Modern computation based on von Neumann architecture is now a mature cutting-edge science. In the von Neumann architecture, processing and memory units are implemented as separate blocks interchanging data intensively and continuously. This data transfer is responsible for a large part of the power consumption. The next generation computer technology is expected to solve problems at the exascale with 1018 calculations each second. Even though these future computers will be incredibly powerful, if they are based on von Neumann type architectures, they will consume between 20 and 30 megawatts of power and will not have intrinsic physically built-in capabilities to learn or deal with complex data as our brain does. These needs can be addressed by neuromorphic computing systems which are inspired by the biological concepts of the human brain. This new generation of computers has the potential to be used for the storage and processing of large amounts of digital information with much lower power consumption than conventional processors. Among their potential future applications, an important niche is moving the control from data centers to edge devices. The aim of this roadmap is to present a snapshot of the present state of neuromorphic technology and provide an opinion on the challenges and opportunities that the future holds in the major areas of neuromorphic technology, namely materials, devices, neuromorphic circuits, neuromorphic algorithms, applications, and ethics. The roadmap is a collection of perspectives where leading researchers in the neuromorphic community provide their own view about the current state and the future challenges for each research area. We hope that this roadmap will be a useful resource by providing a concise yet comprehensive introduction to readers outside this field, for those who are just entering the field, as well as providing future perspectives for those who are well established in the neuromorphic computing community.

Dennis V. Christensen, Regina Dittmann, Bernabe Linares-Barranco, Abu Sebastian, Manuel Le Gallo, Andrea Redaelli, Stefan Slesazeck, Thomas Mikolajick, Sabina Spiga, Stephan Menzel, Ilia Valov, Gianluca Milano, Carlo Ricciardi, Shi-Jun Liang, Feng Miao, Mario Lanza, Tyler J. Quill, Scott T. Keene, Alberto Salleo, Julie Grollier, Danijela Markovic, Alice Mizrahi, Peng Yao, J. Joshua Yang, Giacomo Indiveri, John Paul Strachan, Suman Datta, Elisa Vianello, Alexandre Valentian, Johannes Feldmann, Xuan Li, Wolfram H.P. Pernice, Harish Bhaskaran, Emre Neftci, Srikanth Ramaswamy, Jonathan Tapson, Franz Scherr, Wolfgang Maass, Priyadarshini Panda, Youngeun Kim, Gouhei Tanaka, Simon Thorpe, Chiara Bartolozzi, Thomas A. Cleland, Christoph Posch, Shih-Chii Liu, Arnab Neelim Mazumder, Morteza Hosseini, Tinoosh Mohsenin, Elisa Donati, Silvia Tolu, Roberto Galeazzi, Martin Ejsing Christensen, Sune Holm, Daniele Ielmini, N. Pryds

We present a novel computing architecture which combines the event-based and compute-in-network principles of neuromorphic computing with a traditional dataflow architecture. The result is a fine-grained dynamic dataflow system which avoids the coding issues intrinsic to spiking systems, and is suitable for both procedural workload and deep neural network (DNN) inference. The architecture is particularly suitable for computation of sparse CNNs and low-latency applications. We present results from GrAIOne, the first chip designed using the NeuronFlow architecture, which has 200 704 neurons implemented in a 28nm HPC + process.

Neuronflow is a neuromorphic, many core, data flow architecture that exploits brain-inspired concepts to deliver a scalable event-based processing engine for neuron networks in Live AI applications. Its design is inspired by brain biology, but not necessarily biologically plausible. The main design goal is the exploitation of sparsity to dramatically reduce latency and power consumption as required by sensor processing at the Edge.

Biological neurons are known to have sparse and asynchronous communications using spikes. Despite our incomplete understanding of processing strategies of the brain, its low energy consumption in fulfilling delicate tasks suggests the existence of energy efficient mechanisms. Inspired by these key factors, we introduce SpArNet, a bio-inspired quantization scheme to convert a pre-trained convolutional neural network to a spiking neural network, with the aim of minimizing the computational load for execution on neuromorphic processors. The proposed scheme has significant advantages over the reference CNN in a reduced number of synaptic operations, and can be used for frequent executions of inference tasks. The computational load of SpArNet is adjusted to the spatio-temporal dynamics of the the input data. We have tested the converted network on two applications (autonomous steering and hand gesture recognition), demonstrating a significant reduction on the number of required synaptic operations.

(Neuron 101, 365–369; February 6, 2019) In the original publication of this NeuroView, the member list for the Australian Brain Alliance was omitted. This has now been corrected online. Neuron apologizes for the error.

With this paper we communicated the existence of a surface electrocardiography (ECG) recordings dataset, named WCTECGdb, that aside from the standard 12-lead signals includes the raw electrode biopotential for each of the nine exploring electrodes refereed directly to the right leg. This dataset, comprises of 540 ten second segments recorded from 92 patients at Campbelltown Hospital, NSW Australia, and is now available for download from the Physionet platform. The data included in the dataset confirm that the Wilson's Central Terminal (WCT) has a relatively large amplitude (up to 247% of lead II) with standard ECG characteristics such as a p-wave and a t-wave, and is highly variable during the cardiac cycle. As further examples of application for our data, we assess: (1) the presence of a conductive pathway between the legs and the heart concluding that in some cases is electrically significant and (2) the initial assumption about the limbs potential stating the dominance of the left arm concluding that this is not always the case and that might requires case to case assessment.

Unsupervised feature extraction algorithms form one of the most important building blocks in machine learning systems. These algorithms are often adapted to the event-based domain to perform online learning in neuromorphic hardware. However, not designed for the purpose, such algorithms typically require significant simplification during implementation to meet hardware constraints, creating trade offs with performance. Furthermore, conventional feature extraction algorithms are not designed to generate useful intermediary signals which are valuable only in the context of neuromorphic hardware limitations. In this work a novel event-based feature extraction method is proposed that focuses on these issues. The algorithm operates via simple adaptive selection thresholds which allow a simpler implementation of network homeostasis than previous works by trading off a small amount of information loss in the form of missed events that fall outside the selection thresholds. The behavior of the selection thresholds and the output of the network as a whole are shown to provide uniquely useful signals indicating network weight convergence without the need to access network weights. A novel heuristic method for network size selection is proposed which makes use of noise events and their feature representations. The use of selection thresholds is shown to produce network activation patterns that predict classification accuracy allowing rapid evaluation and optimization of system parameters without the need to run back-end classifiers. The feature extraction method is tested on both the N-MNIST (Neuromorphic-MNIST) benchmarking dataset and a dataset of airplanes passing through the field of view. Multiple configurations with different classifiers are tested with the results quantifying the resultant performance gains at each processing stage.

In this work, we investigate event-based feature extraction through a rigorous framework of testing. We test a hardware efficient variant of Spike Timing Dependent Plasticity (STDP) on a range of spatio-temporal kernels with different surface decaying methods, decay functions, receptive field sizes, feature numbers, and back end classifiers. This detailed investigation can provide helpful insights and rules of thumb for performance vs. complexity trade-offs in more generalized networks, especially in the context of hardware implementation, where design choices can incur significant resource costs. The investigation is performed using a new dataset consisting of model airplanes being dropped free-hand close to the sensor. The target objects exhibit a wide range of relative orientations and velocities. This range of target velocities, analyzed in multiple configurations, allows a rigorous comparison of time-based decaying surfaces (time surfaces) vs. event index-based decaying surface (index surfaces), which are used to perform unsupervised feature extraction, followed by target detection and recognition. We examine each processing stage by comparison to the use of raw events, as well as a range of alternative layer structures, and the use of random features. By comparing results from a linear classifier and an ELM classifier, we evaluate how each element of the system affects accuracy. To generate time and index surfaces, the most commonly used kernels, namely event binning kernels, linearly, and exponentially decaying kernels, are investigated. Index surfaces were found to outperform time surfaces in recognition when invariance to target velocity was made a requirement. In the investigation of network structure, larger networks of neurons with large receptive field sizes were found to perform best. We find that a small number of event-based feature extractors can project the complex spatio-temporal event patterns of the dataset to an almost linearly separable representation in feature space, with best performing linear classifier achieving 98.75% recognition accuracy, using only 25 feature extracting neurons.

In this paper we introduce a novel Salience Affected Artificial Neural Network (SANN) that models the way neuromodulators such as dopamine and noradrenaline affect neural dynamics in the human brain by being distributed diffusely through neocortical regions, allowing both salience signals to modulate cognition immediately, and one time learning to take place through strengthening entire patterns of activation at one go. We present a model that is capable of one-time salience tagging in a neural network trained to classify objects, and returns a salience response during classification (inference). We explore the effects of salience on learning via its effect on the activation functions of each node, as well as on the strength of weights between nodes in the network. We demonstrate that salience tagging can improve classification confidence for both the individual image as well as the class of images it belongs to. We also show that the computation impact of producing a salience response is minimal. This research serves as a proof of concept, and could be the first step towards introducing salience tagging into Deep Learning Networks and robotics.

Inference of Deep Neural Networks for stream signal (Video/Audio) processing in edge devices is still challenging. Unlike the most state of the art inference engines which are efficient for static signals, our brain is optimized for real-time dynamic signal processing. We believe one important feature of the brain (asynchronous state-full processing) is the key to its excellence in this domain. In this work, we show how asynchronous processing with state-full neurons allows exploitation of the existing sparsity in natural signals. This paper explains three different types of sparsity and proposes an inference algorithm which exploits all types of sparsities in the execution of already trained networks. Our experiments in three different applications (Handwritten digit recognition, Autonomous Steering and Hand-Gesture recognition) show that this model of inference reduces the number of required operations for sparse input data by a factor of one to two orders of magnitudes. Additionally, due to fully asynchronous processing this type of inference can be run on fully distributed and scalable neuromorphic hardware platforms.

In this paper we introduce a novel Salience Affected Artificial Neural Network (SANN) that models the way neuromodulators such as dopamine and noradrenaline affect neural dynamics in the human brain by being distributed diffusely through neocortical regions, allowing both salience signals to modulate cognition immediately, and one time learning to take place through strengthening entire patterns of activation at one go. We present a model that is capable of one-time salience tagging in a neural network trained to classify objects, and returns a salience response during classification (inference). We explore the effects of salience on learning via its effect on the activation functions of each node, as well as on the strength of weights between nodes in the network. We demonstrate that salience tagging can improve classification confidence for both the individual image as well as the class of images it belongs to. We also show that the computation impact of producing a salience response is minimal. This research serves as a proof of concept, and could be the first step towards introducing salience tagging into Deep Learning Networks and robotics.

The surface electrocardiography (ECG) uses a virtual reference point to measure the potential of chest electrodes. This reference potential is known as Wilson central terminal (WCT) and is assumed negligible (near zero) in amplitude. Consequently, the precordial leads have been named as the unipolar leads. Although this assumption was found incorrect immediate after this reference potential was introduced, it was difficult to measure its real amplitude. We recently introduced a 15-lead electrocardiography device that can record the traditional ECG leads in combination with the raw potential of limbs and chest electrodes directly referred to the circuit grounding. Consequently, we are able to record the potential of the raw chest electrodes, which we named as true unipolar chest leads. The aim of this study is to have a clear understanding of the WCT potential and its influence on the chest leads. Our records show that the true unipolar leads may be more sensitive for detecting cardiac diseases in the left anterior descending coronary artery in patients with non-ST elevation reported on chest leads.

Since its inception, electrocardiography has been based on the simplifying hypothesis that cardinal limb leads form an equilateral triangle of which, at the center/centroid, the electrical equivalent of the cardiac activity rotates during the cardiac cycle. Therefore, it is thought that the three limbs (right arm, left arm, and left leg) which enclose the heart into a circuit, where each branch directly implies current circulation through the heart, can be averaged together to form a stationary reference (central terminal) for precordials/chest-leads. Our hypothesis is that cardinal limbs do not form a triangle for the majority of the duration of the cardiac cycle. As a corollary, the central point may not lie in the plane identified by the limb leads. Using a simple and efficient algorithm, we demonstrate that the portion of the cardiac cycle where the three limb leads form a triangle is, on average less, than 50%.

As the integrated circuit (IC) technology advances into smaller nanometre feature sizes, a fixed-error noise known as device mismatch is introduced owing to the dissimilarity between transistors, and this degrades the accuracy of analog circuits. In this paper, we present an analog co-processor that uses this fixed-pattern noise to its advantage to perform complex computation. This circuit is an extension of our previously published trainable analogue block (TAB) framework and uses multiple inputs that substantially increase functionality. We present measurement results of our two-input analogue co-processor built using a 130-nm process technology and show its learning capabilities for regression and classification tasks. We also show that the co-processor, comprised of 100 neurons, is a low-power system with a power dissipation of only 1.1 mu W. The IC fabrication process contributes to randomness and variability in ICs, and we show that random device mismatch is favorable for the learning capability of our system as it causes variability among the neuronal tuning curves. The low-power capability of our framework makes it suitable for use in various battery-powered applications ranging from biomedical to military as a front-end analog co-processor.

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.

Objective The Wilson Central Terminal (WCT) is an artificially constructed reference for surface electrocardiography, which is assumed to be near zero and steady during the cardiac cycle; namely it is the simple average of the three recorded limbs (right arm, left arm and left leg) composing the Einthoven triangle and considered to be electrically equidistant from the electrical center of the heart. This assumption has been challenged and disproved in 1954 with an experiment designed just to measure and minimize WCT. Minimization was attempted varying in real time the weight resistors connected to the limbs. Unfortunately, the experiment required a very cumbersome setup and showed that WCT amplitude could not be universally minimized, in other words, the weight resistors change for each person. Taking advantage of modern computation techniques as well as of a special ECG device that aside of the standard 12-lead Electrocardiogram (ECG) can measure WCT components, we propose a software minimization (genetic algorithm) method using data recorded from 72 volunteers. Result We show that while the WCT presents average amplitude relative to lead II of 58.85% (standard deviation of 30.84%), our minimization method yields an amplitude as small as 7.45% of lead II (standard deviation of 9.04%).

We present a hardware architecture that uses the neural engineering framework (NEF) to implement large-scale neural networks on field programmable gate arrays (FPGAs) for performing massively parallel real-time pattern recognition. NEF is a framework that is capable of synthesising large-scale cognitive systems from subnetworks and we have previously presented an FPGA implementation of the NEF that successfully performs nonlinear mathematical computations. That work was developed based on a compact digital neural core, which consists of 64 neurons that are instantiated by a single physical neuron using a time-multiplexing approach. We have now scaled this approach up to build a pattern recognition system by combining identical neural cores together. As a proof of concept, we have developed a handwritten digit recognition system using the MNIST database and achieved a recognition rate of 96.55%. The system is implemented on a state-of-the-art FPGA and can process 5.12 million digits per second. The architecture and hardware optimisations presented offer high-speed and resource-efficient means for performing high-speed, neuromorphic, and massively parallel pattern recognition and classification tasks.

Modern electrocardiography (ECG)1 uses the Wilson Central Terminal (WCT) as a reference point for the majority of leads. WCT is assumed to be near zero and steady during the cardiac cycle. However, due to the measurement encumbrances of the real amplitude of WCT, this assumption has never been verified in clinical practice. Using our own recently developed 15-lead ECG device that can measure WCT components in addition to 12-lead ECGs in a clinical setting, we propose a framework to derive a New Central Terminal (NCT) with demonstrated less variation and near zero amplitude during the cardiac cycle. Our method is based upon application of a Genetic Algorithm (first 1000 samples), and then a linear regression to calculate the NCT for the rest of the recording.

Acute neuronal tissue preparations, brain slices and retinal wholemount, can usually only be maintained for 6 - 8 h following dissection. This limits the experimental time, and increases the number of animals that are utilized per study. This limitation specifically impacts protocols such as calcium imaging that require prolonged pre-incubation with bath-applied dyes. Exponential bacterial growth within 3 - 4 h after slicing is tightly correlated with a decrease in tissue health. This study describes a method for limiting the proliferation of bacteria in acute preparations to maintain viable neuronal tissue for prolonged periods of time (>24 h) without the need for antibiotics, sterile procedures, or tissue culture media containing growth factors. By cycling the extracellular fluid through UV irradiation and keeping the tissue in a custom holding chamber at 15 - 16 °C, the tissue shows no difference in electrophysiological properties, or calcium signaling through intracellular calcium dyes at >24 h postdissection. These methods will not only extend experimental time for those using acute neuronal tissue, but will reduce the number of animals required to complete experimental goals, and will set a gold standard for acute neuronal tissue incubation.