that at more lateral locations, a larger horizontal separation was required for the perception of two sounds. The reverse was true for vertical separation. Furthermore, it was observed that subjects were unable to separate stimulus pairs if they delivered the same interaural differences in time (ITD) and level (ILD). These findings suggested that the auditory system exploited differences in one or both of the binaural cues to resolve the sources, and could not use monaural spectral cues effectively for the task. In experiments 2 and 3, separation of concurrent noise sources was examined upon removal of low-frequency content (and ITDs), onset/offset ITDs, both of these in conjunction, and all ITD information. While onset and offset ITDs did not appear to play a major role, differences in ongoing ITDs were robust cues for separation under these conditions, including those in the envelopes of high-frequency channels.

An ultra-high input impedance ECG amplifier for long-term monitoring of athletes.

We present a new, low-power electrocardiogram (ECG) recording system with an ultra-high input impedance that enables the use of long-lasting, dry electrodes. The system incorporates a low-power Bluetooth module for wireless connectivity and is designed to be suitable for long-term monitoring during daily activities. The new system using dry electrodes was compared with a clinically approved ECG reference system using gelled Ag/AgCl electrodes and performance was found to be equivalent. In addition, the system was used to monitor an athlete during several physical tasks, and a good quality ECG was obtained in all cases, including when the athlete was totally submerged in fresh water.

Fast, Simple and Accurate Handwritten Digit Classification by Training Shallow Neural Network Classifiers with the ‘Extreme Learning Machine’ Algorithm

Recent advances in training deep (multi-layer) architectures have inspired a renaissance in neural network use. For example, deep convolutional networks are becoming the default option for difficult tasks on large datasets, such as image and speech recognition. However, here we show that error rates below 1% on the MNIST handwritten digit benchmark can be replicated with shallow non-convolutional neural networks. This is achieved by training such networks using the ‘Extreme Learning Machine’ (ELM) approach, which also enables a very rapid training time (∼ 10 minutes). Adding distortions, as is common practise for MNIST, reduces error rates even further. Our methods are also shown to be capable of achieving less than 5.5% error rates on the NORB image database. To achieve these results, we introduce several enhancements to the standard ELM algorithm, which individually and in combination can significantly improve performance. The main innovation is to ensure each hidden-unit operates only on a randomly sized and positioned patch of each image. This form of random ‘receptive field’ sampling of the input ensures the input weight matrix is sparse, with about 90% of weights equal to zero. Furthermore, combining our methods with a small number of iterations of a single-batch backpropagation method can significantly reduce the number of hidden-units required to achieve a particular performance. Our close to state-of-the-art results for MNIST and NORB suggest that the ease of use and accuracy of the ELM algorithm for designing a single-hidden-layer neural network classifier should cause it to be given greater consideration either as a standalone method for simpler problems, or as the final classification stage in deep neural networks applied to more difficult problems.