Authors:Dongwoo Lew, Hoyoung Tang, Jongsun Park Abstract: Recently, the accuracy of spike neural network (SNN) has been significantly improved by deploying convolutional neural networks (CNN) and their parameters to SNN. The deep convolutional SNNs, however, suffer from large amounts of computations, which is the major bottleneck for energy efficient SNN processor design. In this paper, we present an input-dependent computation reduction approach, where relatively unimportant neurons are identified and pruned without seriously sacrificing the accuracies. Specifically, a neuron pruning in temporal domain is proposed that prunes less important neurons and skips its future operations based on the layer-wise pruning thresholds of membrane voltages. To find the pruning thresholds, two pruning threshold search algorithms are presented that can efficiently trade-off accuracy and computational complexity with a given computation reduction ratio. The proposed neuron pruning scheme has been implemented using 65 nm CMOS process. The SNN processor achieves a 57% energy reduction and a 2.68× speed up, with up to 0.82% accuracy loss and 7.3% area overhead for CIFAR-10 dataset. PubDate: 2023-11-30T00:00:00Z
Authors:Manuel Jiménez, María J. Avedillo, Bernabé Linares-Barranco, Juan Núñez Abstract: Alternative paradigms to the von Neumann computing scheme are currently arousing huge interest. Oscillatory neural networks (ONNs) using emerging phase-change materials like VO2 constitute an energy-efficient, massively parallel, brain-inspired, in-memory computing approach. The encoding of information in the phase pattern of frequency-locked, weakly coupled oscillators makes it possible to exploit their rich non-linear dynamics and their synchronization phenomena for computing. A single fully connected ONN layer can implement an auto-associative memory comparable to that of a Hopfield network, hence Hebbian learning rule is the most widely adopted method for configuring ONNs for such applications, despite its well-known limitations. An extensive amount of literature is available about learning in Hopfield networks, with information regarding many different learning algorithms that perform better than the Hebbian rule. However, not all of these algorithms are useful for ONN training due to the constraints imposed by their physical implementation. This paper evaluates different learning methods with respect to their suitability for ONNs. It proposes a new approach, which is compared against previous works. The proposed method has been shown to produce competitive results in terms of pattern recognition accuracy with reduced precision in synaptic weights, and to be suitable for online learning. PubDate: 2023-11-29T00:00:00Z
Authors:Mohamed Sadek Bouanane, Dalila Cherifi, Elisabetta Chicca, Lyes Khacef Abstract: Spiking neural networks coupled with neuromorphic hardware and event-based sensors are getting increased interest for low-latency and low-power inference at the edge. However, multiple spiking neuron models have been proposed in the literature with different levels of biological plausibility and different computational features and complexities. Consequently, there is a need to define the right level of abstraction from biology in order to get the best performance in accurate, efficient and fast inference in neuromorphic hardware. In this context, we explore the impact of synaptic and membrane leakages in spiking neurons. We confront three neural models with different computational complexities using feedforward and recurrent topologies for event-based visual and auditory pattern recognition. Our results showed that, in terms of accuracy, leakages are important when there are both temporal information in the data and explicit recurrence in the network. Additionally, leakages do not necessarily increase the sparsity of spikes flowing in the network. We also investigated the impact of heterogeneity in the time constant of leakages. The results showed a slight improvement in accuracy when using data with a rich temporal structure, thereby validating similar findings obtained in previous studies. These results advance our understanding of the computational role of the neural leakages and network recurrences, and provide valuable insights for the design of compact and energy-efficient neuromorphic hardware for embedded systems. PubDate: 2023-11-24T00:00:00Z
Authors:Xiyan Wu, Yong Song, Ya Zhou, Yurong Jiang, Yashuo Bai, Xinyi Li, Xin Yang Abstract: Spiking Neural Networks (SNNs) have shown great promise in processing spatio-temporal information compared to Artificial Neural Networks (ANNs). However, there remains a performance gap between SNNs and ANNs, which impedes the practical application of SNNs. With intrinsic event-triggered property and temporal dynamics, SNNs have the potential to effectively extract spatio-temporal features from event streams. To leverage the temporal potential of SNNs, we propose a self-attention-based temporal-channel joint attention SNN (STCA-SNN) with end-to-end training, which infers attention weights along both temporal and channel dimensions concurrently. It models global temporal and channel information correlations with self-attention, enabling the network to learn ‘what’ and ‘when’ to attend simultaneously. Our experimental results show that STCA-SNNs achieve better performance on N-MNIST (99.67%), CIFAR10-DVS (81.6%), and N-Caltech 101 (80.88%) compared with the state-of-the-art SNNs. Meanwhile, our ablation study demonstrates that STCA-SNNs improve the accuracy of event stream classification tasks. PubDate: 2023-11-10T00:00:00Z
Authors:Pengfei Sun, Yansong Chua, Paul Devos, Dick Botteldooren Abstract: Spiking neural networks (SNNs), which are composed of biologically plausible spiking neurons, and combined with bio-physically realistic auditory periphery models, offer a means to explore and understand human auditory processing-especially in tasks where precise timing is essential. However, because of the inherent temporal complexity in spike sequences, the performance of SNNs has remained less competitive compared to artificial neural networks (ANNs). To tackle this challenge, a fundamental research topic is the configuration of spike-timing and the exploration of more intricate architectures. In this work, we demonstrate a learnable axonal delay combined with local skip-connections yields state-of-the-art performance on challenging benchmarks for spoken word recognition. Additionally, we introduce an auxiliary loss term to further enhance accuracy and stability. Experiments on the neuromorphic speech benchmark datasets, NTIDIDIGITS and SHD, show improvements in performance when incorporating our delay module in comparison to vanilla feedforward SNNs. Specifically, with the integration of our delay module, the performance on NTIDIDIGITS and SHD improves by 14% and 18%, respectively. When paired with local skip-connections and the auxiliary loss, our approach surpasses both recurrent and convolutional neural networks, yet uses 10 × fewer parameters for NTIDIDIGITS and 7 × fewer for SHD. PubDate: 2023-11-09T00:00:00Z
Authors:Ye Yue, Marc Baltes, Nidal Abuhajar, Tao Sun, Avinash Karanth, Charles D. Smith, Trevor Bihl, Jundong Liu Abstract: IntroductionThe field of machine learning has undergone a significant transformation with the progress of deep artificial neural networks (ANNs) and the growing accessibility of annotated data. ANNs usually require substantial power and memory usage to achieve optimal performance. Spiking neural networks (SNNs) have recently emerged as a low-power alternative to ANNs due to their sparsity nature. Despite their energy efficiency, SNNs are generally more difficult to be trained than ANNs.MethodsIn this study, we propose a novel three-stage SNN training scheme designed specifically for segmenting human hippocampi from magnetic resonance images. Our training pipeline starts with optimizing an ANN to its maximum capacity, then employs a quick ANN-SNN conversion to initialize the corresponding spiking network. This is followed by spike-based backpropagation to fine-tune the converted SNN. In order to understand the reason behind performance decline in the converted SNNs, we conduct a set of experiments to investigate the output scaling issue. Furthermore, we explore the impact of binary and ternary representations in SNN networks and conduct an empirical evaluation of their performance through image classification and segmentation tasks.Results and discussionBy employing our hybrid training scheme, we observe significant advantages over both ANN-SNN conversion and direct SNN training solutions in terms of segmentation accuracy and training efficiency. Experimental results demonstrate the effectiveness of our model in achieving our design goals. PubDate: 2023-11-01T00:00:00Z
Authors:Masataka Konishi, Kei M. Igarashi, Keiji Miura Abstract: In deep neural networks, representational learning in the middle layer is essential for achieving efficient learning. However, the currently prevailing backpropagation learning rules (BP) are not necessarily biologically plausible and cannot be implemented in the brain in their current form. Therefore, to elucidate the learning rules used by the brain, it is critical to establish biologically plausible learning rules for practical memory tasks. For example, learning rules that result in a learning performance worse than that of animals observed in experimental studies may not be computations used in real brains and should be ruled out. Using numerical simulations, we developed biologically plausible learning rules to solve a task that replicates a laboratory experiment where mice learned to predict the correct reward amount. Although the extreme learning machine (ELM) and weight perturbation (WP) learning rules performed worse than the mice, the feedback alignment (FA) rule achieved a performance equal to that of BP. To obtain a more biologically plausible model, we developed a variant of FA, FA_Ex-100%, which implements direct dopamine inputs that provide error signals locally in the layer of focus, as found in the mouse entorhinal cortex. The performance of FA_Ex-100% was comparable to that of conventional BP. Finally, we tested whether FA_Ex-100% was robust against rule perturbations and biologically inevitable noise. FA_Ex-100% worked even when subjected to perturbations, presumably because it could calibrate the correct prediction error (e.g., dopaminergic signals) in the next step as a teaching signal if the perturbation created a deviation. These results suggest that simplified and biologically plausible learning rules, such as FA_Ex-100%, can robustly facilitate deep supervised learning when the error signal, possibly conveyed by dopaminergic neurons, is accurate. PubDate: 2023-10-11T00:00:00Z
Authors:Rajalekshmi T. R, Rinku Rani Das, Chithra Reghuvaran, Alex James Abstract: Resistive random access memory is very well known for its potential application in in-memory and neural computing. However, they often have different types of device-to-device and cycle-to-cycle variability. This makes it harder to build highly accurate crossbar arrays. Traditional RRAM designs make use of various filament-based oxide materials for creating a channel that is sandwiched between two electrodes to form a two-terminal structure. They are often subjected to mechanical and electrical stress over repeated read-and-write cycles. The behavior of these devices often varies in practice across wafer arrays over these stresses when fabricated. The use of emerging 2D materials is explored to improve electrical endurance, long retention time, high switching speed, and fewer power losses. This study provides an in-depth exploration of neuro-memristive computing and its potential applications, focusing specifically on the utilization of graphene and 2D materials in RRAM for neural computing. The study presents a comprehensive analysis of the structural and design aspects of graphene-based RRAM, along with a thorough examination of commercially available RRAM models and their fabrication techniques. Furthermore, the study investigates the diverse range of applications that can benefit from graphene-based RRAM devices. PubDate: 2023-10-05T00:00:00Z
Authors:Siying Liu, Vincent C. H. Leung, Pier Luigi Dragotti Abstract: Spiking neural networks (SNNs) are well-suited to process asynchronous event-based data. Most of the existing SNNs use rate-coding schemes that focus on firing rate (FR), and so they generally ignore the spike timing in events. On the contrary, methods based on temporal coding, particularly time-to-first-spike (TTFS) coding, can be accurate and efficient but they are difficult to train. Currently, there is limited research on applying TTFS coding to real events, since traditional TTFS-based methods impose one-spike constraint, which is not realistic for event-based data. In this study, we present a novel decision-making strategy based on first-spike (FS) coding that encodes FS timings of the output neurons to investigate the role of the first-spike timing in classifying real-world event sequences with complex temporal structures. To achieve FS coding, we propose a novel surrogate gradient learning method for discrete spike trains. In the forward pass, output spikes are encoded into discrete times to generate FS times. In the backpropagation, we develop an error assignment method that propagates error from FS times to spikes through a Gaussian window, and then supervised learning for spikes is implemented through a surrogate gradient approach. Additional strategies are introduced to facilitate the training of FS timings, such as adding empty sequences and employing different parameters for different layers. We make a comprehensive comparison between FS and FR coding in the experiments. Our results show that FS coding achieves comparable accuracy to FR coding while leading to superior energy efficiency and distinct neuronal dynamics on data sequences with very rich temporal structures. Additionally, a longer time delay in the first spike leads to higher accuracy, indicating important information is encoded in the timing of the first spike. PubDate: 2023-10-02T00:00:00Z
Authors:Ammar Bitar, Rafael Rosales, Michael Paulitsch Abstract: IntroductionSpiking neural networks (SNNs) are a model of computation that mimics the behavior of biological neurons. SNNs process event data (spikes) and operate more sparsely than artificial neural networks (ANNs), resulting in ultra-low latency and small power consumption. This paper aims to adapt and evaluate gradient-based explainability methods for SNNs, which were originally developed for conventional ANNs.MethodsThe adapted methods aim to create input feature attribution maps for SNNs trained through backpropagation that process either event-based spiking data or real-valued data. The methods address the limitations of existing work on explainability methods for SNNs, such as poor scalability, limited to convolutional layers, requiring the training of another model, and providing maps of activation values instead of true attribution scores. The adapted methods are evaluated on classification tasks for both real-valued and spiking data, and the accuracy of the proposed methods is confirmed through perturbation experiments at the pixel and spike levels.Results and discussionThe results reveal that gradient-based SNN attribution methods successfully identify highly contributing pixels and spikes with significantly less computation time than model-agnostic methods. Additionally, we observe that the chosen coding technique has a noticeable effect on the input features that will be most significant. These findings demonstrate the potential of gradient-based explainability methods for SNNs in improving our understanding of how these networks process information and contribute to the development of more efficient and accurate SNNs. PubDate: 2023-09-27T00:00:00Z
Authors:Yuhang Li, Ruokai Yin, Youngeun Kim, Priyadarshini Panda Abstract: In this study, we explore Human Activity Recognition (HAR), a task that aims to predict individuals' daily activities utilizing time series data obtained from wearable sensors for health-related applications. Although recent research has predominantly employed end-to-end Artificial Neural Networks (ANNs) for feature extraction and classification in HAR, these approaches impose a substantial computational load on wearable devices and exhibit limitations in temporal feature extraction due to their activation functions. To address these challenges, we propose the application of Spiking Neural Networks (SNNs), an architecture inspired by the characteristics of biological neurons, to HAR tasks. SNNs accumulate input activation as presynaptic potential charges and generate a binary spike upon surpassing a predetermined threshold. This unique property facilitates spatio-temporal feature extraction and confers the advantage of low-power computation attributable to binary spikes. We conduct rigorous experiments on three distinct HAR datasets using SNNs, demonstrating that our approach attains competitive or superior performance relative to ANNs, while concurrently reducing energy consumption by up to 94%. PubDate: 2023-09-14T00:00:00Z
Authors:Yijian Pei, Changqing Xu, Zili Wu, Yi Liu, Yintang Yang Abstract: Spiking neural network (SNN) is a brain-inspired model with more spatio-temporal information processing capacity and computational energy efficiency. However, with the increasing depth of SNNs, the memory problem caused by the weights of SNNs has gradually attracted attention. In this study, we propose an ultra-low latency adaptive local binary spiking neural network (ALBSNN) with accuracy loss estimators, which dynamically selects the network layers to be binarized to ensure a balance between quantization degree and classification accuracy by evaluating the error caused by the binarized weights during the network learning process. At the same time, to accelerate the training speed of the network, the global average pooling (GAP) layer is introduced to replace the fully connected layers by combining convolution and pooling. Finally, to further reduce the error caused by the binary weight, we propose binary weight optimization (BWO), which updates the overall weight by directly adjusting the binary weight. This method further reduces the loss of the network that reaches the training bottleneck. The combination of the above methods balances the network's quantization and recognition ability, enabling the network to maintain the recognition capability equivalent to the full precision network and reduce the storage space by more than 20%. So, SNNs can use a small number of time steps to obtain better recognition accuracy. In the extreme case of using only a one-time step, we still can achieve 93.39, 92.12, and 69.55% testing accuracy on three traditional static datasets, Fashion- MNIST, CIFAR-10, and CIFAR-100, respectively. At the same time, we evaluate our method on neuromorphic N-MNIST, CIFAR10-DVS, and IBM DVS128 Gesture datasets and achieve advanced accuracy in SNN with binary weights. Our network has greater advantages in terms of storage resources and training time. PubDate: 2023-09-13T00:00:00Z
Authors:Zihan Yin, Md Abdullah-Al Kaiser, Lamine Ousmane Camara, Mark Camarena, Maryam Parsa, Ajey Jacob, Gregory Schwartz, Akhilesh Jaiswal Abstract: Neuromorphic image sensors draw inspiration from the biological retina to implement visual computations in electronic hardware. Gain control in phototransduction and temporal differentiation at the first retinal synapse inspired the first generation of neuromorphic sensors, but processing in downstream retinal circuits, much of which has been discovered in the past decade, has not been implemented in image sensor technology. We present a technology-circuit co-design solution that implements two motion computations—object motion sensitivity and looming detection—at the retina's output that could have wide applications for vision-based decision-making in dynamic environments. Our simulations on Globalfoundries 22 nm technology node show that the proposed retina-inspired circuits can be fabricated on image sensing platforms in existing semiconductor foundries by taking advantage of the recent advances in semiconductor chip stacking technology. Integrated Retinal Functionality in Image Sensors (IRIS) technology could drive advances in machine vision applications that demand energy-efficient and low-bandwidth real-time decision-making. PubDate: 2023-09-01T00:00:00Z
Authors:Longbin Zhang, Xiaochen Zhang, Xueyu Zhu, Ruoli Wang, Elena M. Gutierrez-Farewik Abstract: IntroductionResearch interest in exoskeleton assistance strategies that incorporate the user's torque capacity is growing rapidly. However, the predicted torque capacity from users often includes uncertainty from various sources, which can have a significant impact on the safety of the exoskeleton-user interface.MethodsTo address this challenge, this paper proposes an adaptive control framework for a knee exoskeleton that uses muscle electromyography (EMG) signals and joint kinematics. The framework predicted the user's knee flexion/extension torque with confidence bounds to quantify the uncertainty based on a neuromusculoskeletal (NMS) solver-informed Bayesian Neural Network (NMS-BNN). The predicted torque, with a specified confidence level, controlled the assistive torque provided by the exoskeleton through a TCP/IP stream. The performance of the NMS-BNN model was also compared to that of the Gaussian process (NMS-GP) model.ResultsOur findings showed that both the NMS-BNN and NMS-GP models accurately predicted knee joint torque with low error, surpassing traditional NMS models. High uncertainties were observed at the beginning of each movement, and at terminal stance and terminal swing in self-selected speed walking in both NMS-BNN and NMS-GP models. The knee exoskeleton provided the desired assistive torque with a low error, although lower torque was observed during terminal stance of fast walking compared to self-selected walking speed.DiscussionThe framework developed in this study was able to predict knee flexion/extension torque with quantifiable uncertainty and to provide adaptive assistive torque to the user. This holds significant potential for the development of exoskeletons that provide assistance as needed, with a focus on the safety of the exoskeleton-user interface. PubDate: 2023-08-30T00:00:00Z
Authors:Jongkil Park, Sohmyung Ha, Theodore Yu, Emre Neftci, Gert Cauwenberghs Abstract: Neuromorphic cognitive computing offers a bio-inspired means to approach the natural intelligence of biological neural systems in silicon integrated circuits. Typically, such circuits either reproduce biophysical neuronal dynamics in great detail as tools for computational neuroscience, or abstract away the biology by simplifying the functional forms of neural computation in large-scale systems for machine intelligence with high integration density and energy efficiency. Here we report a hybrid which offers biophysical realism in the emulation of multi-compartmental neuronal network dynamics at very large scale with high implementation efficiency, and yet with high flexibility in configuring the functional form and the network topology. The integrate-and-fire array transceiver (IFAT) chip emulates the continuous-time analog membrane dynamics of 65 k two-compartment neurons with conductance-based synapses. Fired action potentials are registered as address-event encoded output spikes, while the four types of synapses coupling to each neuron are activated by address-event decoded input spikes for fully reconfigurable synaptic connectivity, facilitating virtual wiring as implemented by routing address-event spikes externally through synaptic routing table. Peak conductance strength of synapse activation specified by the address-event input spans three decades of dynamic range, digitally controlled by pulse width and amplitude modulation (PWAM) of the drive voltage activating the log-domain linear synapse circuit. Two nested levels of micro-pipelining in the IFAT architecture improve both throughput and efficiency of synaptic input. This two-tier micro-pipelining results in a measured sustained peak throughput of 73 Mspikes/s and overall chip-level energy efficiency of 22 pJ/spike. Non-uniformity in digitally encoded synapse strength due to analog mismatch is mitigated through single-point digital offset calibration. Combined with the flexibly layered and recurrent synaptic connectivity provided by hierarchical address-event routing of registered spike events through external memory, the IFAT lends itself to efficient large-scale emulation of general biophysical spiking neural networks, as well as rate-based mapping of rectified linear unit (ReLU) neural activations. PubDate: 2023-08-28T00:00:00Z
Authors:Srijanie Dey, Alexander G. Dimitrov Abstract: With the ongoing growth in the field of neuro-inspired computing, newly arriving computational architectures demand extensive validation and testing against existing benchmarks to establish their competence and value. In our work, we break down the validation step into two parts—(1) establishing a methodological and numerical groundwork to establish a comparison between neuromorphic and conventional platforms and, (2) performing a sensitivity analysis on the obtained model regime to assess its robustness. We study the neuronal dynamics based on the Leaky Integrate and Fire (LIF) model, which is built upon data from the mouse visual cortex spanning a set of anatomical and physiological constraints. Intel Corp.'s first neuromorphic chip “Loihi” serves as our neuromorphic platform and results on it are validated against the classical simulations. After setting up a model that allows a seamless mapping between the Loihi and the classical simulations, we find that Loihi replicates classical simulations very efficiently with high precision. This model is then subjected to the second phase of validation, through sensitivity analysis, by assessing the impact on the cost function as values of the significant model parameters are varied. The work is done in two steps—(1) assessing the impact while changing one parameter at a time, (2) assessing the impact while changing two parameters at a time. We observe that the model is quite robust for majority of the parameters with slight change in the cost function. We also identify a subset of the model parameters changes which make the model more sensitive and thus, need to be defined more precisely. PubDate: 2023-08-24T00:00:00Z
Authors:Andrea Castagnetti, Alain Pegatoquet, Benoît Miramond Abstract: In recent years, Deep Convolutional Neural Networks (DCNNs) have outreached the performance of classical algorithms for image restoration tasks. However, most of these methods are not suited for computational efficiency. In this work, we investigate Spiking Neural Networks (SNNs) for the specific and uncovered case of image denoising, with the goal of reaching the performance of conventional DCNN while reducing the computational cost. This task is challenging for two reasons. First, as denoising is a regression task, the network has to predict a continuous value (i.e., the noise amplitude) for each pixel of the image, with high precision. Moreover, state of the art results have been obtained with deep networks that are notably difficult to train in the spiking domain. To overcome these issues, we propose a formal analysis of the information conversion processing carried out by the Integrate and Fire (IF) spiking neurons and we formalize the trade-off between conversion error and activation sparsity in SNNs. We then propose, for the first time, an image denoising solution based on SNNs. The SNN networks are trained directly in the spike domain using surrogate gradient learning and backpropagation through time. Experimental results show that the proposed SNN provides a level of performance close to the state of the art with CNN based solutions. Specifically, our SNN achieves 30.18 dB of signal-to-noise ratio on the Set12 dataset, which is only 0.25 dB below the performance of the equivalent DCNN. Moreover we show that this performance can be achieved with low latency, i.e., using few timesteps, and with a significant level of sparsity. Finally, we analyze the energy consumption for different network latencies and network sizes. We show that the energy consumption of SNNs increases with longer latencies, making them more energy efficient compared to CNNs only for very small inference latencies. However, we also show that by increasing the network size, SNNs can provide competitive denoising performance while reducing the energy consumption by 20%. PubDate: 2023-08-11T00:00:00Z
Authors:Sangwoo Hwang, Yujin Hwang, Duhee Kim, Junhee Lee, Han Kyoung Choe, Junghyup Lee, Hongki Kang, Jaeha Kung Abstract: Recent developments in artificial neural networks and their learning algorithms have enabled new research directions in computer vision, language modeling, and neuroscience. Among various neural network algorithms, spiking neural networks (SNNs) are well-suited for understanding the behavior of biological neural circuits. In this work, we propose to guide the training of a sparse SNN in order to replace a sub-region of a cultured hippocampal network with limited hardware resources. To verify our approach with a realistic experimental setup, we record spikes of cultured hippocampal neurons with a microelectrode array (in vitro). The main focus of this work is to dynamically cut unimportant synapses during SNN training on the fly so that the model can be realized on resource-constrained hardware, e.g., implantable devices. To do so, we adopt a simple STDP learning rule to easily select important synapses that impact the quality of spike timing learning. By combining the STDP rule with online supervised learning, we can precisely predict the spike pattern of the cultured network in real-time. The reduction in the model complexity, i.e., the reduced number of connections, significantly reduces the required hardware resources, which is crucial in developing an implantable chip for the treatment of neurological disorders. In addition to the new learning algorithm, we prototype a sparse SNN hardware on a small FPGA with pipelined execution and parallel computing to verify the possibility of real-time replacement. As a result, we can replace a sub-region of the biological neural circuit within 22 μs using 2.5 × fewer hardware resources, i.e., by allowing 80% sparsity in the SNN model, compared to the fully-connected SNN model. With energy-efficient algorithms and hardware, this work presents an essential step toward real-time neuroprosthetic computation. PubDate: 2023-08-10T00:00:00Z
Authors:Hong Zhang, Yang Li, Bin He, Xiongfei Fan, Yue Wang, Yu Zhang Abstract: IntroductionThe spiking neural network (SNN) is a bionic model that is energy-efficient when implemented on neuromorphic hardwares. The non-differentiability of the spiking signals and the complicated neural dynamics make direct training of high-performance SNNs a great challenge. There are numerous crucial issues to explore for the deployment of direct training SNNs, such as gradient vanishing and explosion, spiking signal decoding, and applications in upstream tasks.MethodsTo address gradient vanishing, we introduce a binary selection gate into the basic residual block and propose spiking gate (SG) ResNet to implement residual learning in SNNs. We propose two appropriate representations of the gate signal and verify that SG ResNet can overcome gradient vanishing or explosion by analyzing the gradient backpropagation. For the spiking signal decoding, a better decoding scheme than rate coding is achieved by our attention spike decoder (ASD), which dynamically assigns weights to spiking signals along the temporal, channel, and spatial dimensions.Results and discussionThe SG ResNet and ASD modules are evaluated on multiple object recognition datasets, including the static ImageNet, CIFAR-100, CIFAR-10, and neuromorphic DVS-CIFAR10 datasets. Superior accuracy is demonstrated with a tiny simulation time step of four, specifically 94.52% top-1 accuracy on CIFAR-10 and 75.64% top-1 accuracy on CIFAR-100. Spiking RetinaNet is proposed using SG ResNet as the backbone and ASD module for information decoding as the first direct-training hybrid SNN-ANN detector for RGB images. Spiking RetinaNet with a SG ResNet34 backbone achieves an mAP of 0.296 on the object detection dataset MSCOCO. PubDate: 2023-08-08T00:00:00Z
Open Access journal
[96 journals]
