Surrogate Gradient Research Articles

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.

Fully Spiking Actor Network With Intralayer Connections for Reinforcement Learning.

With the help of special neuromorphic hardware, spiking neural networks (SNNs) are expected to realize artificial intelligence (AI) with less energy consumption. It provides a promising energy-efficient way for realistic control tasks by combining SNNs with deep reinforcement learning (DRL). In this article, we focus on the task where the agent needs to learn multidimensional deterministic policies to control, which is very common in real scenarios. Recently, the surrogate gradient method has been utilized for training multilayer SNNs, which allows SNNs to achieve comparable performance with the corresponding deep networks in this task. Most existing spike-based reinforcement learning (RL) methods take the firing rate as the output of SNNs, and convert it to represent continuous action space (i.e., the deterministic policy) through a fully connected (FC) layer. However, the decimal characteristic of the firing rate brings the floating-point matrix operations to the FC layer, making the whole SNN unable to deploy on the neuromorphic hardware directly. To develop a fully spiking actor network (SAN) without any floating-point matrix operations, we draw inspiration from the nonspiking interneurons found in insects and employ the membrane voltage of the nonspiking neurons to represent the action. Before the nonspiking neurons, multiple population neurons are introduced to decode different dimensions of actions. Since each population is used to decode a dimension of action, we argue that the neurons in each population should be connected in time domain and space domain. Hence, the intralayer connections are used in output populations to enhance the representation capacity. This mechanism exists extensively in animals and has been demonstrated effectively. Finally, we propose a fully SAN with intralayer connections (ILC-SAN). Extensive experimental results demonstrate that the proposed method outperforms the state-of-the-art performance on continuous control tasks from OpenAI gym. Moreover, we estimate the theoretical energy consumption when deploying ILC-SAN on neuromorphic chips to illustrate its high energy efficiency.

Effective Surrogate Gradient Learning With High-Order Information Bottleneck for Spike-Based Machine Intelligence.

Brain-inspired computing technique presents a promising approach to prompt the rapid development of artificial general intelligence (AGI). As one of the most critical aspects, spiking neural networks (SNNs) have demonstrated superiority for AGI, such as low power consumption. Effective training of SNNs with high generalization ability, high robustness, and low power consumption simultaneously is a significantly challenging problem for the development and success of applications of spike-based machine intelligence. In this research, we present a novel and flexible learning framework termed high-order spike-based information bottleneck (HOSIB) leveraging the surrogate gradient technique. The presented HOSIB framework, including second-order and third-order formation, i.e., second-order information bottleneck (SOIB) and third-order information bottleneck (TOIB), comprehensively explores the common latent architecture and the spike-based intrinsic information and discards the superfluous information in the data, which improves the generalization capability and robustness of SNN models. Specifically, HOSIB relies on the information bottleneck (IB) principle to prompt the sparse spike-based information representation and flexibly balance its exploitation and loss. Extensive classification experiments are conducted to empirically show the promising generalization ability of HOSIB. Furthermore, we apply the SOIB and TOIB algorithms in deep spiking convolutional networks to demonstrate their improvement in robustness with various categories of noise. The experimental results prove the HOSIB framework, especially TOIB, can achieve better generalization ability, robustness and power efficiency in comparison with the current representative studies.

Biomimetic oculomotor control with spiking neural networks

Spiking neural networks (SNNs) are comprised of artificial neurons that, like their biological counterparts, communicate via electrical spikes. SNNs have been hailed as the next wave of deep learning as they promise low latency and low-power consumption when run on neuromorphic hardware. Current deep neural network models for computer vision often require power-hungry GPUs to train and run, making them great candidates to replace with SNNs. We develop and train a biomimetic, SNN-driven, neuromuscular oculomotor controller for a realistic biomechanical model of the human eye. Inspired by the ON and OFF bipolar cells of the retina, we use event-based data flow in the SNN to direct the necessary extraocular muscle-driven eye movements. We train our SNN models from scratch, using modified deep learning techniques. Classification tasks are straightforward to implement with SNNs and have received the most research attention, but visual tracking is a regression task. We use surrogate gradients and introduce a linear layer to convert membrane voltages from the final spiking layer into the desired outputs. Our SNN foveation network enhances the biomimetic properties of the virtual eye model and enables it to perform reliable visual tracking. Overall, with event-based data processed by an SNN, our oculomotor controller successfully tracks a visual target while activating 87.3% fewer neurons than a conventional neural network.

Sequential motion optimization with short-term adaptive moment estimation for deep learning problems

This work proposes a variant of Sequential Motion Optimization (SMO) framework called Sequential Motion Optimization with Short-term Adaptive Moment Estimation (SMO-SAdam) to optimize neural network parameters. SMO-SAdam embeds the mechanism of Adaptive Moment Estimation (Adam) into SMO to enhance its optimization effectiveness for solving large-scale problems. Particularly, first and second moments of solution candidates including the leader and the follower are calculated by a so-called short-term adaptive moment estimation and incorporated into SMO process to accelerate their parameter optimization updates. Surrogate gradients at update steps for calculating the required moments are theoretically extracted by chain rule through SMO motion chain. Numerical results on both standard and scientific deep learning (DL) benchmarks considering training and generalization performance confirm the superiority of the devised SMO-SAdam in comparison to state-of-the-art algorithms. Moreover, it is noteworthy that SMO-SAdam outstandingly solves scientific deep learning problems having highly complex loss landscapes, which indicates its great potential to be successfully employed to other applications in this field and different non-convex optimization issues as well.

A Privacy and Energy-Aware Federated Framework for Human Activity Recognition

Human activity recognition (HAR) using wearable sensors enables continuous monitoring for healthcare applications. However, the conventional centralised training of deep learning models on sensor data poses challenges related to privacy, communication costs, and on-device efficiency. This paper proposes a federated learning framework integrating spiking neural networks (SNNs) with long short-term memory (LSTM) networks for energy-efficient and privacy-preserving HAR. The hybrid spiking-LSTM (S-LSTM) model synergistically combines the event-driven efficiency of SNNs and the sequence modelling capability of LSTMs. The model is trained using surrogate gradient learning and backpropagation through time, enabling fully supervised end-to-end learning. Extensive evaluations of two public datasets demonstrate that the proposed approach outperforms LSTM, CNN, and S-CNN models in accuracy and energy efficiency. For instance, the proposed S-LSTM achieved an accuracy of 97.36% and 89.69% for indoor and outdoor scenarios, respectively. Furthermore, the results also showed a significant improvement in energy efficiency of 32.30%, compared to simple LSTM. Additionally, we highlight the significance of personalisation in HAR, where fine-tuning with local data enhances model accuracy by up to 9% for individual users.

Energy-Efficient and High-Performance Ship Classification Strategy Based on Siamese Spiking Neural Network in Dual-Polarized SAR Images

Ship classification using the synthetic aperture radar (SAR) images has a significant role in remote sensing applications. Aiming at the problems of excessive model parameters numbers and high energy consumption in the traditional deep learning methods for the SAR ship classification, this paper provides an energy-efficient SAR ship classification paradigm that combines spiking neural networks (SNNs) with Siamese network architecture, for the first time in the field of SAR ship classification, which is called the Siam-SpikingShipCLSNet. It combines the advantage of SNNs in energy consumption and the advantage of the idea in performances that use the Siamese neuron network to fuse the features from dual-polarized SAR images. Additionally, we migrated the feature fusion strategy from CNN-based Siamese neural networks to the SNN domain and analyzed the effects of various spiking feature fusion methods on the Siamese SNN. Finally, an end-to-end error backpropagation optimization method based on the surrogate gradient has been adopted to train this model. Experimental results tested on the OpenSARShip2.0 dataset have demonstrated the correctness and effectiveness of the proposed SAR ship classification strategy, which has the advantages of the higher accuracy, fewer parameters and lower energy consumption compared with the mainstream deep learning method of the SAR ship classification.

Dynamic layer-span connecting spiking neural networks with backpropagation training

Spiking Neural Network (SNN) is one of the mainstream frameworks for brain-like computing and neuromorphic computing, which has the potential to overcome current AI challenges, for example, low-power learning dynamic processes. However, there is still a huge gap in performance between SNN and artificial neural networks (ANN) in traditional supervised learning. One solution for this problem is to propose a better spiking neuron model to improve its memory ability for temporal data. This paper proposed a leaky integrate-and-fire (LIF) neuron model with dynamic postsynaptic potential and a layer-span connecting method for SNN trained using backpropagation. The dynamic postsynaptic potential LIF model allows the neurons dynamically release neurotransmitters in an SNN model, which mimics the activity of biological neurons. The layer-span connecting method enhances the long-distance memory ability of SNN. We also first introduced a cosh-based surrogate gradient for the backpropagation training of SNNs. We compared the SNN with cosh-based surrogate gradient (CSNN), CSNN with dynamic postsynaptic potential (Dyn-CSNN), layer-span connecting CSNN (Las-CSNN), and SNN model with all the proposed methods (DlaCSNN-BP) in three image classification and one text classification datasets. The experimental results exhibited that proposed SNN methods could outperform most of the previously proposed SNNs and ANNs in the same network structure. Among them, the proposed DlaCSNN-BP got the best classification performance. This result indicates that our proposed method can effectively improve the effect of SNN in supervised learning and reduce the gap with deep learning. This work also provides more possibilities for putting SNN into practical application.

BP-SRM: A directly training algorithm for spiking neural network constructed by spike response model

Spiking neural networks (SNNs) have attracted widespread attention due to their unique bio-interpretability and low-power properties, but the non-differentiability of discrete spike sequences fired by spiking neurons brings difficulties to the learning of SNNs. Recently, surrogate gradient (SG) and back-propagation through time (BPTT) have provided an excellent idea for training SNNs constructed by leaky integrate-and-fire (LIF) neuron models. The LIF neuron model has been widely used in previous SNNs due to its simplicity and low computational cost, however, this also limits its simulation of biological neuron dynamics, reducing the biological interpretability of SNNs. In this paper, we generalize SG and BPTT to SNNs constructed by spike response model (SRM) and propose the BP-SRM algorithm. Specifically, we address why BPTT succeeded in LIF neurons-based SNNs but failed in SRM neurons-based SNNs in previous researches. Then we establish an iterative form of the SRM neuron model by selecting different state variables. Based on iterable SRM, we get the spatiotemporal dependencies between the state variables of SRM neurons-based SNNs, which allows us to derive the gradient and update weights in SNN by BP-SRM. Then we design temporal channel normalization for BP-SRM and verify the performance of the SNNs on static image dataset, dynamic image dataset and engineering dataset, including Fashion-MNIST, Weather Dataset, N-MNIST, American Sign Language Dataset and bearing fault diagnosis dataset. The experiment results indicate that BP-SRM achieves the state-of-the-art performance of SNNs.

NUTS-BSNN: A non-uniform time-step binarized spiking neural network with energy-efficient in-memory computing macro

This work introduces a network architecture NUTS-BSNN: A Non-uniform Time-step Binarized Spiking Neural Network. NUTS-BSNN is a fully binarized spiking neural network with all binary weights, including the input and output layers. In the input and output layers, the weights are represented as stochastic series of numbers, while in the hidden layers, they are approximated to binary values for using simple XNOR-based computations. To compensate for the information loss due to binarization, we increased the convolutions at the input layer sequentially computed over multiple time-steps. The results from these operations are accumulated before generating spikes for the subsequent layers to increase the overall performance. We chose 14 time-steps for accumulation to achieve a good tradeoff between performance and inference latency. The proposed technique was evaluated using three datasets by direct training method and using a surrogate gradient algorithm. We achieved classification accuracies of 93.25%, 88.71%, and 70.31% on the Fashion-MNIST, CIFAR-10, and CIFAR-100 datasets, respectively. Further, we present an in-memory computing architecture for NUTS-BSNN, which limits resource and power consumption for hardware implementation.

An improved probabilistic spiking neural network with enhanced discriminative ability

The non-differentiability of the spike activity has been a hindrance to the development of high-performance spiking neural networks (SNNs). Current learning algorithms mainly focus on achieving attractive SNNs based on surrogate gradient or conversion, yet their performance is still limited. The probability-based SNNs use the probabilistic mechanism to smooth out spike activity, showing a promising way for training SNNs. This work optimizes the probabilistic mechanism and proposes the probabilistic firing mechanism (PFM) for spiking neurons. PFM enables differentiable spike activity and can be adapted to a variety of spiking neurons. In addition, to eliminate the negative influence of probabilistic uncertainty, the attention discrimination mechanism (ADM) is proposed, which enables the neurons to respond efficiently by adaptively distinguishing the salient elements of the input current. By fusing PFM, ADM, and Leaky Integrate-and-Fire (LIF) neurons, we constructed the Probabilistic Attention Leaky Integrate-and-Fire (PALIF) neuron and Probabilistic Attention Spiking Neural Network (PASNN). Ablation studies confirm the effectiveness of PFM and ADM, and indicate that PASNN is suitable for low-latency scenarios. Experiments on both static image and neuromorphic datasets, including CIFAR10, CIFAR100, N-MNIST, and CIFAR10-DVS, demonstrate that PASNN achieves competitive performance in terms of accuracy and inference speed.

Surrogate gradient scaling for directly training spiking neural networks

Surrogate gradient scaling for directly training spiking neural networks

Spiking neural networks for frame-based and event-based single object localization

Spiking neural networks (SNNs) have shown much promise as an energy-efficient alternative to artificial neural networks (ANNs). Such methods trained by surrogate gradient (SG) descent are now capable of dealing with frame- or event-based vision tasks beyond classification (e.g. object detection, semantic segmentation). However, important questions about their behavior w.r.t fundamental design choices remain under-explored (e.g. how a specific neural coding scheme impacts the performance/robustness? Does a higher temporal latency necessarily imply better results?). In this paper, we focus on single object localization as a context to analyze deep convolutional SNNs on (1) the importance of temporal latency (i.e. the number of time-steps) on performance ; (2) their robustness against sensor corruptions ; and (3) the impact of neural coding schemes on performance with static images. We design a simple SNN baseline for frame- and event-based single object localization and compare it against a similar ANN architecture. Our experiments show that our approach can achieve competitive or better performance in accuracy and robustness against common sensor corruptions with significantly lower energy consumption. More importantly, our experimental analysis draws conclusions significantly different from well-known studies focused on SNNs trained with bio-plausible learning rules, which helps in the design of SG-trained architectures, and offers insight to design priorities in future neuromorphic technologies.

Hybrid Spiking Fully Convolutional Neural Network for Semantic Segmentation

The spiking neural network (SNN) exhibits distinct advantages in terms of low power consumption due to its event-driven nature. However, it is limited to simple computer vision tasks because the direct training of SNNs is challenging. In this study, we propose a hybrid architecture called the spiking fully convolutional neural network (SFCNN) to expand the application of SNNs in the field of semantic segmentation. To train the SNN, we employ the surrogate gradient method along with backpropagation. The accuracy of mean intersection over union (mIoU) for the VOC2012 dataset is higher than that of existing spiking FCNs by almost 30%. The accuracy of mIoU can reach 39.6%. Moreover, the proposed hybrid SFCNN achieved excellent segmentation performance for other datasets such as COCO2017, DRIVE, and Cityscapes. Our hybrid SFCNN is a valuable and interesting contribution to extending the functionality of SNNs, especially for power-constrained applications.

SPIDEN: deep Spiking Neural Networks for efficient image denoising.

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%.

Hybrid photonic deep convolutional residual spiking neural networks for text classification.

Spiking neural networks (SNNs) offer powerful computation capability due to its event-driven nature and temporal processing. However, it is still limited to shallow structure and simple tasks due to the training difficulty. In this work, we propose a deep convolutional residual spiking neural network (DCRSNN) for text classification tasks. In the DCRSNN, the feature extraction is achieved via a convolution SNN with residual connection, using the surrogate gradient direct training technique. Classification is performed by a fully-connected network. We also suggest a hybrid photonic DCRSNN, in which photonic SNNs are used for classification with a converted training method. The accuracy of hard and soft reset methods, as well as three different surrogate functions, were evaluated and compared across four different datasets. Results indicated a maximum accuracy of 76.36% for MR, 91.03% for AG News, 88.06% for IMDB and 93.99% for Yelp review polarity. Soft reset methods used in the deep convolutional SNN yielded slightly better accuracy than their hard reset counterparts. We also considered the effects of different pooling methods and observation time windows and found that the convergence accuracy achieved by convolutional SNNs was comparable to that of convolutional neural networks under the same conditions. Moreover, the hybrid photonic DCRSNN also shows comparable testing accuracy. This work provides new insights into extending the SNN applications in the field of text classification and natural language processing, which is interesting for the resources-restrained scenarios.

FireFly: A High-Throughput Hardware Accelerator for Spiking Neural Networks With Efficient DSP and Memory Optimization

Spiking neural networks (SNNs) have been widely used due to their strong biological interpretability and high energy efficiency. With the introduction of the backpropagation algorithm and surrogate gradient, the structure of spiking neural networks has become more complex, and the performance gap with artificial neural networks has gradually decreased. However, most SNN hardware implementations for field-programmable gate arrays (FPGAs) cannot meet arithmetic or memory efficiency requirements, which significantly restricts the development of SNNs. They do not delve into the arithmetic operations between the binary spikes and synaptic weights or assume unlimited on-chip RAM resources by using overly expensive devices on small tasks. To improve arithmetic efficiency, we analyze the neural dynamics of spiking neurons, generalize the SNN arithmetic operation to the multiplex-accumulate operation, and propose a high-performance implementation of such operation by utilizing the DSP48E2 hard block in Xilinx Ultrascale FPGAs. To improve memory efficiency, we design a memory system to enable efficient synaptic weights and membrane voltage memory access with reasonable on-chip RAM consumption. Combining the above two improvements, we propose an FPGA accelerator that can process spikes generated by the firing neuron on-the-fly (FireFly). FireFly is the first SNN accelerator that incorporates DSP optimization techniques into SNN synaptic operations. FireFly is implemented on several FPGA edge devices with limited resources but still guarantees a peak performance of 5.53TOP/s at 300MHz. As a lightweight accelerator, FireFly achieves the highest computational density efficiency compared with existing research using large FPGA devices.

Implementation of Field-Programmable Gate Array Platform for Object Classification Tasks Using Spike-Based Backpropagated Deep Convolutional Spiking Neural Networks.

This paper investigates the performance of deep convolutional spiking neural networks (DCSNNs) trained using spike-based backpropagation techniques. Specifically, the study examined temporal spike sequence learning via backpropagation (TSSL-BP) and surrogate gradient descent via backpropagation (SGD-BP) as effective techniques for training DCSNNs on the field programmable gate array (FPGA) platform for object classification tasks. The primary objective of this experimental study was twofold: (i) to determine the most effective backpropagation technique, TSSL-BP or SGD-BP, for deeper spiking neural networks (SNNs) with convolution filters across various datasets; and (ii) to assess the feasibility of deploying DCSNNs trained using backpropagation techniques on low-power FPGA for inference, considering potential configuration adjustments and power requirements. The aforementioned objectives will assist in informing researchers and companies in this field regarding the limitations and unique perspectives of deploying DCSNNs on low-power FPGA devices. The study contributions have three main aspects: (i) the design of a low-power FPGA board featuring a deployable DCSNN chip suitable for object classification tasks; (ii) the inference of TSSL-BP and SGD-BP models with novel network architectures on the FPGA board for object classification tasks; and (iii) a comparative evaluation of the selected spike-based backpropagation techniques and the object classification performance of DCSNNs across multiple metrics using both public (MNIST, CIFAR10, KITTI) and private (INHA_ADAS, INHA_KLP) datasets.

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.

A convolutional spiking neural network with adaptive coding for motor imagery classification

Motor imagery (MI) signal classification is crucial for brain-computer interfaces (BCI). The third-generation neural network, spiking neural network (SNN), has rich neurodynamic properties in the spatiotemporal domain, and therefore it is more suitable for processing EEG signals. However, the feature extraction capability of the SNN previously applied to MI signal classification is limited by its structure, and the model’s classification accuracy is not comparable to the state-of-the-art algorithms. In this paper, we propose a spiking neural network model called SCNet, which combines the feature extraction capability of CNN with the biological interpretability of SNN, making the model structurally closer to the biological neuronal dynamical system and improving the classification accuracy. SCNet reduces information loss by adaptive coding with learnability and solves the training difficulties of spiking neural networks by surrogate gradient learning. We evaluated the performance of the proposed SCNet on three typically representative motor imagery datasets. The validation shows that the model outperforms state-of-the-art SNN-based MI classification methods and various ANN and machine learning methods. The experimental results demonstrate the generality and effectiveness of the proposed motor imagery EEG signal classification model. Better classification results can be obtained by designing a well-structured spiking neural network.