Search

This paper introduces a novel method for designing and simulating neuromorphic circuits for inference tasks, utilizing spiking neural networks (SNNs) trained ex-situ to offer practical advantages in efficiency and accuracy. The high effectiveness of this approach is demonstrated through efficient implementations for solving XOR and MNIST problems, achieved via specialized 2D crossbar circuit architectures integrated with operational amplifiers and Lapicque neuron models. Initially trained via a software application, the networks’ weights are then transferred to neuromorphic circuit designs simulated with PySpice and Ngspice tools. Detailed simulations exhibit robust inference capabilities, achieving 100% accuracy on XOR problems and over 89% on MNIST datasets. Comprehensive listings of hardware components are provided for direct use in real-world applications. By validating the practicality of ex-situ trained SNNs in neuromorphic hardware, this work underscores their potential for future advanced computational tasks in neuromorphic systems and opens avenues for developing more efficient and powerful artificial intelligence (AI) systems.

Effectively interfacing synthetic systems with a tangible world using a growing number and variety of sensors under the constraints of precision, resilience, and adaptability is indispensable. In particular, system integration employing leading-edge technologies is both rewarding and challenging because of signal swings, manufacturing deviations, and noise. Utilizing the general self-X concept and the transition from a common amplitude-domain to biology-inspired adaptive spike-domain processing offers a viable solution. In this work, a neuromorphic concept and the first prototype of an adaptive spiking sensory front-end with self-X properties were designed and fabricated using XFAB CMOS $0.35~\mu $ m technology. The chip includes synapse, neuron, self-adaptive spike-to-rank coding (SA-SRC), and adaptive coincidence detection (ACD) cells with areas 0.086 mm x 0.046 mm, 0.06 mm x 0.041 mm, 0.75mm x 1.3 mm, and 0.123 mm x 0.336 mm, respectively. The cells were applied to achieve an adaptive sensor signal-to-spike converter (ASSC) feeding SA-SRC, followed by a decoder to a 4-bit digital code. Characterization achieved differential non-linearity (DNL), integral non-linearity (INL), missing codes, effective number of bits (ENOB), and signal-to-noise and distortion ratio (SINAD), with values of 0.41 LSB, 0.3 LSB, no missing codes, 3.82 bits, and 24.79 dB respectively. The system consumed $321~\mu $ W of power, required 158 pJ per conversion, and had a conversion speed of 492 ns. A final angular decoder system application with Tunnel Magnetoresistance (TMR) sensors revealed our spiking sensory front-end’s ability to reduce the angle measurement error from 24.95 to 12.72 degrees due to adaptation after system perturbation.

In asynchronous Spiking Neural Networks (SNNs), the voltage division between passive resistive random-access memory (RRAM) arrays and neuron circuits presents a significant challenge, affecting the overall network accuracy and power efficiency. This study introduces the quantized-weight-splitting method (QWSM) as a novel solution to address this challenge. The QWSM optimizes and splits the quantized weights using the static read distortion owing to voltage division. The QWSM successfully guarantees inference accuracy and reduces power consumption. To validate the QWSM, the fabricated RRAM devices were measured, and a fitting model was carefully developed to describe their behavior, showing a strong correlation with measured data. In SNN simulations using the fitting model, the inference accuracy was improved across various weight quantization levels when the QWSM was applied. Moreover, the QWSM led to a substantial reduction in the average power consumption. Specifically, Compared to a network configured to have the smallest combined conductance of RRAMs for low-power operation, the network applying the QWSM showed a 12.56 -% reduction in average power per synapse. This power-saving feature, combined with improved accuracy, positions the QWSM as a valuable tool for efficient SNN design using passive RRAM arrays. Our findings highlight the potential of the QWSM in advancing neuromorphic computing with better energy efficiency and accuracy robustness.

Spiking Neural Networks (SNNs) have emerged as a prominent paradigm for brain-inspired computing, capable of processing temporal information and event-driven data in an efficient and biologically plausible manner. However, their revolutionary and complex nature is one of the key reasons why SNNs are not yet a widely used approach in contrast to traditional Artificial Neural Networks (ANNs). In this paper, we present a comprehensive SNN framework that offers user-friendly implementation. It has been designed so that it is compatible with other well-known software tools for data science, being easy to integrate with them. We showcase the versatility of the framework by applying it to various well-known benchmarking datasets, including image processing of handwritten numbers, time-series forecasting and an advance use case for speech recognition, achieving competitive results compared to traditional ANNs. Our SNN framework aims to bridge the gap between neuroscience and artificial intelligence, empowering researchers and practitioners with an accessible tool to explore the potential of neuro-inspired computing in advancing the field of AI.

While deep learning and backpropagation continue to dominate the field of machine learning in terms of benchmarks and versatility, recent neuroscientific advances shed light on more biologically plausible approaches. Spiking neural networks (SNNs), modelled after action potential dynamics, offer inherent time sensitivity and more efficiency in terms of performance to complexity. While investigating paradigms to support such alternatives, we attempt to answer whether reservoir computing can benefit from a spiking network based implementation with elements of biologically realistic models. This is done by varying both hyper-parameters and reservoir generation approaches and comparing implementations to spot potential improvements. We demonstrate how customized training of SNNs can result in competitive performance levels at lower operational complexity and be readily applied to other paradigms, such as the development of reservoir dynamics.

The massive deployment of the Internet of Things (IoT) combined with the need to reduce its impact on global energy consumption calls for the design of intelligent sensors. These sensors must be able to process information in situ to reduce the amount of data to be transferred, and to perform computing at low energy cost. Therefore, the design of intelligent sensors requires a compromise between performance and energy consumption. Spiking neural networks (SNNs) are good candidates for achieving the goal of an efficient intelligent sensor, as they use event-based computation. Interest for hardware implementation of SNNs has bloomed since the first experimental observation of memristors in 2008. In this paper, we study, by simulation means, the impact of the main technological parameters of Ferroelectric Tunnel Junctions (FTJs) synapses and of analog Leaky Integrate-and-Fire (LIF) neurons on the system learning capabilities. This allows us to determine which parameters are critical for the design of such systems and to suggest mitigation solutions as well as guidelines on how to build a SNN-based smart vision sensor in the context of unsupervised or reward-modulated learning. In particular, we show that splitting up the passive crossbar array of memristors could help dealing with the detrimental effect of the input voltage offset of the postsynaptic neurons.

Spiking neural network models characterize the emergent collective dynamics of circuits of biological neurons and help engineer neuro-inspired solutions across fields. Most dynamical systems’ models of spiking neural networks typically exhibit one of two major types of interactions: First, the response of a neuron’s state variable to incoming pulse signals (spikes) may be additive and independent of its current state. Second, the response may depend on the current neuron’s state and multiply a function of the state variable. Here we reveal that deterministic spiking neural network models with additive coupling are equivalent to models with multiplicative coupling for simultaneously modified intrinsic neuron time evolution. As a consequence, the same collective dynamics can be attained by state-dependent multiplicative and constant (state-independent) additive coupling. Such a mapping enables the transfer of theoretical results between spiking neural network models with different types of interaction mechanisms and at the same time extends the option space for hardware implementation or modeling. By allowing to choose the coupling type or neuron type that is the simplest one to implement in a given practical situation where a specific dynamic or functionality is required, it potentially allows simpler or more effective engineering applications.

Spiking Neural Networks (SNNs) are promising candidates for low-power and low-latency embedded artificial intelligence. However, those networks require event-based data produced by neuromorphic sensors which are not widely available, except for a few specialized devices like neuromorphic retinas. For other data types, a solution lies in the use of conventional sensors in conjunction with encoding layers. However, when performed in software, this solution can be detrimental to energy consumption or latency. Here we introduce a flexible design methodology for efficiently implementing, optimizing, and evaluating digital architectures of spike encoding integrating algorithms available in the literature. In order to quickly evaluate different hardware architectures and to tailor the solution to the application needs, our approach relies on High-Level Synthesis (HLS) tools and Python scripting. We illustrate the methodology by generating various digital architectures of two encoding algorithms taken from the literature and we evaluate their energy consumption and timing performances on Field Programmable Gate Arrays. This work could overcome the lack of neuromorphic sensors and accelerate the development of lower-power hardware SNNs.

How spiking neuronal networks encode memories in their different time and spatial scales constitute a fundamental topic in neuroscience and neuro-inspired engineering. Much attention has been paid to large networks and long-term memory, for example in models of associative memory. Smaller circuit motifs may play an important complementary role on shorter time scales, where broader network effects may be of less relevance. Yet, compact computational models of spiking neural networks that exhibit short-term volatile memory and actively hold information until their energy source is switched off, seem not fully understood. Here we propose that small spiking neural circuit motifs may act as volatile memory components. A minimal motif consists of only two interconnected neurons – one self-connected excitatory neuron and one inhibitory neuron – and realizes a single-bit volatile memory. An excitatory, delayed self-connection promotes a bistable circuit in which a self-sustained periodic orbit generating spike trains co-exists with the quiescent state of no neuron spiking. Transient external inputs may straightforwardly induce switching between those states. Moreover, the inhibitory neuron may act as an autonomous turn-off switch. It integrates incoming excitatory pulses until a threshold is reached after which the inhibitory neuron emits a spike that then inhibits further spikes in the excitatory neuron, terminating the memory. Our results show how external bits of information (excitatory signal), can be actively held in memory for a pre-defined amount of time. We show that such memory operations are robust against parameter variations and exemplify how sequences of multidimensional input signals may control the dynamics of a many-bits memory circuit in a desired way.

Sensors are an essential element in a wide range of applications. As the number of sensors increases, so does the amount of data collected with them. This raises the challenge of efficiently processing this data. Spiking Neural Networks (SNNs) represents a promising approach to solve this problem through event-based, parallelized data processing. For SNNs to be genuinely efficient, some fundamental challenges arise, like converting analog signals to spike events. An emerging possibility is the use of Resonate-and-Fire (R&F) neurons, capable of reacting to specific frequency components of input signals. In this work, we present a possible analog implementation for a R&F neuron and show the practical encoding of analog signals into a spiking domain using actual measurements. The coding method allows analog sensor signals to be directly applied to SNNs for efficient data processing. In the future, this approach can potentially enable the direct integration of analog Spiking Neural Networks into sensors.

Spiking Neural Networks (SNNs), the third generation of artificial neural networks, have been widely employed. However, the realization of advanced artificial intelligence is challenging due to the dearth of efficient spatiotemporal information integration models. Inspired by brain neuroscientists, this paper proposes a novel spiking neural network - Blended Glial Cell’s Spiking Neural Network (BGSNN). BGSNN introduces glial cells as spatiotemporal information processing units based on neurons and synapses, and also provides four new network dynamics connection models which extend the information processing dimension, enhance the network global information integration in the spatiotemporal domain, as well as the plasticity of neurons and synapses. In this paper, a BGSNN application - Sudoku solver is designed and implemented on the “WenTian” neuromorphic prototype. On the Easybrain dataset, the BGSNN solver achieves 100% accuracy, outperforming the same structure SNN solver by 97% at the Evil difficulty level, and has faster converges speed compared with the SOTA Sudoku solver LSGA. On the kaggle dataset, the BGSNN solver achieves over 99.99% accuracy, outperforming the publicly available optimal DNN solver under this dataset by 3.82%. In addition, BGSNN exhibits good parallelism and sparsity, decreasing computation by at least 92.9% compared to serial solvers and reducing sparsity by 88% compared to the equal fully dense DNN. BGSNN improves the expression, feedback, and regulation capabilities of neural networks while maintaining the advantages of SNN parallel sparsity, making it simpler to implement advanced artificial intelligence.

Depth estimation is an important computer vision task, useful in particular for navigation in autonomous vehicles, or for object manipulation in robotics. Here, we propose to solve it using StereoSpike, an end-to-end neuromorphic approach, combining two event-based cameras and a Spiking Neural Network (SNN) with a modified U-Net-like encoder-decoder architecture. More specifically, we used the Multi Vehicle Stereo Event Camera Dataset (MVSEC). It provides a depth ground-truth, which was used to train StereoSpike in a supervised manner, using surrogate gradient descent. We propose a novel readout paradigm to obtain a dense analog prediction–the depth of each pixel– from the spikes of the decoder. We demonstrate that this architecture generalizes very well, even better than its non-spiking counterparts, leading to near state-of-the-art test accuracy. To the best of our knowledge, it is the first time that such a large-scale regression problem is solved by a fully spiking neural network. Finally, we show that very low firing rates (< 5%) can be obtained via regularization, with a minimal cost in accuracy. This means that StereoSpike could be efficiently implemented on neuromorphic chips, opening the door for low power and real time embedded systems.

We present an end-to-end trainable modular event-driven neural architecture that uses local synaptic and threshold adaptation rules to perform transformations between arbitrary spatio-temporal spike patterns. The architecture represents a highly abstracted model of existing Spiking Neural Network (SNN) architectures. The proposed Optimized Deep Event-driven Spiking neural network Architecture (ODESA) can simultaneously learn hierarchical spatio-temporal features at multiple arbitrary time scales. ODESA performs online learning without the use of error back-propagation or the calculation of gradients. Through the use of simple local adaptive selection thresholds at each node, the network rapidly learns to appropriately allocate its neuronal resources at each layer for any given problem without using an error measure. These adaptive selection thresholds are the central feature of ODESA, ensuring network stability and remarkable robustness to noise as well as to the selection of initial system parameters. Network activations are inherently sparse due to a hard Winner-Take-All (WTA) constraint at each layer. We evaluate the architecture on existing spatio-temporal datasets, including the spike-encoded IRIS, latency-coded MNIST, Oxford Spike pattern and TIDIGITS datasets, as well as a novel set of tasks based on International Morse Code that we created. These tests demonstrate the hierarchical spatio-temporal learning capabilities of ODESA. Through these tests, we demonstrate ODESA can optimally solve practical and highly challenging hierarchical spatio-temporal learning tasks with the minimum possible number of computing nodes.

Visual depth perception is the basic function of the visual nervous system. To a stimulus in the stereo space, the visual nervous system could generate a perception about depth of its position. Experimental observations have demonstrated that interocular blur differences could lead to illusory perceptions about depths of moving stimuli. However, to stimuli with interocular blur differences, influences of different factors on illusory depth perceptions are still unclear. To explore these influences, this paper constructs a plastic two-layer k-winner-take-all (k-WTA) spiking network, simulating primary visual cortical responses. With incompatible stimuli presented into two eyes in experiments, binocular rivalry could occur in the primary visual cortex and interact with depth perception. To simulate binocular rivalry, the network consists of two parallel visual channels driven by left-eye and right-eye stimuli and competing with each other through mutual inhibition. In simulations, the horizontally moving stimulus is filtered with different Gaussian filters to generate paired monocular stimuli with interocular blur differences. The blurry strength, the moving direction and the moving speed perform as varying factors of moving stimuli. The network updates its dynamics through probabilistic inference, reflecting impacts of each factor on both neural responses and binocular rivalry. The modified responses could simulate illusory depth perceptions of stimuli as observed in experiments. To stimuli with interocular blur differences, varying factors could modify binocular rivalry in the network, inducing distinguishing illusory depth perceptions. Based on probabilistic inference, our model could provide possible explanations to illusory depth perceptions with interocular blur differences.

We propose a new learning algorithm to train spiking neural networks (SNN) using conventional artificial neural networks (ANN) as proxy. We couple two SNN and ANN networks, respectively, made of integrate-and-fire (IF) and ReLU neurons with the same network architectures and shared synaptic weights. The forward passes of the two networks are totally independent. By assuming IF neuron with rate-coding as an approximation of ReLU, we backpropagate the error of the SNN in the proxy ANN to update the shared weights, simply by replacing the ANN final output with that of the SNN. We applied the proposed proxy learning to deep convolutional SNNs and evaluated it on two benchmarked datasets of Fashion-MNIST and Cifar10 with 94.56% and 93.11% classification accuracy, respectively. The proposed networks could outperform other deep SNNs trained with tandem learning, surrogate gradient learning, or converted from deep ANNs. Converted SNNs require long simulation times to reach reasonable accuracies while our proxy learning leads to efficient SNNs with much smaller simulation times. The source codes of the proposed method are publicly available at https://github.com/SRKH/ProxyLearning.

We present a mixed-signal spiking neural networks processor with 8-bit synaptic weight on-chip learning in 40 nm CMOS that consists of a 10k mixed-signal synapse circuit and 100 analog leaky integrate-and-fire (LIF) neuron circuits. The processor has no clock signal except in peripheral circuits for I/O, and neuron and synapse circuits can operate asynchronously in the continuous-time domain, just like biological neurons. We demonstrate the energy efficiency of 6.24–18.7 TOPS/W in a multitarget spike learning task.

This paper presents a Spiking Neural Network(SNN) architecture to distinguish two musical instruments: piano and violin. The acoustic characteristics of music such as frequency and time convey a lot of information that help humans in distinguishing music instruments within few seconds. SNNs are neural networks that work effectively with temporal data. In this study, 2-layer SNN temporal based architecture is implemented for instrument (piano and violin) recognition. Further, this research investigates the behaviour of spiking neurons for piano and violin samples through different spike based statistics. Additionally, a Gamma metric that utilises spike time information and Root Mean Square Error (RMSE) from the membrane potential are used for classification and recognition. SNN achieved an overall classification accuracy of 92.38% and 93.19%, indicating the potential of SNNs in this inherently temporal recognition and classification domain. On the other hand, we implemented rate-coding techniques using machine learning (ML) techniques. Through this research, we demonstrated that SNN are more effective than conventional ML methods for capturing important the acoustic characteristics of music such as frequency and time. Overall, this research showed the potential capability of temporal coding over rate coding techniques while processing spatial and temporal data.

Despite recent developments in deep learning and their success in computer vision, model efficiency is increasingly becoming a vital factor for their deployment in various real-world applications. To provide a more effective form of computational capabilities, bio-inspired spiking neural networks (SNNs) have attracted considerable interest in recent years. SNNs are known as the third-generation of neural networks, which mimic how information is transmitted and processed in the human brain. SNNs enable sparse yet efficient information transmission through spike trains, leading to exceptional computational and energy efficiency. Nevertheless, critical challenges in SNNs to date are two-fold: (a) latency: the number of time steps required to achieve competitive results and (b) synaptic operations: the total number of spikes generated during inference. Many researchers have attempted to address these challenges, but their applications have been mainly limited to image classification. In this work, we present a threshold voltage balancing method for object detection in SNNs, which utilizes Bayesian optimization to find optimal threshold voltages in SNNs. We specifically design Bayesian optimization to consider important characteristics of SNNs such as latency and number of synaptic operations. Furthermore, we introduce two-phase threshold voltages to provide faster and more accurate object detection while providing high energy efficiency. According to experimental results, the proposed methods achieve the state-of-the-art object detection accuracy in SNNs, and converge 2x and 1.85x faster than conventional methods on PASCAL VOC and MS COCO, respectively. Moreover, the total number of synaptic operations is reduced by 40.33% and 45.31% on PASCAL VOC and MS COCO, respectively.

Learning binary weights to minimize the difference between target and actual outputs can be considered as a parameter optimization task within the given constraints, and thus, it belongs to the application domain of the Lagrange multiplier method (LMM). Based on the LMM, we propose a novel event-based weight binarization (eWB) algorithm for spiking neural networks (SNNs) with binary synaptic weights (−1, 1). The algorithm features (i) event-based asymptotic weight binarization using local data only, (ii) full compatibility with event-based end-to-end learning algorithms (e.g., event-driven random backpropagation (eRBP) algorithm), and (iii) the capability to address various constraints (including the binary weight constraint). As a proof of concept, we combine eWB with eRBP (eWB-eRBP) to obtain a single algorithm for learning binary weights to generate correct classifications. Fully connected SNNs were trained using eWB-eRBP and achieved an accuracy of 95.35% on MNIST. To the best of our knowledge, this is the first report on completely binary SNNs trained using an event-based learning algorithm. Given that eRBP with full-precision (32-bit) weights exhibited 97.20% accuracy, the binarization comes at the cost of an accuracy reduction of approximately 1.85%. The python code is available online: https://github.com/galactico7/eWB.

Spiking neural networks (SNNs) represent a promising alternative to conventional neural networks. In particular, the so-called Spike-by-Spike (SbS) neural networks provide exceptional noise robustness and reduced complexity. However, deep SbS networks require a memory footprint and a computational cost unsuitable for embedded applications. To address this problem, this work exploits the intrinsic error resilience of neural networks to improve performance and to reduce hardware complexity. More precisely, we design a vector dot-product hardware unit based on approximate computing with configurable quality using hybrid custom floating-point and logarithmic number representation. This approach reduces computational latency, memory footprint, and power dissipation while preserving inference accuracy. To demonstrate our approach, we address a design exploration flow using high-level synthesis and a Xilinx SoC-FPGA. The proposed design reduces $20.5\times $ computational latency and $8\times $ weight memory footprint, with less than 0.5% of accuracy degradation on a handwritten digit recognition task.

Over the past years Spiking Neural Networks (SNNs) models became attractive as a possible bridge to enable low-power event-driven neuromorphic hardware. SNNs have a high computational power due to the implicit employment of the biologically inspired input times. SNNs employ various parameters such as neuron threshold, synaptic delays, and weights in their structures. However, SNNs applications are still limited and elementary compared with other neural network architectures such as the Convolution Neural Networks (CNNs). In this research, a new SNN-based model named Adaptive Threshold Module (ATM) and its algorithm are proposed. The proposed ATM and algorithm depend on the adaptation of the internal spiking neuron threshold level. Adapting the threshold of the neurons is employed to control the spiking neuron firing rate to uniquely extract the main features of the input pattern that is in the shape of spike trains. It is shown that this technique works as an automated feature extraction method of input patterns in an efficient and faster way than other methods. The proposed method can preserve all information of the input spike trains. Simulations of the proposed model and the algorithm, using the challenging speech TIDIGITS dataset, sound RWCP dataset, and Poisson distribution spike trains, show encouraging results. The ATM can make SNN provide an accuracy surpassing that of the current state-of-the-art SNN algorithms and conventional non-spiking learning models.

In recent years, neuroscientists have discovered that the neural information is encoded by spike trains with precise times. Supervised learning algorithm based on the precise times for spiking neurons becomes an important research field. Although many existing algorithms have the excellent learning ability, most of their mechanisms still have some complex computations and certain limitations. Moreover, the discontinuity of spiking process also makes it very difficult to build an efficient algorithm. This paper proposes a supervised learning algorithm for spiking neurons using the kernel function of spike trains based on a unit of pair-spike. Firstly, we comprehensively divide the intervals of spike trains. Then, we construct an optimal selection and computation method of spikes based on the unit of pair-spike. This method avoids some wrong computations and reduces the computational cost by using each effective input spike only once in every epoch. Finally, we use the kernel function defined by an inner product operator to solve the computing problem of discontinue spike process and multiple output spikes. The proposed algorithm is successfully applied to many learning tasks of spike trains, where the effect of our optimal selection and computation method is verified and the influence of learning factors such as learning kernel, learning rate, and learning epoch is analyzed. Moreover, compared with other algorithms, all experimental results show that our proposed algorithm has the higher learning accuracy and good learning efficiency.

A phase change memory (PCM) model suitable for neuromorphic circuit simulations is developed. A crystallization ratio module is used to track the memory state in the SET process, and an active region radius module is developed to track the continuously varying amorphous region in the RESET process. To converge the simulations with bi-stable memory states, a predictive filament module is proposed using a previous state in iterations of nonlinear circuit matrix under a voltage-driven mode. Both DC and transient analysis are successfully converged in circuits with voltage sources. The spiking-time-dependent-plasticity (STDP) characteristics essential for synaptic PCM are successfully reproduced with SPICE simulations verifying the model's promising applications in neuromorphic circuit designs. Further on, the developed PCM model is applied to propose a neuron circuit topology with lateral inhibitions which is more bionic and capable of distinguishing fuzzy memories. Finally, unsupervised learning of handwritten digits on neuromorphic circuits is simulated to verify the integrity of models in a large-scale-integration circuits. For the first time in literature an emerging memory model is developed and applied successfully in neuromorphic circuit designs, and the model is applicable to flexible designs of neuron circuits for further performance improvements.

A Spiking Neural Network (SNN) can be trained indirectly by first training an Artificial Neural Network (ANN) with the conventional backpropagation algorithm, then converting it into an equivalent SNN. To reduce the computational cost of the resulting SNN as measured by the number of spikes, we present Logarithmic Temporal Coding (LTC), where the number of spikes used to encode an activation grows logarithmically with the activation value; and the accompanying Exponentiate-and-Fire (EF) neuron model, which only involves efficient bit-shift and addition operations. Moreover, we improve the training process of ANN to compensate for approximation errors due to LTC. Experimental results indicate that the resulting SNN achieves competitive performance in terms of classification accuracy at significantly lower computational cost than related work.

Automatic decision and lane-keeping tasks are the most fundamental but important tasks in robot controlling researches. Concerned about the computing limitations of mobile robot platforms, an easily trainable method with low computational consumption and low latency is needed to solve this problem. An unsupervised conditional reflex learning network was proposed in this paper, which uses the conditional pattern to learn the conditional pattern and make decisions in an unsupervised manner. We used the convolutional spiking neural network to extract hidden features of road lanes, and then used the dopamine modulation mechanism to learn the decision-making information from the acquired features. In order to evaluate the quality of automatic decision-making models, two metrics were designed which are total deviation distance per second (TDDPS) and target achievement rate during training (TAR). In the process of training, neither labels were given to the convolutional spiking neural network, nor artificial decision-making information was assigned to the dopamine modulation layer. Simulation experiments showed that the proposed model has a state-of-the-art performance in a relatively complicated scenario and solves three limitations mentioned in previous works. Our work brought more biological inspiration into decision support systems, with the hope that the proposed method can promote the development of the bionic decision support system in hardware, especially in neuromorphic hardware.

The computation of rank ordering plays a fundamental role in cognitive tasks and offers a basic building block for computing arbitrary digital functions. Spiking neural networks have been demonstrated to be capable of identifying the largest k out of N analog input signals through their collective nonlinear dynamics. By finding partial rank orderings, they perform k-winners-take-all computations. Yet, for any given study so far, the value of k is fixed, often to k equal one. Here we present a concept for spiking neural networks that are capable of (re)configurable computation by choosing k via one global system parameter. The spiking network acts via pulse-suppression induced by inhibitory pulse-couplings. Couplings are proportional to each units' state variable (neuron voltage), constituting an uncommon but straightforward type of leaky integrate-and-fire neural network. The result of a computation is encoded as a stable periodic orbit with k units spiking at some frequency and others at lower frequency or not at all. Orbit stability makes the resulting analog-to-digital computation robust to sufficiently small variations of both, parameters and signals. Moreover, the computation is completed quickly within a few spike emissions per neuron. These results indicate how reconfigurable k-winners-take-all computations may be implemented and effectively exploited in simple hardware relying only on basic dynamical units and spike interactions resembling simple current leakages to a common ground.

Neuromorphic computing systems are an emerging field that takes its inspiration from the biological neural architectures and computations inside the mammalian nervous system. The spiking neural networks (SNNs) mimic real biological neural networks by conveying information through the communication of short pulses between neurons. Since each neuron in these networks is connected to thousands of others, high bandwidth is required. Moreover, since the spike times are used to encode information in SNN, very low communication latency is also necessary. On the other hand, the combination of Two-dimensional Networks-on-Chip (2D-NoC) and Three-dimensional Integrated Circuits (3D-ICs) can provide a scalable interconnection fabric in large-scale parallel SNN systems. Although the SNNs have some intrinsic fault-tolerance properties, they are still susceptible to a significant amount of faults; especially, when we talk about integrating the large-scale SNN models in hardware. Consequently, the need for efficient solutions capable of avoiding any malfunctions or inaccuracies, as well as early fault-tolerance assessment, is becoming increasingly necessary for the design of future large-scale reliable neuromorphic systems. This paper first presents an analytical model to assess the effect of faulty connections on the performance of a 3D-NoC-based spiking neural network under different neural network topologies. Second, we present a fault-tolerant shortest-path k-means-based multicast routing algorithm (FTSP-KMCR) and architecture for spike routing in 3D-NoC of spiking neurons (3DFT-SNN). Evaluation results show that the proposed SP-KMCR algorithm reduces the average latency by 12.2% when compared to the previously proposed algorithm. In addition, the proposed fault-tolerant methodology enables the system to sustain correct traffic communication with a fault rate up to 20%, while only suffering 16.23% longer latency and 5.49% extra area cost when compared to the baseline architecture.

We propose a multi-modal perceptual system that is inspired by the inner working of the human brain; in particular, the hierarchical structure of the sensory cortex and the spatial-temporal binding criteria. The system is context independent and can be applied to many on-going problems in social robotics, including but not limited to person recognition, emotion recognition, and multi-modal robot doctor to name a few. The system encapsulates the parallel distributed processing of real-world stimuli through different sensor modalities and encoding them into features vectors which in turn are processed via a number of dedicated processing units (DPUs) through hierarchical paths. DPUs are algorithmic realizations of the cell assemblies in neuroscience. A plausible and realistic perceptual system is presented via the integration of the outputs from these units by spiking neural networks. We will also discuss other components of the system including top-down influences and the integration of information through temporal binding with fading memory and suggest two alternatives to realize these criteria. Finally, we will demonstrate the implementation of this architecture on a hardware platform as a social robot and report experimental studies on the system.

The paper develops a neuromorphic system on a Spartan 6 field programmable gate array (FPGA) board to generate locomotion patterns (gaits) for three different legged robots (biped, quadruped, and hexapod). The neuromorphic system consists of a reconfigurable FPGA-based architecture for a 3G artificial neural network (spiking neural network), which acts as a Central Pattern Generator (CPG). The locomotion patterns, are then generated by the CPG through a general neural architecture, which parameters are offline estimated by means of grammatical evolution and Victor-Purpura distance-based fitness function. The neuromorphic system is fully validated on real biped, quadruped, and hexapod robots.