Performing brain simulations that match the size and dynamic nature of real brains is arduous but essential for understanding neural mechanisms underlying animal behaviour. To address this challenge, this paper proposes an 8-bit floating-point format (minifloat) for the purpose of efficient simulation of biological spiking neural networks in digital hardware. We present models employing minifloat variables, as well as multiplication and addition arithmetics. Other low-precision data types are considered to elucidate the feasibility and advantages of minifloat. Despite the inherent floating-point errors, minifloat models effectively simulate balanced networks that reproduce activity patterns observed in cortical networks. Our results suggest that low-precision floating-point data types are a viable alternative for spiking neural network simulations that could also improve scalability and data throughput.

Animal nervous systems can detect changes in their environments within hundredths of a second. They do so by discerning abrupt shifts in sensory neural activity. Many neuroscience studies have employed change-point detection (CPD) algorithms to estimate such abrupt shifts in neural activity. But very few studies have suggested that spiking neurons themselves are online change-point detectors. We show that a leaky integrate-and-fire (LIF) neuron implements an online CPD algorithm for a compound Poisson process. We quantify the CPD performance of an LIF neuron under various regions of its parameter space. We show that CPD can be a recursive algorithm where the output of one algorithm can be input to another. Then we show that a simple feedforward network of LIF neurons can quickly and reliably detect very small changes in input spiking rates. For example, our network detects a 5% change in input rates within 20 ms on average, and false-positive detections are extremely rare. In a rigorous statistical context, we interpret the salient features of the LIF neuron: its membrane potential, synaptic weight, time constant, resting potential, action potentials, and threshold. Our results potentially generalize beyond the LIF neuron model and its associated CPD problem. If spiking neurons perform change-point detection on their inputs, then the electrophysiological properties of their membranes must be related to the spiking statistics of their inputs. We demonstrate one example of this relationship for the LIF neuron and compound Poisson processes and suggest how to test this hypothesis more broadly. Maybe neurons are not noisy devices whose action potentials must be averaged over time or populations. Instead, neurons might implement sophisticated, optimal, and online statistical algorithms on their inputs.

This paper presents the Field Programmable Gate Array (FPGA) implementation of an event-driven unsupervised Feature Extraction using Adaptive Selection Thresholds (FEAST) algorithm for pattern recognition tasks. The novelty of the design lies in splitting the FEAST learning rule into two different sets of tasks and executing them independently in a time-multiplexed fashion, using a minimum number of hardware resources. The proposed hardware architecture, operated at 200 MHz clock frequency, can process 183 × 103events/sec in training mode and 196×103events/sec in inference mode. The FEAST hardware model was tested with the Poker DVS dataset, obtaining a test accuracy of 95%.

The setup includes a host PC, camera and microphone sensors, and a Pynq-Z2 FPGA board. The neuromorphic cochlear model and RAMAN accelerator for neural network inference are deployed on the FPGA. The ARM processor on the FPGA sends the image received, cochleagram and the classified outputs to the PC to be visualized.

Deep Neural Network (DNN) based inference at the edge is challenging as these compute, and data-intensive algorithms need to be implemented at low cost and low power while meeting the latency constraints of the target applications. Sparsity, in both activations and weights inherent to DNNs, is a key knob to leverage. In this paper, we present RAMAN, a Re-configurable and spArse tinyML Accelerator for infereNce on edge, architected to exploit the sparsity to reduce area (storage), power as well as latency. RAMAN can be configured to support a wide range of DNN topologies -consisting of different convolution layer types and a range of layer parameters (feature-map size and the number of channels). RAMAN can also be configured to support accuracy vs. power/latency tradeoffs using techniques deployed at compile-time and run-time. We present the salient features of the architecture, provide implementation results and compare the same with the state-of-the-art. RAMAN employs novel dataflow inspired by Gustavson’s algorithm that has optimal input activation (IA) and output activation (OA) reuse to minimize memory access and the overall data movement cost. The dataflow allows RAMAN to locally reduce the partial sum (Psum) within a processing element array to eliminate the Psum writeback traffic. Additionally, we suggest a method to reduce peak activation memory by overlapping IA and OA on the same memory space, which can reduce storage requirements by up to 50%. RAMAN was implemented on a low-power and resource-constrained Efinix Ti60 FPGA with 37.2K LUTs and 8.6K register utilization. RAMAN processes all layers of the MobileNetV1 model at 98.47 GOp/s/W and the DS-CNN model at 79.68 GOp/s/W by leveraging both weight and activation sparsity.

Neural action potentials (APs) are difficult to interpret as signal encoders and/or computational primitives. Their relationships with stimuli and behaviors are obscured by the staggering complexity of nervous systems themselves. We can reduce this complexity by observing that "simpler" neuron-less organisms also transduce stimuli into transient electrical pulses that affect their behaviors. Without a complicated nervous system, APs are often easier to understand as signal/response mechanisms. We review examples of nonneural stimulus transductions in domains of life largely neglected by theoretical neuroscience: bacteria, protozoans, plants, fungi, and neuron-less animals. We report properties of those electrical signals-for example, amplitudes, durations, ionic bases, refractory periods, and particularly their ecological purposes. We compare those properties with those of neurons to infer the tasks and selection pressures that neurons satisfy. Throughout the tree of life, nonneural stimulus transductions time behavioral responses to environmental changes. Nonneural organisms represent the presence or absence of a stimulus with the presence or absence of an electrical signal. Their transductions usually exhibit high sensitivity and specificity to a stimulus, but are often slow compared to neurons. Neurons appear to be sacrificing the specificity of their stimulus transductions for sensitivity and speed. We interpret cellular stimulus transductions as a cell's assertion that it detected something important at that moment in time. In particular, we consider neural APs as fast but noisy detection assertions. We infer that a principal goal of nervous systems is to detect extremely weak signals from noisy sensory spikes under enormous time pressure. We discuss neural computation proposals that address this goal by casting neurons as devices that implement online, analog, probabilistic computations with their membrane potentials. Those proposals imply a measurable relationship between afferent neural spiking statistics and efferent neural membrane electrophysiology.

We will demonstrate an FPGA implementation of an event- based vision sensor using a commercially available frame- based camera. The demonstration setup consists of the host PC that includes the Quartus Prime software which is used to program and configure the FPGA, a commercially available 8-megapixel MIPI (Mobile Industry Processor Interface) camera kit, and a Cyclone V DE10-nano FPGA board, as shown in Fig. 1. The camera kit has been used to capture conventional frame- based images [1]. It is mounted on the FPGA board via the 2×20 pin general-purpose input-output port connector interface of the FPGA board. The FPGA board is used to process the digital pixel data, which are received from the camera, and to generate events. The generated events are displayed on a VGA monitor with predefined colors for each event behavior.

We are currently experiencing the dawn of what is known as the fourth industrial revolution. At the center of this historical happening, as one of the key enabling technologies, lies a discipline that deals with data and whose goal is to extract information and related knowledge that is hidden in it, in order to make predictions and, subsequently, take decisions. Machine learning (ML) is the name that is used as an umbrella to cover a wide range of theories, methods, algorithms, and architectures that are used to this end. The articles in this special issue cover promising developments in the related areas of machine learning and deep neural networks and offers possible paths for the future.

We propose a biologically inspired binaural sound localization system using a deep convolutional neural network (CNN) for reverberant environments. It utilizes a binaural Cascade of Asymmetric Resonators with Fast-Acting Compression (CAR-FAC) cochlear system to analyze binaural signals, a lateral inhibition function to sharpen temporal information of cochlear channels, and instantaneous correlation function on the two cochlear channels to encode binaural cues. The generated 2-D instantaneous correlation matrix (correlogram) encodes both interaural phase difference (IPD) cues and spectral information in a unified framework. Additionally, a sound onset detector is exploited to generate the correlograms only during sound onsets to remove interference from echoes. The onset correlograms are analyzed using a deep CNN for regression to the azimuthal angle of the sound. The proposed system was evaluated using experimental data in a reverberant environment, and displayed a root mean square localization error (RMSE) of 3.68° in the -90° to 90° range.

Filters in cochlea models use different coefficients to break sound into a two-dimensional time-frequency representation. On digital hardware with a single sampling rate, the number of bits required to represent these coefficients require substantial computational resources such as memory storage. In this paper, we present a cochlea model operating at multiple sampling rates. As a result, fewer bits are required to represent filter coefficients on hardware as opposed to all the filters operating at a single sampling rate. Additionally, with a 108-filter cochlea implementation, up to nine times fewer coefficients are used than a single sampling rate approach across all filter sections. We present an implementation of 108 filters in Matlab and on an Altera Cyclone V FPGA with a low logic level utilization of 2.57%. Our model can thus be extended to include other auditory processing models such as loudness, pitch perception and timbre recognition on a single FPGA.

We propose a biologically inspired binaural sound localization system for reverberant environments. It uses two 100-channel cochlear models to analyze binaural signals, and each channel of the left cochlea is compared with each channel of the right cochlea in parallel to generate a 2-D instantaneous correlation matrix (correlogram). The correlogram encodes both binaural cues and spectral information in a unified framework. A sound onset detector is used to generate the correlogram only during the sound onsets, and the onset correlogram is analyzed using a linear regression approach as well as an extreme learning machine (ELM). The proposed system is evaluated using experimental data in reverberation environments, and we obtained an average absolute error of 16.5° for linear regression and 12.8° for ELM regression in the -90° to 90° range.

As the interest in event-based vision sensors for mobile and aerial applications grows, there is an increasing need for high-speed and highly robust algorithms for performing visual tasks using event-based data. As event rate and network structure have a direct impact on the power consumed by such systems, it is important to explore the efficiency of the event-based encoding used by these sensors. The work presented in this paper represents the first study solely focused on the effects of both spatial and temporal downsampling on event-based vision data and makes use of a variety of data sets chosen to fully explore and characterize the nature of downsampling operations. The results show that both spatial downsampling and temporal downsampling produce improved classification accuracy and, additionally, a lower overall data rate. A finding is particularly relevant for bandwidth and power constrained systems. For a given network containing 1000 hidden layer neurons, the spatially downsampled systems achieved a best case accuracy of 89.38% on N-MNIST as opposed to 81.03% with no downsampling at the same hidden layer size. On the N-Caltech101 data set, the downsampled system achieved a best case accuracy of 18.25%, compared with 7.43% achieved with no downsampling. The results show that downsampling is an important preprocessing technique in event-based visual processing, especially for applications sensitive to power consumption and transmission bandwidth.

The MNIST dataset has become a standard benchmark for learning, classification and computer vision systems. Contributing to its widespread adoption are the understandable and intuitive nature of the task, the relatively small size and storage requirements and the accessibility and ease-of-use of the database itself. The MNIST database was derived from a larger dataset known as the NIST Special Database 19 which contains digits, uppercase and lowercase handwritten letters. This paper introduces a variant of the full NIST dataset, which we have called Extended MNIST (EMNIST), which follows the same conversion paradigm used to create the MNIST dataset. The result is a dataset that constitutes a more challenging classification task involving letters and digits, and one that shares the same image structure and parameters as the original MNIST task, allowing for direct compatibility with all existing classifiers and systems. Benchmark results using an online ELM algorithm are presented along with a validation of the conversion process through the comparison of the classification results on NIST digits and the MNIST digits.

We present a digital implementation of the Spike Timing Dependent Plasticity (STDP) learning rule. The proposed digital implementation consists of an exponential decay (exp-decay) generator array and a STDP adaptor array. The weight values are stored in a digital memory, and the STDP adaptor w ill send these values to the exp-decay generator using a digital spike of which the duration is modulated according to these values. The exp-decay generator will then generate an exponential decay, which will be used by the STDP adaptor for performing the weight adaption. The exponential decay, which is computational expensive, is efficiently implemented by using a novel stochastic approach. This stochastic approach was fully analysed and characterised. We use a time multiplexing approach to achieve 8192 (8k) virtual STDP adaptors and exp-decay generators with only one physical adaptor and exp-decay generator respectively. We have validated our stochastic STDP approach with measurement results of a balanced excitation experiment. In that experiment, the competition (induced by STDP) between the synapses can establish a bimodal distribution of the synaptic weights: either towards zero (weak) or the maximum (strong) values. Our stochastic approach is therefore ideal for implementing the STDP learning rule in large-scale spiking neural networks running in real time.

This paper compares the classification performance and training times of feed-forward neural networks with one hidden layer trained with the two network weight optimisation methods. The first weight optimisation method used the extreme learning machine (ELM) algorithm. The second weight optimisation method used the back-propagation (BP) algorithm. Using identical network topologies the two weight optimization methods were directly compared using the MNIST handwritten digit recognition database. Our results show that, while the ELM weight optimization method was much faster to train for a given network topology, a much larger number of hidden units were required to provide a comparable performance level to the BP algorithm. When the extra computation due to larger number of hidden units was taken in to account for the ELM network, the computation times of the two methods to achieve a similar performance level was not so different.

In the biological nervous system, large neuronal populations work collaboratively to encode sensory stimuli. These neuronal populations are characterised by a diverse distribution of tuning curves, ensuring that the entire range of input stimuli is encoded. Based on these principles, we have designed a neuromorphic system called a Trainable Analogue Block (TAB), which encodes given input stimuli using a large population of neurons with a heterogeneous tuning curve profile. Heterogeneity of tuning curves is achieved using random device mismatches in VLSI (Very Large Scale Integration) process and by adding a systematic offset to each hidden neuron. Here, we present measurement results of a single test cell fabricated in a 65nm technology to verify the TAB framework. We have mimicked a large population of neurons by re-using measurement results from the test cell by varying offset. We thus demonstrate the learning capability of the system for various regression tasks. The TAB system may pave the way to improve the design of analogue circuits for commercial applications, by rendering circuits insensitive to random mismatch that arises due to the manufacturing process.

The front end of the human auditory system, the cochlea, converts sound signals from the outside world into neural impulses transmitted along the auditory pathway for further processing. The cochlea senses and separates sound in a nonlinear active fashion, exhibiting remarkable sensitivity and frequency discrimination. Although several electronic models of the cochlea have been proposed and implemented, none of these are able to reproduce all the characteristics of the cochlea, including large dynamic range, large gain and sharp tuning at low sound levels, and low gain and broad tuning at intense sound levels. Here, we implement the `Cascade of Asymmetric Resonators' (CAR) model of the cochlea on an FPGA. CAR represents the basilar membrane filter in the `Cascade of Asymmetric Resonators with Fast-Acting Compression' (CAR-FAC) cochlear model. CAR-FAC is a neuromorphic model of hearing based on a pole-zero filter cascade model of auditory filtering. It uses simple nonlinear extensions of conventional digital filter stages that are well suited to FPGA implementations, so that we are able to implement up to 1224 cochlear sections on Virtex-6 FPGA to process sound data in real time. The FPGA implementation of the electronic cochlea described here may be used as a front-end sound analyser for various machine-hearing applications.

We will demonstrate a 100-CAR-section cochlear model running in real time on an FPGA. Although our result suggests that an electronic cochlea with 1224 cochlear sections can be implemented on an average FPGA [1], the data rate limit of USB 2.0 does not permit us to implement more than 100 filter sections and display the output on a PC. Future work will explore alternatives to increase the bandwidth such as a PCI interface or USB 3.0 that will enable us to implement more filter sections. Nonetheless, our work demonstrates the capability of the CAR model to process sound in real-time.

We present an FPGA design framework for large-scale spiking neural networks, particularly the ones with a high-density of connections or all-to-all connections. The proposed FPGA design framework is based on a reconfigurable neural layer, which is implemented using a time-multiplexing approach to achieve up to 200,000 virtual neurons with one physical neuron using only a fraction of the hardware resources in commercial-off-the-shelf FPGAs (even entry level ones). Rather than using a mathematical computational model, the physical neuron was efficiently implemented with a conductance-based model, of which the parameters were randomised between neurons to emulate the variance in biological neurons. Besides these building blocks, the proposed time-multiplexed reconfigurable neural layer has an address buffer, which will generate a fixed random weight for each connection on the fly for incoming spikes. This structure effectively reduces the usage of memory. After presenting the architecture of the proposed neural layer, we present a network with 23 proposed neural layers, each containing 64k neurons, yielding 1.5 M neurons and 92 G synapses with a total spike throughput of 1.2T spikes/s, while running in real-time on a Virtex 6 FPGA.

We present a compact mixed-signal implementation of synaptic plasticity for both Spike Timing Dependent Plasticity (STDP) and Spike Timing Dependent Delay Plasticity (STDDP). The proposed mixed-signal implementation consists of an a VLSI time window generator and a digital adaptor. The weight and delay values are stored in a digital memory, and the adaptor will send these values to the time window generator using a digital spike of which the duration is modulated according to these values. The analogue time window generator will then generate a time window, which is required for the implementation of STDP and STDDP. The digital adaptor will carry out the weight/delay adaption using this time window. The aVLSI time window generator is compact (50 μm2in IBM 130nm process) and we use a time multiplexing approach to achieve up to 65536 (64k) virtual digital adaptors with one physical adaptor, consuming only a fraction of the hardware resource on a Virtex 6 FPGA. Since the digital adaptor has been implemented on an FPGA, it can be easily reconfigured for different adaptation algorithms, which leaves it open for future development. Our mixed-signal implementation is therefore practical for implementing the synaptic plasticity in large-scale spiking neural networks running in real time. We show circuit simulation results illustrating both weight and delay adaptation.

We present an analogue Very Large Scale Integration (aVLSI) implementation that uses first-order log-domain low-pass filters to implement a generalised conductance-based silicon neuron. It consists of a single synapse, which is capable of linearly summing both the excitatory and inhibitory post-synaptic currents (EPSC and IPSC) generated by the spikes arriving from different sources, a soma with a positive feedback circuit, a refractory period and spike-frequency adaptation circuit, and a high-speed synchronous Address Event Representation (AER) handshaking circuit. To increase programmability, the inputs to the neuron are digital spikes, the durations of which are modulated according to their weights. The proposed neuron is a compact design (∼170 µm2in the IBM 130nm process). Our aVLSI generalised conductance-based neuron is therefore practical for large-scale reconfigurable spiking neural networks running in real time. Circuit simulations show that this neuron can emulate different spiking behaviours observed in biological neurons.

Bayesian spiking neurons (BSNs) provide a probabilistic interpretation of how neurons perform inference and learning. Online learning in BSNs typically involves parameter estimation based on maximum-likelihood expectation-maximization (ML-EM) which is computationally slow and limits the potential of studying networks of BSNs. An online learning algorithm, fast learning (FL), is presented that is more computationally efficient than the benchmark ML-EM for a fixed number of time steps as the number of inputs to a BSN increases (e.g., 16.5 times faster run times for 20 inputs). Although ML-EM appears to converge 2.0 to 3.6 times faster than FL, the computational cost of ML-EM means that ML-EM takes longer to simulate to convergence than FL. FL also provides reasonable convergence performance that is robust to initialization of parameter estimates that are far from the true parameter values. However, parameter estimation depends on the range of true parameter values. Nevertheless, for a physiologically meaningful range of parameter values, FL gives very good average estimation accuracy, despite its approximate nature. The FL algorithm therefore provides an efficient tool, complementary to ML-EM, for exploring BSN networks in more detail in order to better understand their biological relevance. Moreover, the simplicity of the FL algorithm means it can be easily implemented in neuromorphic VLSI such that one can take advantage of the energy-efficient spike coding of BSNs.

As advances in integrated circuit (IC) fabrication technology reduce feature sizes to dimensions on the order of nanometers, IC designers are facing many of the problems that evolution has had to overcome in order to perform meaningful and accurate computations in biological neural circuits. In this paper, we explore the current state of IC technology including the many new and exciting opportunities “beyond CMOS.” We review the role of noise in both biological and engineered systems and discuss how “stochastic facilitation” can be used to perform useful and precise computation. We explore nondeterministic methodologies for computation in hardware and introduce the concept of stochastic electronics (SE); a new way to design circuits and increase performance in highly noisy and mismatched fabrication environments. This approach is illustrated with several circuit examples whose results demonstrate its exciting potential.

This paper introduces the process for creating the Sydney York Morphological and Acoustic Recordings of Ears (SYMARE) database. The SYMARE database supports research exploring the relationship between the morphology of human outer ears and their acoustic filtering properties-a relationship that is viewed by many as holding the key to human spatial hearing and the future of 3D personal audio. The SYMARE database is comprised of acoustically measured head-related impulse responses for 61 listeners (48 male/13 female), multiple high-resolution surface mesh models (upper torso, head and ears) for these listeners obtained from magnetic resonance imaging (MRI) data, and the corresponding simulated HRIR data for these listeners generated using the Fast Multipole Boundary Element Method (FM-BEM). In this work, we compare acoustically measured HRIR data for 61 listeners with the listeners' corresponding simulated HRIR data generated using the FM-BEM.

We present a bio-potential front-end amplifier capable of recording unipolar ECG signals without making use of the Wilson Central Terminal. The circuit is compatible with both standard and dry electrodes, and is low power (requiring less than 25 mW powered at 12 V). It is therefore well suited for long-term applications The information contained in the new unipolar recordings may yield unique diagnostic capabilities as it avoids the need to measure differential signals or make use of the averaging effect imposed by the Wilson Central Terminal. Our tests on a small population (5 subjects) showed that the system also allows direct, real-time software calculation of signals corresponding to standard ECG leads, which have a correlation in excess of 92% with standard signals recorded in parallel. Comparisons between these recordings reveal that one of the most evident alterations super-imposed by the Wilson Central Terminal is a shift in the time domain of important features of the ECG signals such as the R-peak.

We present a standard pipelined ADC design using Nauta structure differential op-amps as an alternative to traditional analog op-amps. The six stage pipelined ADC is capable of running at 100MS/s and achieves 8 bit resolution under simulations. The research is focused on the path to scaling to deep sub-micron CMOS and finding alternatives to the reduced gain and low output voltage swing of traditional analog op-amp designs. The Nauta structure op-amp allows us to produce one of the smallest reported areas for a 180nm pipelined ADC occupying only 0.3mm2for a 10 bit 100MS/s pipelined ADC.

We present a voltage domain implementation of a programmable delay axon circuit together with measurements from it. It was designed to be a building block for a polychronous spiking neural network. The axonal delay can be programmed by presenting an input spike followed by a post-synaptic spike at the desired delay. An analogue memory was used to store this value. We also use spike timing dependent delay plasticity (STDDP) to reduce the errors in delay that result from the delay programming step. Measurements show that the proposed circuit is capable of learning and retaining delays in the range of 2 ms to 50 ms for many minutes.

This letter discusses temporal order coding and detection in nervous systems. Detection of temporal order in the external world is an adaptive function of nervous systems. In addition, coding based on the temporal order of signals can be used as an internal code. Such temporal order coding is a subset of temporal coding. We discuss two examples of processing the temporal order of external events: the auditory location detection system in birds and the visual direction detection system in flies. We then discuss how somatosensory stimulus intensities are translated into a temporal order code in the human peripheral nervous system. We next turn our attention to input order coding in the mammalian cortex. We review work demonstrating the capabilities of cortical neurons for detecting input order. We then discuss research refuting and demonstrating the representation of stimulus features in the cortex by means of input order. After some general theoretical considerations on input order detection and coding, we conclude by discussing the existing and potential use of input order coding in neuromorphic engineering.

This paper introduces the process for creating the Sydney York Morphological and Acoustic Recordings of Ears (SYMARE) database. The SYMARE database supports research exploring the relationship between the morphology of human outer ears and their acoustic filtering properties - a relationship that is viewed by many as holding the key to human spatial hearing and the future of 3D personal audio. The SYMARE database is comprised of acoustically measured head-related impulse responses for 60 listeners, multiple high-resolution surface mesh models (upper torso, head and ears) for these listeners obtained from magnetic resonance imaging (MRI) data, and the corresponding simulated HRIR data for these listeners generated using the Fast Multipole Boundary Element Method (FM-BEM). In this work, we compare acoustically measured HRIR data for ten listeners with the listeners' corresponding simulated HRIR data generated using the FM-BEM.

Recent work in neuroscience is revealing how the blowfly rapidly detects orientation using neural circuits distributed directly behind its photo receptors. These circuits like all biological systems rely on timing, competition, feedback, and energy optimization. The recent realization of the passive memristor device, the so-called fourth fundamental passive element of circuit theory, assists with making low power biologically inspired parallel analog computation achievable. Building on these developments, we present a memristor-based neuromorphic competitive control (mNCC) circuit, which utilizes a single sensor and can control the output of N actuators delivering optimal scalable performance, and immunity from device variation and environmental noise.

Bayesian Spiking Neurons (BSNs) provide a probabilistic interpretation of how neurons can perform inference and learning. Learning in a single BSN can be formulated as an online maximum-likelihood expectation-maximisation (ML-EM) algorithm. This form of learning is quite slow. Here, an alternative to this learning algorithm, called Fast Learning (FL), is presented. The FL algorithm is shown to have acceptable convergence performance when compared to the ML-EM algorithm. Moreover, for our implementations the FL algorithm is approximately 25 times faster than the ML-EM algorithm. Although only approximate, the FL algorithm therefore makes learning in hierarchical BSN networks much more tractable.

We present a digital phase interpolator (PI) design for 65nm CMOS that avoids conventional analog structures, accurately achieves 2-bits phase resolution across a range of rise time and input delays from trise: 48ps → 200ps using a ratio trise/tdelayof at least 1 or greater. Increased accuracy is available for certain rise times using ratios increasing between 1 and 10 as verified by simulations across process corners using extracted parasitic capacitances but ignoring MOSFET mismatch effects. Power consumption was estimated at 30nW/MHz → 38nW/MHz across a range of process variation corners in these operating conditions. Monte Carlo simulations across process and MOSFET mismatch conditions show large variations in estimated accuracy. Monte Carlo trials show the PI achieves a worst case DNL error (mean±3σ) of 1.06 LSB using trise/tdelayratio of 5.3 and 48ps rise time, and a worst case DNL error (mean ±2σ) of 0.49 LSB for trise/tdelayratio of 4 and 84ps rise time.

A new parallel ADC architecture is presented which makes use of neuromorphic principles to be fast, accurate, and robust to noise and circuit mismatch. The architecture uses spiking integrate-and-fire neurons as base elements, with lateral inhibition to decohere the parallel pathways, and alternate on-and off-triggered paths to maintain a constant spike rate. Results from a proof-of-concept circuit reinforce the analytical conclusion that this circuit can make a practical ADC.

We present measurements from an aVLSI programmable axonal propagation delay circuit. It is intended to be used in the implementation of polychronous spiking neural networks. The delay can be programmed by presenting an input spike followed by a training spike at the desired delay. To fine tune and maintain the delay using an analogue memory, we use continuous spike timing dependent delay adaptation. Measurements presented here show that the axon circuit is capable of learning and retaining delays in the 2.5-20 ms range, as long as the neuron is stimulated at least once every few seconds.

This paper describes the design of an auto-associative memory based on a spiking neural network (SNN). The architecture is able to effectively utilize the massive interconnect resources available in FPGA architectures as a good match to the axons in biological neural networks. A complete implementation of the memory on a single FPGA is presented. The signal processing circuitry is composed from simple, parallel building blocks and the training logic is implemented using an on-chip soft processor.

A novel technique for time domain spatial sound reproduction using compressed sensing is presented. The presented technique is based on the application of compressed sensing theory, which is used to improve the accuracy of the reconstructed sound field. In addition, singular value decomposition is also applied, which acts to significantly reduce the size of the data set to process, thus making it efficient and realisable for real-time applications. Results are presented from the preliminary performance evaluation of the compressed sensing technique in comparison to the Higher Order Ambisonic reconstruction technique.

We present an implementation of a programmable axonal propagation delay circuit which uses one first-order log-domain low-pass filter. Delays may be programmed in the 5-50ms range. It is designed to be a building block for time-delay spiking neural networks. It consists of a leaky-integrate-and-fire core, a spike generator circuit, and a delay adaptation circuit.

In this paper we present results from several implementations of silicon cochleae whose dynamics are governed by the Hopf equation. These silicon cochleae exhibit the majority of active, nonlinear characteristics of the biological cochlea such as large-signal compression, two-tone suppression, the creation of distortion products and so forth. Here we explore the coupling between resonant sections of the basilar membrane to investigate phenomena such as masking and the characteristic frequency response curve of the cochlea at a particular place along the basilar membrane. We see that the interaction of resonant sections can account for these phenomena and that we can use these observations to partially explain the connectivity of the afferent and efferent fibres to the outer hair cells. This work not only gives us valuable insight into the dynamical behaviour of the early auditory system but it also highlights the benefits of building circuits of these complex systems in order to produce models whose parameters can be tuned and whose outputs can be observed and measured in realtime.

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.

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.

We present an electronic neuron that uses first-order log-domain low-pass filters to implement the Mihalas-Niebur model. The neuron consists of a leaky-integrate-and-fire core and building blocks to implement an adaptive threshold and spike induced currents. Simulation results show that this modular neuron can emulate different spiking behaviours observed in biological neurons.

We have designed, constructed and tested an electromechanical acoustic sensor as a conceptual model for the active process in the mammalian cochlea. The sensor is based on a mechanical resonator - a stretched latex band - which is tonically (tensionally) modulated by an electromechanical actuator. A feedback circuit senses the motion of the resonator, and modulates its tension at twice the frequency of its motion. An amplifier is thereby formed, which is self-tuned to the circuit's resonance, and which produces gain by regeneration and degenerate parametric pumping. The system is simple in structure; physiologically plausible as a model for the basilar membrane and associated hair cells; and reproduces several well-known features of the cochlear amplifier.

We describe an electromechanical acoustic sensor which we have constructed and tested as a conceptual model for the active process in the mammalian cochlea. The sensor is based on a mechanical resonator - a stretched latex band - which is tonically (tensionally) modulated by an electromechanical actuator. A feedback circuit senses the motion of the resonator, and modulates its tension at twice the frequency of its motion. An amplifier is thereby formed, which is self-tuned to the circuit's resonance, and which produces gain by regeneration and degenerate parametric pumping. We show that this system is surprisingly simple in structure; is physiologically plausible as a model for the basilar membrane and associated structures; and that it reproduces several well-known features of the cochlear amplifier, such as the variation in center frequency and bandwidth with gain, and a cubic transfer characteristic in open loop.

We present a new method and performance data related to volumetric acoustic intensity imaging using a spherical microphone array (SMA) consisting of a dual, concentric rigid and open SMA. The dual, concentric array was designed to improve the frequency range of a standard SMA. We apply standard techniques associated with interior spherical near-field acoustic holography (NAH) and, in particular, consider issues related to the optimal use of information from both arrays for NAH projection and the advantages that thus accrue from utilising a dual, concentric SMA.

A neuromorphic sound localisation system is proposed. It employs two microphones and a pair of silicon cochleae with address event interface for front-end processing. This allows subsequent processing to be implemented with spike-based algorithms. The system is adaptive and supports online learning. Its localisation capability was tested with white noise and pure tone stimuli, with an average error of around 3° in the −45° to 45° range.

We present a first-order nonhomogeneous Markov model for the interspike-interval density of a continuously stimulated spiking neuron. The model allows the conditional interspike-interval density and the stationary interspike-interval density to be expressed as products of two separate functions, one of which describes only the neuron characteristics and the other of which describes only the signal characteristics. The approximation shows particularly clearly that signal autocorrelations and cross-correlations arise as natural features of the interspike-interval density and are particularly clear for small signals and moderate noise. We show that this model simplifies the design of spiking neuron cross-correlation systems and describe a four-neuron mutual inhibition network that generates a cross-correlation output for two input signals.

In this paper we present a 2-D silicon cochlea which includes an automatic quality factor control (AQC) loop. This control-loop is an improved version of that presented in [1] where the control-loop imposes both a ceiling and a threshold level on the output amplitude of the basilar membrane (BM) resonators. In this improved version we include only a single set-point in our control-loop. This allows us to tune the BM resonators close to a Hopf bifurcation. We present test results from a fabricated integrated circuit which, when compared with biological data, demonstrates the feasibility of our active 2-D cochlea model.

Recent work in cochlear amplifier modeling has focused on systems which show the dynamics of a Hopf bifurcation. We show that these systems are examples of a generic amplifier topology, the self-tuned regenerative amplifier (STRA). The STRA is a feedback-stabilized regenerative amplifier that can be operated in a region of supercritical stability. The signatures of Hopf amplification, such as a cubic nonlinear small-signal response at resonance, are general features of the topology. The topology is shown to include a degenerate parametric amplifier, which may explain its low noise and insensitivity to input-feedback phase mismatch.

We describe a circuit of two spiking neurons which extracts mathematically accurate cross-correlations from the signal inputs. It differs from prior circuits such as coincidence detectors or enhanced motion detectors in that it does not require anapriorifixed delay between input signals to be selected. The output, in the form of a differential spike histogram, displays a mathematical cross-correlation in the conventional correlation vs. time form.

Address-event-representation (AER) is a communication protocol for transferring asynchronous events between VLSI chips, originally developed for bio-inspired processing systems (for example, image processing). The event information in an AER system is transferred using a high-speed digital parallel bus. This paper presents an experiment using AER for sensing, processing and finally actuating a robot. The AER output of a silicon cochlea is processed by an AER filter implemented on a FPGA to produce rhythmic walking in a humanoid robot (Redbot). We have implemented both the AER rhythm detector and the central pattern generator (CPG) on a Spartan II FPGA which is part of a USB-AER platform developed by some of the authors

In this paper we present a basilar membrane resonator design for an active 2D cochlea. It incorporates some of the non-linear behaviour exhibited in the real cochlea by utilizing a quality factor control loop. This control loop varies the gain and the frequency selectivity of the resonator based on the amplitude of the input signal.

We present a method for simulating a near-field sound source in virtual auditory space (VAS). The method scales individualised HRTFs, measured at a distance of 1m to arbitrary distances in the near-field. It uses a model of the acoustic scattering for a point-source on a rigid sphere to calculate a distance variation function (DVF) to apply to the HRTFs. A sound localisation experiment was conducted in VAS with three subjects to evaluate the acoustic spatial fidelity of this method. Results show that with the modified HRTFs directional localisation is generally maintained at different distances and there is reasonable correlation between the perceived distance and target distance for distances up to 50cm from the centre of the subject's head.

We present a psychoacoustic experiment that explores the ability of various listeners to discriminate between the virtual auditory space (VAS) stimuli generated using different binaural impulse response functions recorded in the Sydney Opera House. The binaural head-related impulse response (HRIR) functions were recorded for a group of subjects sitting in the same seat, P34, using a log sine sweep sound source located at the centre of the stage. The VAS stimuli generated using these HRIRs consist mostly of a variety of musical excerpts, speech, and white noise. Experimental results using an ABX test procedure show that out of a total of 1350 trials, 10 subjects responded correctly in 1230 of the test trials, indicating a discrimination performance greater than 90%. We also present data indicating the types of perceptual cues that aid in binaural sound discrimination process.

We present measured frequency-gain functions and the spike rate outputs of the different sections in a spiking silicon cochlea chip. The chip consists of a matched pair of silicon cochleae with an address event interface for the output. Each section of the cochlea is modelled by a second-order low-pass filter followed by a simplified inner hair cell circuit and a spiking neuron circuit. When the neuron spikes, an address event is generated on the asynchronous data bus. These spike outputs are analogous to the spikes on an auditory nerve, connecting the cochlea with the brain.

Using various layout techniques and circuit configurations, the effects of matching on a log-domain filter were analyzed. It is shown here that one of three possible Tau cell configurations to implement the same 2nd order low pass filter clearly outperforms the others. Furthermore, application of a common centroid layout technique has had no noticeable improvement on filter matching

Female crickets can locate males by phonotaxis to the mating song they produce. The behaviour and underlying physiology has been studied in some depth showing that the cricket auditory system solves this complex problem in a unique manner. We present an analogue very large scale integrated (aVLSI) circuit model of this process and show that results from testing the circuit agree with simulation and what is known from the behaviour and physiology of the cricket auditory system. The aVLSI circuitry is now being extended to use on a robot along with previously modelled neural circuitry to better understand the complete sensorimotor pathway. 1 In t r o d u c t i o n Understanding how insects carry out complex sensorimotor tasks can help in the design of simple sensory and robotic systems. Often insect sensors have evolved into intricate filters matched to extract highly specific data from the environment which solves a particular problem directly with little or no need for further processing [1]. Examples include head stabilisation in the fly, which uses vision amongst other senses to estimate self-rotation and thus to stabilise its head in flight, and phonotaxis in the cricket. Because of the narrowness of the cricket body (only a few millimetres), the Interaural Time Difference (ITD) for sounds arriving at the two sides of the head is very small (1020s). Even with the tympanal membranes (eardrums) located, as they are, on the forelegs of the cricket, the ITD only reaches about 40s, which is too low to detect directly from timings of neural spikes. Because the wavelength of the cricket calling song is significantly greater than the width of the cricket body the Interaural Intensity Difference (IID) is also very low. In the absence of ITD or IID information, the cricket uses phase to determine direction. This is possible because the male cricket produces an almost pure tone for its calling song. * + School of Electrical and Information Engineering, Institute of Perception, Action and Behaviour. Figure 1: The cricket auditory system. Four acoustic inputs channel sounds directly or through tracheal tubes onto two tympanal membranes. Sound from contralateral inputs has to pass a (double) central membrane (the medial septum), inducing a phase delay and reduction in gain. The sound transmission from the contralateral tympanum is very weak, making each eardrum effectively a 3 input system. The physics of the cricket auditory system is well understood [2]; the system (see Figure 1) uses a pair of sound receivers with four acoustic inputs, two on the forelegs, which are the external surfaces of the tympana, and two on the body, the prothoracic or acoustic spiracles [3]. The connecting tracheal tubes are such that interference occurs as sounds travel inside the cricket, producing a directional response at the tympana to frequencies near to that of the calling song. The amplitude of vibration of the tympana, and hence the firing rate of the auditory afferent neurons attached to them, vary as a sound source is moved around the cricket and the sounds from the different inputs move in and out of phase. The outputs of the two tympana match when the sound is straight ahead, and the inputs are bilaterally symmetric with respect to the sound source. However, when sound at the calling song frequency is off-centre the phase of signals on the closer side comes better into alignment, and the signal increases on that side, and conversely decreases on the other. It is that crossover of tympanal vibration amplitudes which allows the cricket to track a sound source (see Figure 6 for example). A simplified version of the auditory system using only two acoustic inputs was implemented in hardware [4], and a simple 8-neuron network was all that was required to then direct a robot to carry out phonotaxis towards a species-specific calling song [5]. A simple simulator was also created to model the behaviour of the auditory system of Figure 1 at different frequencies [6]. Data from Michelsen et al. [2] (Figures 5 and 6) were digitised, and used together with average and "typical" values from the paper to choose gains and delays for the simulation. Figure 2 shows the model of the internal auditory system of the cricket from sound arriving at the acoustic inputs through to transmission down auditory receptor fibres. The simulator implements this model up to the summing of the delayed inputs, as well as modelling the external sound transmission. Results from the simulator were used to check the directionality of the system at different frequencies, and to gain a better understanding of its response. It was impractical to check the effect of leg movements or of complex sounds in the simulator due to the necessity of simulating the sound production and transmission. An aVLSI chip was designed to implement the same model, both allowing more complex experiments, such as leg movements to be run, and experiments to be run in the real world. Figure 2: A model of the auditory system of the cricket, used to build the simulator and the aVLSI implementation (shown in boxes). These experiments with the simulator and the circuits are being published in [6] and the reader is referred to those papers for more details. In the present paper we present the details of the circuits used for the aVLSI implementation.

In this work we study the influence and relationship of five different acoustical cues to the human sound localisation process. These cues are: interaural time delay, interaural level difference, interaural spectrum, monaural spectrum, and band-edge spectral contrast. Of particular interest was the synthesis and integration of the different cues to produce a coherent and robust percept of spatial location. The relative weighting and role of the different cues was investigated using band-pass filtered white noise with a frequency range (in kHz) of: 0.3-5, 0.3-7, 0.3-10, 0.3-14, 3-8, 4-9, and 7-14. These stimuli provided varying amounts of spectral information and physiologically detectable temporal information, thus probing the localisation process under varying sound conditions. Three subjects with normal hearing in both ears have performed five trials of 76 test positions for each of these stimuli in an anechoic room. All subjects showed systematic mislocalisation on most of these stimuli. The location to which they are mislocalised varies among subjects but in a systematic manner related to the five different acoustical cues. These cues have been correlated with the subject's localisation responses on an individual basis with the results suggesting that the internal weighting of the spectral cues may vary with the sound condition.

An analogue electronic model of a cochlea was developed some years ago by R. F. Lyon using a cascade of filters. Combined with an analogue gradient descent circuit, such an artificial cochlea can be used to extract the Linear Predictive Coding (LPC) of the speech signal in continuous time. We propose in this article an analogue VLSI circuit implementing a so-called 'Cochlear LPC' (CLPC). We discuss the expected advantages of the CLPC, such as an optimal time-frequency resolution and the continuous time processing. Furthermore we present some speech recognition results obtained using a computer model of the CLPC pre-processing combined with an HMM classifier. These results compare favourably with those obtained from a standard LPC/HMM system.