Runkai Li, Jia Xiong, Xiuyuan He, Jieru Zhao, Jiaqi Lv, Haowen Fang, Lei Qi, Xi Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Spiking neural networks (SNNs) have attracted much attention for their high energy efficiency and for recent advances in their classification performance. However, unlike traditional deep learning approaches, the analysis and study of the robustness of SNNs to adversarial examples remain relatively underdeveloped. In this work, we focus on advancing the adversarial attack side of SNNs and make three major contributions. First, we show that successful white-box adversarial attacks on SNNs are highly dependent on the underlying surrogate gradient estimation technique, even in the case of adversarially trained SNNs. Second, using the best single surrogate gradient estimation technique, we analyze the transferability of adversarial attacks on SNNs and other state-of-the-art architectures like Vision Transformers (ViTs), as well as CNNs. Our analyzes reveal two key areas where SNN adversarial attacks can be enhanced: no white-box attack effectively exploits the use of multiple surrogate gradient estimators for SNNs, and no single model attack is effective at generating adversarial examples misclassified by both SNNs and non-SNN models simultaneously. For our third contribution, we develop a new attack, the Mixed Dynamic Spiking Estimation (MDSE) attack to address these issues. MDSE utilizes a dynamic gradient estimation scheme to fully exploit multiple surrogate gradient estimator functions. In addition, our novel attack generates adversarial examples capable of fooling both SNN and non-SNN models simultaneously. The MDSE attack is as much as 91.4 % more effective on SNN/ViT model ensembles and provides a 3 × boost in attack effectiveness on adversarially trained SNN ensembles, compared to conventional white-box attacks like Auto-PGD. Our experiments are broad and rigorous, covering three datasets (CIFAR-10, CIFAR-100 and ImageNet) and nineteen classifier models (seven for each CIFAR dataset and five models for ImageNet). We will release a fully publicly available code repository for the models and attacks upon publication.

Spiking neural networks (SNNs) are bio-plausible computing models with high energy efficiency. The temporal dynamics of neurons and synapses enable them to detect temporal patterns and generate sequences. While Backpropagation Through Time (BPTT) is traditionally used to train SNNs, it is not suitable for online learning of embedded applications due to its high computation and memory cost as well as extended latency. In this work, we present Spatiotemporal Online Learning for Synaptic Adaptation (SOLSA), which is specifically designed for online learning of SNNs composed of Leaky Integrate and Fire (LIF) neurons with exponentially decayed synapses and soft reset. The algorithm not only learns the synaptic weight but also adapts the temporal filters associated to the synapses. Compared to the BPTT algorithm, SOLSA has much lower memory requirement and achieves a more balanced temporal workload distribution. Moreover, SOLSA incorporates enhancement techniques such as scheduled weight update, early stop training and adaptive synapse filter, which speed up the convergence and enhance the learning performance. When compared to other non-BPTT based SNN learning, SOLSA demonstrates an average learning accuracy improvement of 14.2%. Furthermore, compared to BPTT, SOLSA achieves a 5% higher average learning accuracy with a 72% reduction in memory cost.

Biologically inspired Spiking Neural Networks (SNNs) have attracted significant attention for their ability to provide extremely energy-efficient machine intelligence through event-driven operation and sparse activities. As artificial intelligence (AI) becomes ever more democratized, there is an increasing need to execute SNN models on edge devices. Existing works adopt weight pruning to reduce SNN model size and accelerate inference. However, these methods mainly focus on how to obtain a sparse model for efficient inference, rather than training efficiency. To overcome these drawbacks, in this paper, we propose a Neurogenesis Dynamics-inspired Spiking Neural Network training acceleration framework, NDSNN. Our framework is computational efficient and trains a model from scratch with dynamic sparsity without sacrificing model fidelity. Specifically, we design a new drop-and-grow strategy with decreasing number of non-zero weights, to maintain extreme high sparsity and high accuracy. We evaluate NDSNN using VGG-16 and ResNet-19 on CIFAR-10, CIFAR-100 and TinyImageNet. Experimental results show that NDSNN achieves up to 20.52% improvement in accuracy on Tiny-ImageNet using ResNet-19 (with a sparsity of 99%) as compared to other SOTA methods (e.g., Lottery Ticket Hypothesis (LTH), SET-SNN, RigL-SNN). In addition, the training cost of NDSNN is only 40.89% of the LTH training cost on ResNet-19 and 31.35% of the LTH training cost on VGG-16 on CIFAR-10.

Graph Convolutional Networks (GCNs) are pivotal in extracting latent information from graph data across various domains, yet their acceleration on mainstream GPUs is challenged by workload imbalance and memory access irregularity. To address these challenges, we present Accel-GCN, a GPU accelerator architecture for GCNs. The design of Accel-GCN encompasses: (i) a lightweight degree sorting stage to group nodes with similar degree; (ii) a block-level partition strategy that dynamically adjusts warp workload sizes, enhancing shared memory locality and workload balance, and reducing metadata overhead compared to designs like GNNAdvisor; (iii) a combined warp strategy that improves memory coalescing and computational parallelism in the column dimension of dense matrices. Utilizing these principles, we formulate a kernel for SpMM in GCNs that employs block-level partitioning and combined warp strategy. This approach augments performance and multi-level memory efficiency and optimizes memory bandwidth by exploiting memory coalescing and alignment. Evaluation of Accel-GCN across 18 benchmark graphs reveals that it outperforms cuSPARSE, GNNAdvisor, and graph-BLAST by factors of 1.17×, 1.86×, and 2.94× respectively. The results underscore Accel-GCN as an effective solution for enhancing GCN computational efficiency. The implementation can be found on Github*.

Deep reinforcement learning (DRL) has been applied for optimal control of autonomous UAV trajectory generation. The energy and payload capacity of small UAVs impose constraints on the complexity and size of the neural network. While Model compression has the potential to optimize the trained neural network model for efficient deployment on em-bedded platforms, pruning a neural network for DRL is more difficult due to the slow convergence in the training before and after pruning. In this work, we focus on improving the speed of DRL training and pruning. New reward function and action exploration are first introduced, resulting in convergence speedup by 34.14%. The framework that integrates pruning and DRL training is then presented with an emphasize on how to reduce the training cost. The pruning does not only improve computational performance of inference, but also reduces the training effort with-out compromising the quality of the trajectory. Finally, experimental results are presented. We show that the integrated training and pruning framework reduces 67.16% of the weight and improves trajectory success rate by 1.7%. It achieves a 4.43x reduction of the floating-point operations for the inference, resulting a measured 41.85% run time reduction.

Neuromorphic computing and spiking neural networks (SNN) mimic the behavior of biological systems and have drawn interest for their potential to perform cognitive tasks with high energy efficiency. However, some factors such as temporal dynamics and spike timings prove critical for information processing but are often ignored by existing works, limiting the performance and applications of neuromorphic computing. On one hand, due to the lack of effective SNN training algorithms, it is difficult to utilize the temporal neural dynamics. Many existing algorithms still treat neuron activation statistically. On the other hand, utilizing temporal neural dynamics also poses challenges to hardware design. Synapses exhibit temporal dynamics, serving as memory units that hold historical information, but are often simplified as a connection with weight. Most current models integrate synaptic activations in some storage medium to represent membrane potential and institute a hard reset of membrane potential after the neuron emits a spike. This is done for its simplicity in hardware, requiring only a “clear” signal to wipe the storage medium, but destroys temporal information stored in the neuron.In this work, we derive an efficient training algorithm for Leaky Integrate and Fire neurons, which is capable of training a SNN to learn complex spatial temporal patterns. We achieved competitive accuracy on two complex datasets. We also demonstrate the advantage of our model by a novel temporal pattern association task. Codesigned with this algorithm, we have developed a CMOS circuit implementation for a memristor-based network of neuron and synapses which retains critical neural dynamics with reduced complexity. This circuit implementation of the neuron model is simulated to demonstrate its ability to react to temporal spiking patterns with an adaptive threshold.

Although widely used in machine learning, backpropagation cannot directly be applied to SNN training and is not feasible on a neuromorphic processor that emulates biological neuron and synapses. This work presents a spike-based backpropagation algorithm with biological plausible local update rules and adapts it to fit the constraint in a neuromorphic hardware. The algorithm is implemented on Intel’s Loihi chip enabling low power in-hardware supervised online learning of multilayered SNNs for mobile applications. We test this implementation on MNIST, Fashion-MNIST, CIFAR-10 and MSTAR datasets with promising performance and energy-efficiency, and demonstrate a possibility of incremental online learning with the implementation.

Spiking neural network (SNN) has drawn research interests as it mimics dynamic activities of human brain and has the potential to perform real-time cognitive tasks. However, latency, throughput and flexibility of existing hardware implemented SNNs are limited. The conventional rate coding is inefficient in terms of accuracy and latency. Oversimplified SNN models adopted by neuromorphic hardware discard characteristics such as neuron dynamics and filter effects etc., which are critical for neural information processing. Recent research advancements show that the potential of SNN can be better utilized by moving beyond rate-based model and considering temporal information embedded in the spike sequences. However, these works employ complex biologically realistic SNN models, posing challenges to hardware complexity. Furthermore, most existing neuromorphic hardware are developed for specific SNN models, or aiming at replicating biological behaviors. There is a lack of general methodology for SNN design optimization. Novel hardware architecture and systematic optimization techniques are required for efficient FPGA implementation and support flexible SNN models.

Grounding free-form textual queries necessitates an understanding of these textual phrases and its relation to the visual cues to reliably reason about the described locations. Spatial attention networks are known to learn this relationship and focus its gaze on salient objects in the image. Thus, we propose to utilize spatial attention networks for image-level visual-textual fusion preserving local (word) and global (phrase) information to refine region proposals with an in-network Region Proposal Network (RPN) and detect single or multiple regions for a phrase query. We focus only on the phrase query - ground truth pair (referring expression) for a model independent of the constraints of the datasets i.e. additional attributes, context etc. For such referring expression dataset ReferIt game, our Multi-region Attention-assisted Grounding network (MAGNet) achieves over 12% improvement over the state-of-the-art. Without the context from image captions and attribute information in Flickr30k Entities, we still achieve competitive results compared to the state-of-the-art.

The recently discovered spatial-temporal information processing capability of bio-inspired Spiking neural networks (SNN) has enabled some interesting models and applications. However designing large-scale and high-performance model is yet a challenge due to the lack of robust training algorithms. A bio-plausible SNN model with spatial-temporal property is a complex dynamic system. Synapses and neurons behave as filters capable of preserving temporal information. As such neuron dynamics and filter effects are ignored in existing training algorithms, the SNN downgrades into a memoryless system and loses the ability of temporal signal processing. Furthermore, spike timing plays an important role in information representation, but conventional rate-based spike coding models only consider spike trains statistically, and discard information carried by its temporal structures. To address the above issues, and exploit the temporal dynamics of SNNs, we formulate SNN as a network of infinite impulse response (IIR) filters with neuron nonlinearity. We proposed a training algorithm that is capable to learn spatial-temporal patterns by searching for the optimal synapse filter kernels and weights. The proposed model and training algorithm are applied to construct associative memories and classifiers for synthetic and public datasets including MNIST, NMNIST, DVS 128 etc. Their accuracy outperforms state-of-the-art approaches.

There is an increasing demand to process streams of temporal data in energy-limited scenarios such as embedded devices, driven by the advancement and expansion of Internet of Things (IoT) and Cyber-Physical Systems (CPS). Spiking neural network has drawn attention as it enables low power consumption by encoding and processing information as sparse spike events, which can be exploited for event-driven computation. Recent works also show SNNs' capability to process spatial temporal information. Such advantages can be exploited by power-limited devices to process real-time sensor data. However, most existing SNN training algorithms focus on vision tasks and temporal credit assignment is not addressed. Furthermore, widely adopted rate encoding ignores temporal information, hence it's not suitable for representing time series. In this work, we present an encoding scheme to convert time series into sparse spatial temporal spike patterns. A training algorithm to classify spatial temporal patterns is also proposed. Proposed approach is evaluated on multiple time series datasets in the UCR repository and achieved performance comparable to deep neural networks.

The trajectory prediction is a critical and challenging problem in the design of an autonomous driving system. Many AI-oriented companies, such as Google Waymo, Uber and DiDi, are investigating more accurate vehicle trajectory prediction algorithms. However, the prediction performance is governed by lots of entangled factors, such as the stochastic behaviors of surrounding vehicles, historical information of self-trajectory, and relative positions of neighbors, etc. In this paper, we propose a novel graph-based information sharing network (GISNet) that allows the information sharing between the target vehicle and its surrounding vehicles. Meanwhile, the model encodes the historical trajectory information of all the vehicles in the scene. Experiments are carried out on the public NGSIM US-101 and I-80 Dataset and the prediction performance is measured by the Root Mean Square Error (RMSE). The quantitative and qualitative experimental results show that our model significantly improves the trajectory prediction accuracy, by up to 50.00%, compared to existing models.

In natural language processing (NLP), the "Transformer" architecture was proposed as the first transduction model replying entirely on self-attention mechanisms without using sequence-aligned recurrent neural networks (RNNs) or convolution, and it achieved significant improvements for sequence to sequence tasks. The introduced intensive computation and storage of these pre-trained language representations has impeded their popularity into computation and memory constrained devices. The field-programmable gate array (FPGA) is widely used to accelerate deep learning algorithms for its high parallelism and low latency. However, the trained models are still too large to accommodate to an FPGA fabric. In this paper, we propose an efficient acceleration framework, Ftrans, for transformer-based large scale language representations. Our framework includes enhanced block-circulant matrix (BCM)-based weight representation to enable model compression on large-scale language representations at the algorithm level with few accuracy degradation, and an acceleration design at the architecture level. Experimental results show that our proposed framework significantly reduce the model size of NLP models by up to 16 times. Our FPGA design achieves 27.07× and 81 × improvement in performance and energy efficiency compared to CPU, and up to 8.80× improvement in energy efficiency compared to GPU.

Driven by the expanse of Internet of Things (IoT) and Cyber-Physical Systems (CPS), there is an increasing demand to process streams of temporal data on embedded devices with limited energy and power resources. Among all potential solutions, neuromorphic computing with spiking neural networks (SNN) that mimic the behavior of brain, have recently been placed at the forefront. Encoding information into sparse and distributed spike events enables low-power implementations, and the complex spatial temporal dynamics of synapses and neurons enable SNNs to detect temporal pattern. However, most existing hardware SNN implementations use simplified neuron and synapse models ignoring synapse dynamic, which is critical for temporal pattern detection and other applications that require temporal dynamics. To adopt a more realistic synapse model in neuromorphic platform its significant computation overhead must be addressed. In this work, we propose an FPGA-based SNN with biologically realistic neuron and synapse for temporal information processing. An encoding scheme to convert continuous real-valued information into sparse spike events is presented. The event-driven implementation of synapse dynamic model and its hardware design that is optimized to exploit the sparsity are also presented. Finally, we train the SNN on various temporal pattern-learning tasks and evaluate its performance and efficiency as compared to rate-based models and artificial neural networks on different embedded platforms. Experiments show that our work can achieve 10X speed up and 196X gains in energy efficiency compared with GPU.

Continuous trajectory control of fixed-wing unmanned aerial vehicles (UAVs) is complicated when considering hidden dynamics. Due to UAV multi degrees of freedom, tracking methodologies based on conventional control theory, such as Proportional-Integral-Derivative (PID) has limitations in response time and adjustment robustness, while a model based approach that calculates the force and torques based on UAV's current status is complicated and rigid. We present an actor-critic reinforcement learning framework that controls UAV trajectory through a set of desired waypoints. A deep neural network is constructed to learn the optimal tracking policy and reinforcement learning is developed to optimize the resulting tracking scheme. The experimental results show that our proposed approach can achieve 58.14% less position error, 21.77% less system power consumption and 9.23% faster attainment than the baseline. The actor network consists of only linear operations, hence Field Programmable Gate Arrays (FPGA) based hardware acceleration can easily be designed for energy efficient real-time control.

Small Unmanned Aircraft Systems (sUAS) will be an important component of the smart city and intelligent transportation environments of the near future. The demand for sUAS related applications, such as commercial delivery and land surveying, is expected to grow rapidly in next few years. In general, sUAS traffic scheduling and management functions are needed to coordinate the launching of sUAS from different launch sites and plan their trajectories to avoid conflict while considering several other constraints such as expected arrival time, minimum flight energy, and availability of communication resources. However, as the airbone sUAS density grows in a certain area, it is difficult to foresee the potential airspace and communications resource conflicts and make immediate decisions to avoid them. To address this challenge, we present a temporal and spatial routing algorithm for sUAS trajectory management in a high density urban area. It plans sUAS movements in a spatial and temporal maze with the consideration of obstacles that are either static or dynamic in time. The routing allows the sUAS to avoid static no-fly areas (i.e. static obstacles) or other in-flight sUAS and areas that have congested communication resources (i.e. dynamic obstacles). The algorithm is evaluated using an agent-based simulation platform. The simulation results show that the proposed algorithm outperforms reference route management algorithms in many areas, especially in processing speed and memory efficiency. Detailed comparisons are provided for the sUAS flight time, the overall throughput, the conflict rate and communication resource utilization. The results demonstrate that our proposed algorithm can be used as a solution to improve the efficiency of airspace and communication resource utilization for next generation smart city and smart transportation.

Bio-inspired neuromorphic hardware is a research direction to approach brain's computational power and energy efficiency. Spiking neural networks (SNN) encode information as sparsely distributed spike trains and employ spike-timingdependent plasticity (STDP) mechanism for learning. Existing hardware implementations of SNN are limited in scale or do not have in-hardware learning capability. In this work, we propose a low-cost scalable Network-on-Chip (NoC) based SNN hardware architecture with fully distributed in-hardware STDP learning capability. All hardware neurons work in parallel and communicate through the NoC. This enables chip-level interconnection, scalability and reconfigurability necessary for deploying different applications. The hardware is applied to learn MNIST digits as an evaluation of its learning capability. We explore the design space to study the trade-offs between speed, area and energy. How to use this procedure to find optimal architecture configuration is also discussed.