A Tandem Learning Rule for Effective Training and Rapid Inference of Deep Spiking Neural Networks

Effective Active Learning Method for Spiking Neural Networks

A large quantity of labeled data is required to train high-performance deep spiking neural networks (SNNs), but obtaining labeled data is expensive. Active learning is proposed to reduce the quantity of labeled data required by deep learning models. However, conventional active learning methods in SNNs are not as effective as that in conventional artificial neural networks (ANNs) because of the difference in feature representation and information transmission. To address this issue, we propose an effective active learning method for a deep SNN model in this article. Specifically, a loss prediction module ActiveLossNet is proposed to extract features and select valuable samples for deep SNNs. Then, we derive the corresponding active learning algorithm for deep SNN models. Comprehensive experiments are conducted on CIFAR-10, MNIST, Fashion-MNIST, and SVHN on different SNN frameworks, including seven-layer CIFARNet and 20-layer ResNet-18. The comparison results demonstrate that the proposed active learning algorithm outperforms random selection and conventional ANN active learning methods. In addition, our method converges faster than conventional active learning methods.

Backpropagation-Based Learning Techniques for Deep Spiking Neural Networks: A Survey

With the adoption of smart systems, artificial neural networks (ANNs) have become ubiquitous. Conventional ANN implementations have high energy consumption, limiting their use in embedded and mobile applications. Spiking neural networks (SNNs) mimic the dynamics of biological neural networks by distributing information over time through binary spikes. Neuromorphic hardware has emerged to leverage the characteristics of SNNs, such as asynchronous processing and high activation sparsity. Therefore, SNNs have recently gained interest in the machine learning community as a brain-inspired alternative to ANNs for low-power applications. However, the discrete representation of the information makes the training of SNNs by backpropagation-based techniques challenging. In this survey, we review training strategies for deep SNNs targeting deep learning applications such as image processing. We start with methods based on the conversion from an ANN to an SNN and compare these with backpropagation-based techniques. We propose a new taxonomy of spiking backpropagation algorithms into three categories, namely, spatial, spatiotemporal, and single-spike approaches. In addition, we analyze different strategies to improve accuracy, latency, and sparsity, such as regularization methods, training hybridization, and tuning of the parameters specific to the SNN neuron model. We highlight the impact of input encoding, network architecture, and training strategy on the accuracy-latency tradeoff. Finally, in light of the remaining challenges for accurate and efficient SNN solutions, we emphasize the importance of joint hardware-software codevelopment.

A Hybrid Neural Coding Approach for Pattern Recognition With Spiking Neural Networks

Recently, brain-inspired spiking neural networks (SNNs) have demonstrated promising capabilities in solving pattern recognition tasks. However, these SNNs are grounded on homogeneous neurons that utilize a uniform neural coding for information representation. Given that each neural coding scheme possesses its own merits and drawbacks, these SNNs encounter challenges in achieving optimal performance such as accuracy, response time, efficiency, and robustness, all of which are crucial for practical applications. In this study, we argue that SNN architectures should be holistically designed to incorporate heterogeneous coding schemes. As an initial exploration in this direction, we propose a hybrid neural coding and learning framework, which encompasses a neural coding zoo with diverse neural coding schemes discovered in neuroscience. Additionally, it incorporates a flexible neural coding assignment strategy to accommodate task-specific requirements, along with novel layer-wise learning methods to effectively implement hybrid coding SNNs. We demonstrate the superiority of the proposed framework on image classification and sound localization tasks. Specifically, the proposed hybrid coding SNNs achieve comparable accuracy to state-of-the-art SNNs, while exhibiting significantly reduced inference latency and energy consumption, as well as high noise robustness. This study yields valuable insights into hybrid neural coding designs, paving the way for developing high-performance neuromorphic systems.

Training multi-layer spiking neural networks with plastic synaptic weights and delays

Spiking neural networks are usually considered as the third generation of neural networks, which hold the potential of ultra-low power consumption on corresponding hardware platforms and are very suitable for temporal information processing. However, how to efficiently train the spiking neural networks remains an open question, and most existing learning methods only consider the plasticity of synaptic weights. In this paper, we proposed a new supervised learning algorithm for multiple-layer spiking neural networks based on the typical SpikeProp method. In the proposed method, both the synaptic weights and delays are considered as adjustable parameters to improve both the biological plausibility and the learning performance. In addition, the proposed method inherits the advantages of SpikeProp, which can make full use of the temporal information of spikes. Various experiments are conducted to verify the performance of the proposed method, and the results demonstrate that the proposed method achieves a competitive learning performance compared with the existing related works. Finally, the differences between the proposed method and the existing mainstream multi-layer training algorithms are discussed.

Chip-In-Loop SNN Proxy Learning: a new method for efficient training of spiking neural networks

The primary approaches used to train spiking neural networks (SNNs) involve either training artificial neural networks (ANNs) first and then transforming them into SNNs, or directly training SNNs using surrogate gradient techniques. Nevertheless, both of these methods encounter a shared challenge: they rely on frame-based methodologies, where asynchronous events are gathered into synchronous frames for computation. This strays from the authentic asynchronous, event-driven nature of SNNs, resulting in notable performance degradation when deploying the trained models on SNN simulators or hardware chips for real-time asynchronous computation. To eliminate this performance degradation, we propose a hardware-based SNN proxy learning method that is called Chip-In-Loop SNN Proxy Learning (CIL-SPL). This approach effectively eliminates the performance degradation caused by the mismatch between synchronous and asynchronous computations. To demonstrate the effectiveness of our method, we trained models using public datasets such as N-MNIST and tested them on the SNN simulator or hardware chip, comparing our results to those classical training methods.

CQ + Training: Minimizing Accuracy Loss in Conversion From Convolutional Neural Networks to Spiking Neural Networks

Spiking neural networks (SNNs) are attractive for energy-constrained use-cases due to their binarized activation, eliminating the need for weight multiplication. However, its lag in accuracy compared to traditional convolutional network networks (CNNs) has limited its deployment. In this paper, we propose CQ+ training (extended "clamped" and "quantized" training), an SNN-compatible CNN training algorithm that achieves state-of-the-art accuracy for both CIFAR-10 and CIFAR-100 datasets. Using a 7-layer modified VGG model (VGG-*), we achieved 95.06% accuracy on the CIFAR-10 dataset for equivalent SNNs. The accuracy drop from converting the CNN solution to an SNN is only 0.09% when using a time step of 600. To reduce the latency, we propose a parameterized input encoding method and a threshold training method, which further reduces the time window size to 64 while still achieving an accuracy of 94.09%. For the CIFAR-100 dataset, we achieved an accuracy of 77.27% using the same VGG-* structure and a time window of 500. We also demonstrate the transformation of popular CNNs, including ResNet (basic, bottleneck, and shortcut block), MobileNet v1/2, and Densenet, to SNNs with near-zero conversion accuracy loss and a time window size smaller than 60. The framework was developed in PyTorch and is publicly available.

Direct training high-performance spiking neural networks for object recognition and detection

Introduction

The 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.

Methods

To 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 discussion

The 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.

Direct learning-based deep spiking neural networks: a review

The spiking neural network (SNN), as a promising brain-inspired computational model with binary spike information transmission mechanism, rich spatially-temporal dynamics, and event-driven characteristics, has received extensive attention. However, its intricately discontinuous spike mechanism brings difficulty to the optimization of the deep SNN. Since the surrogate gradient method can greatly mitigate the optimization difficulty and shows great potential in directly training deep SNNs, a variety of direct learning-based deep SNN works have been proposed and achieved satisfying progress in recent years. In this paper, we present a comprehensive survey of these direct learning-based deep SNN works, mainly categorized into accuracy improvement methods, efficiency improvement methods, and temporal dynamics utilization methods. In addition, we also divide these categorizations into finer granularities further to better organize and introduce them. Finally, the challenges and trends that may be faced in future research are prospected.

Spiking neural network with working memory can integrate and rectify spatiotemporal features

In the real world, information is often correlated with each other in the time domain. Whether it can effectively make a decision according to the global information is the key indicator of information processing ability. Due to the discrete characteristics of spike trains and unique temporal dynamics, spiking neural networks (SNNs) show great potential in applications in ultra-low-power platforms and various temporal-related real-life tasks. However, the current SNNs can only focus on the information a short time before the current moment, its sensitivity in the time domain is limited. This problem affects the processing ability of SNN in different kinds of data, including static data and time-variant data, and reduces the application scenarios and scalability of SNN. In this work, we analyze the impact of such information loss and then integrate SNN with working memory inspired by recent neuroscience research. Specifically, we propose Spiking Neural Networks with Working Memory (SNNWM) to handle input spike trains segment by segment. On the one hand, this model can effectively increase SNN's ability to obtain global information. On the other hand, it can effectively reduce the information redundancy between adjacent time steps. Then, we provide simple methods to implement the proposed network architecture from the perspectives of biological plausibility and neuromorphic hardware friendly. Finally, we test the proposed method on static and sequential data sets, and the experimental results show that the proposed model can better process the whole spike train, and achieve state-of-the-art results in short time steps. This work investigates the contribution of introducing biologically inspired mechanisms, e.g., working memory, and multiple delayed synapses to SNNs, and provides a new perspective to design future SNNs.

Explaining cocktail party effect and McGurk effect with a spiking neural network improved by Motif-topology

Network architectures and learning principles have been critical in developing complex cognitive capabilities in artificial neural networks (ANNs). Spiking neural networks (SNNs) are a subset of ANNs that incorporate additional biological features such as dynamic spiking neurons, biologically specified architectures, and efficient and useful paradigms. Here we focus more on network architectures in SNNs, such as the meta operator called 3-node network motifs, which is borrowed from the biological network. We proposed a Motif-topology improved SNN (M-SNN), which is further verified efficient in explaining key cognitive phenomenon such as the cocktail party effect (a typical noise-robust speech-recognition task) and McGurk effect (a typical multi-sensory integration task). For M-SNN, the Motif topology is obtained by integrating the spatial and temporal motifs. These spatial and temporal motifs are first generated from the pre-training of spatial (e.g., MNIST) and temporal (e.g., TIDigits) datasets, respectively, and then applied to the previously introduced two cognitive effect tasks. The experimental results showed a lower computational cost and higher accuracy and a better explanation of some key phenomena of these two effects, such as new concept generation and anti-background noise. This mesoscale network motifs topology has much room for the future.

A Co-Designed Neuromorphic Chip With Compact (17.9K F2) and Weak Neuron Number-Dependent Neuron/Synapse Modules

Many efforts have been made to improve the neuron integration efficiency on neuromorphic chips, such as using emerging memory devices and shrinking CMOS technology nodes. However, in the fully connected (FC) neuromorphic core, increasing the number of neurons will lead to a square increase in synapse & dendrite costs and a high-slope linear increase in soma costs, resulting in an explosive growth of core hardware costs. We propose a co-designed neuromorphic core (SRCcore) based on the quantized spiking neural network (SNN) technology and compact chip design methodology. The cost of the neuron/synapse module in SRCcore weakly depends on the neuron number, which effectively relieves the growth pressure of the core area caused by increasing the neuron number. In the proposed BICS chip based on SRCcore, although the neuron/synapse module implements 1∼16 times of neurons and 1∼66 times of synapses, it only costs an area of 1.79 × 10⁷ F², which is 7.9%∼38.6% of that in previous works. Based on the weight quantization strategy matched with SRCcore, quantized SNNs achieve 0.05%∼2.19% higher accuracy than previous works, thus supporting the design and application of SRCcore. Finally, a cross-modeling application is demonstrated based on the chip. We hope this work will accelerate the development of cortical-scale neuromorphic systems.

Bio-mimetic high-speed target localization with fused frame and event vision for edge application

Evolution has honed predatory skills in the natural world where localizing and intercepting fast-moving prey is required. The current generation of robotic systems mimics these biological systems using deep learning. High-speed processing of the camera frames using convolutional neural networks (CNN) (frame pipeline) on such constrained aerial edge-robots gets resource-limited. Adding more compute resources also eventually limits the throughput at the frame rate of the camera as frame-only traditional systems fail to capture the detailed temporal dynamics of the environment. Bio-inspired event cameras and spiking neural networks (SNN) provide an asynchronous sensor-processor pair (event pipeline) capturing the continuous temporal details of the scene for high-speed but lag in terms of accuracy. In this work, we propose a target localization system combining event-camera and SNN-based high-speed target estimation and frame-based camera and CNN-driven reliable object detection by fusing complementary spatio-temporal prowess of event and frame pipelines. One of our main contributions involves the design of an SNN filter that borrows from the neural mechanism for ego-motion cancelation in houseflies. It fuses the vestibular sensors with the vision to cancel the activity corresponding to the predator's self-motion. We also integrate the neuro-inspired multi-pipeline processing with task-optimized multi-neuronal pathway structure in primates and insects. The system is validated to outperform CNN-only processing using prey-predator drone simulations in realistic 3D virtual environments. The system is then demonstrated in a real-world multi-drone set-up with emulated event data. Subsequently, we use recorded actual sensory data from multi-camera and inertial measurement unit (IMU) assembly to show desired working while tolerating the realistic noise in vision and IMU sensors. We analyze the design space to identify optimal parameters for spiking neurons, CNN models, and for checking their effect on the performance metrics of the fused system. Finally, we map the throughput controlling SNN and fusion network on edge-compatible Zynq-7000 FPGA to show a potential 264 outputs per second even at constrained resource availability. This work may open new research directions by coupling multiple sensing and processing modalities inspired by discoveries in neuroscience to break fundamental trade-offs in frame-based computer vision.

Progressive Tandem Learning for Pattern Recognition With Deep Spiking Neural Networks

Spiking neural networks (SNNs) have shown clear advantages over traditional artificial neural networks (ANNs) for low latency and high computational efficiency, due to their event-driven nature and sparse communication. However, the training of deep SNNs is not straightforward. In this paper, we propose a novel ANN-to-SNN conversion and layer-wise learning framework for rapid and efficient pattern recognition, which is referred to as progressive tandem learning. By studying the equivalence between ANNs and SNNs in the discrete representation space, a primitive network conversion method is introduced that takes full advantage of spike count to approximate the activation value of ANN neurons. To compensate for the approximation errors arising from the primitive network conversion, we further introduce a layer-wise learning method with an adaptive training scheduler to fine-tune the network weights. The progressive tandem learning framework also allows hardware constraints, such as limited weight precision and fan-in connections, to be progressively imposed during training. The SNNs thus trained have demonstrated remarkable classification and regression capabilities on large-scale object recognition, image reconstruction, and speech separation tasks, while requiring at least an order of magnitude reduced inference time and synaptic operations than other state-of-the-art SNN implementations. It, therefore, opens up a myriad of opportunities for pervasive mobile and embedded devices with a limited power budget.

Spiking VGG7: Deep Convolutional Spiking Neural Network with Direct Training for Object Recognition

We propose a deep convolutional spiking neural network (DCSNN) with direct training to classify concrete bridge damage in a real engineering environment. The leaky-integrate-and-fire (LIF) neuron model is employed in our DCSNN that is similar to VGG. Poisson encoding and convolution encoding strategies are considered. The gradient surrogate method is introduced to realize the supervised training for the DCSNN. In addition, we have examined the effect of observation time step on the network performance. The testing performance for two different spike encoding strategies are compared. The results show that the DCSNN using gradient surrogate method can achieve a performance of 97.83%, which is comparable to traditional CNN. We also present a comparison with STDP-based unsupervised learning and a converted algorithm, and the proposed DCSNN is proved to have the best performance. To demonstrate the generalization performance of the model, we also use a public dataset for comparison. This work paves the way for the practical engineering applications of the deep SNNs.

SSTDP: Supervised Spike Timing Dependent Plasticity for Efficient Spiking Neural Network Training

Spiking Neural Networks (SNNs) are a pathway that could potentially empower low-power event-driven neuromorphic hardware due to their spatio-temporal information processing capability and high biological plausibility. Although SNNs are currently more efficient than artificial neural networks (ANNs), they are not as accurate as ANNs. Error backpropagation is the most common method for directly training neural networks, promoting the prosperity of ANNs in various deep learning fields. However, since the signals transmitted in the SNN are non-differentiable discrete binary spike events, the activation function in the form of spikes presents difficulties for the gradient-based optimization algorithms to be directly applied in SNNs, leading to a performance gap (i.e., accuracy and latency) between SNNs and ANNs. This paper introduces a new learning algorithm, called SSTDP, which bridges the gap between backpropagation (BP)-based learning and spike-time-dependent plasticity (STDP)-based learning to train SNNs efficiently. The scheme incorporates the global optimization process from BP and the efficient weight update derived from STDP. It not only avoids the non-differentiable derivation in the BP process but also utilizes the local feature extraction property of STDP. Consequently, our method can lower the possibility of vanishing spikes in BP training and reduce the number of time steps to reduce network latency. In SSTDP, we employ temporal-based coding and use Integrate-and-Fire (IF) neuron as the neuron model to provide considerable computational benefits. Our experiments show the effectiveness of the proposed SSTDP learning algorithm on the SNN by achieving the best classification accuracy 99.3% on the Caltech 101 dataset, 98.1% on the MNIST dataset, and 91.3% on the CIFAR-10 dataset compared to other SNNs trained with other learning methods. It also surpasses the best inference accuracy of the directly trained SNN with 25~32× less inference latency. Moreover, we analyze event-based computations to demonstrate the efficacy of the SNN for inference operation in the spiking domain, and SSTDP methods can achieve 1.3~37.7× fewer addition operations per inference. The code is available at: https://github.com/MXHX7199/SNN-SSTDP.