Search Results (85)

This paper presents the Post-Stimulus Time-Dependent Event Descriptor (P-TED), a novel “pure event” feature descriptor designed for neuromorphic vision data. Unlike conventional frame-based approaches or hybrid methods that transform event data into intermediate representations, P-TED operates directly on asynchronous event streams, thereby preserving the intrinsic low-latency and high-temporal-resolution advantages of event-based sensors. The descriptor integrates two complementary feature sets: a motion feature vector, which aggregates spatial relationships within a Moore neighbourhood to quantify stimulus direction, and a pattern feature vector, which employs rate encoding to capture temporal excitation signatures. The efficacy of the P-TED framework is validated through three distinct experiments: object and character recognition (MNIST-DVS and CIFAR10-DVS), mobile robot movement analysis, and complex non-rigid robotic hand gesture recognition (RoShamBo). Experimental results demonstrate that the P-TED achieves a significant reduction in classification latency, requiring only 2.7 ms compared to the 10.3 ms recorded by the state-of-the-art Distribution-Aware Retinal Transform (DART) framework. Additionally, P-TED exhibits superior robustness in disambiguating symmetric and mirrored motions, as well as in maintaining stability under non-linear fluctuations in event density caused by changing scale. This work establishes P-TED as a high-speed, computationally efficient, and explainable solution for real-time neuromorphic robotic vision systems. Full article

Despite the rapid proliferation of unmanned aerial vehicles (UAVs) across industrial, agricultural, and scientific domains, their deployment remains constrained by limited endurance, aerodynamic inefficiency, and acoustic emissions, all mediated by a shared onboard energy budget. Existing biomimetic UAV reviews have generally treated aerodynamics, structures, sensing, control, and energy systems as parallel topics rather than as interacting components of a unified aerial architecture. Drawing primarily on literature published between 2015 and June 2026 and identified through searches of Web of Science, Scopus, and Google Scholar, this review addresses this gap by examining bio-inspired technologies across six principal domains: aeroacoustic and passive flow control, aerodynamic efficiency, multifunctional structural composites, neuromorphic sensing and control, ionic energy storage, and energy harvesting. Its principal contribution is a cross-domain synergy analysis identifying five performance couplings and one structural enabling architecture through which these domains interact physically and functionally. Representative examples include serration-based propeller geometries that can simultaneously reduce noise and power demand; morphing wing surfaces that serve as both aerodynamic structures and triboelectric harvesting substrates; and neuromorphic spiking neural networks that have been reported, in specific event-vision inference benchmarks, to reduce inference energy by three to four orders of magnitude relative to embedded graphics processing unit (GPU)-based implementations. Mechanical harvesting outputs nonetheless remain orders of magnitude below propulsion requirements and are thus positioned as supplementary. Four systemic barriers (unquantified mass–energy balance, undocumented durability, aeroelastic co-design gaps, and heterogeneous metrics) are evaluated, and the resulting synthesis indicates that advancing bio-inspired UAVs requires a transition from structural imitation to functional, system-level biomimetics. Full article

Federated learning enables collaborative model training without centralizing private data, yet its deployment on edge devices remains constrained by communication overhead, heterogeneous data distributions, and privacy leakage from shared model updates. Spiking neural networks offer an energy-efficient alternative to conventional artificial neural networks by transmitting sparse binary events rather than dense activations, but existing federated spiking learning methods still suffer from inefficient gradient exchange and insufficient privacy protection under non-independent and identically distributed data. This paper proposes PriSparK, a privacy-preserving and communication-efficient federated spiking neural learning framework that jointly exploits temporal event sparsity, adaptive client-side spike-gradient compression, and privacy-calibrated aggregation. The core of PriSparK is a novel Event-Sparse Differentially Private Federated Spiking Optimization algorithm, which converts local surrogate gradients into spike-saliency-aware sparse updates, dynamically allocates communication budgets across layers and time steps, and injects calibrated Gaussian noise after clipping in a low-dimensional event subspace. To mitigate accuracy degradation caused by aggressive compression and privacy perturbation, PriSparK further introduces a membrane-aware error-feedback mechanism and a heterogeneity-adaptive server aggregation rule that weights client updates according to spike activity stability and local distribution drift. Experiments on neuromorphic and vision benchmarks, including N-MNIST, DVS128 Gesture, CIFAR-10, and Fashion-MNIST, show that PriSparK achieves competitive or superior accuracy compared with federated artificial neural and spiking baselines while substantially reducing uplink communication. Under strong privacy constraints, PriSparK maintains stable convergence and improves the accuracy–communication–privacy trade-off, demonstrating its potential for privacy-sensitive edge intelligence with event-driven neural computation. Full article

To address the challenges of severe optical attenuation and dynamic feature extraction for moving target depth retrieval in complex underwater remote sensing environments, this paper proposes a dynamic target depth estimation method based on multi-source data fusion. Taking optical RGB imagery and neuromorphic vision (NeuroIV) data as joint inputs, the proposed method constructs a three-channel feature extraction and fusion network. By leveraging a hypergraph structure, it establishes association weights between dynamic (temporal) and static (spatial) nodes to capture spatiotemporal correlations. To efficiently process the high-dimensional multi-modal data, the traditional dot-product attention is replaced with element-wise multiplication, significantly reducing computational complexity. Furthermore, a lightweight deformable attention pyramid (DAP) and diffusion model is introduced to refine depth image edges, effectively suppressing discontinuities and abruptness in the estimation results. Compared to single-modality optical imagery, the fused multi-modal data yields a superior signal-to-noise ratio and foreground contrast, achieving an improvement of over 20% in the MAE index. These results validate the effectiveness and superiority of the proposed multi-modal fusion strategy for dynamic target observation and depth retrieval in aquatic environments. Full article

Federated learning enables privacy-preserving training of spiking neural networks on distributed neuromorphic and event-based data, but it also exposes the global model to stealthy backdoor attacks injected by malicious clients. Compared with conventional artificial neural networks, federated spiking neural networks are more difficult to sanitize because malicious behavior may be encoded not only in spatial filters but also in spike timing, membrane dynamics, and temporal firing sparsity. Existing backdoor defenses usually require repeated federated retraining, access to client data, trigger synthesis, or computationally expensive model repair, which limits their practicality in low-power neuromorphic deployment. This paper proposes FedTSR, a fast post-training backdoor removal framework for spiking neural networks trained through federated learning. FedTSR introduces two coordinated algorithms: temporal synaptic risk estimation, which identifies backdoor-sensitive synaptic groups by measuring abnormal spike-response contributions across simulation timesteps using a small clean calibration set, and spike-consistency recalibration, which restores benign task performance through lightweight membrane-potential alignment and firing-rate regularization without restarting federated training. The proposed method is trigger-agnostic, client-independent, and compatible with surrogate-gradient trained spiking models. Here, trigger-agnostic means that FedTSR does not use trigger identity during repair; it should not be interpreted as certified robustness against every adaptive trigger family. Experiments on real neuromorphic and vision benchmarks, including N-MNIST, CIFAR10-DVS, DVS Gesture, and CIFAR-10, show that FedTSR substantially reduces attack success rate while preserving clean accuracy under multiple backdoor patterns, poisoning ratios, and non-IID federated settings. The results indicate that exploiting temporal spike dynamics provides an efficient and effective route for sanitizing compromised federated spiking neural networks. Full article

Low-latency visual perception is essential for interactive machine vision on edge AI devices, but conventional frame-based image sensors impose frame period delays and generate dense image data that increase memory bandwidth and processing latency. Although Dynamic Vision Sensors (DVSs) are known to provide low latency, sparse output, and high dynamic range, these sensor-level properties do not automatically translate into practical application-level latency reduction on resource-constrained edge platforms. This paper presents a latency-driven sensing algorithm co-design approach for DVS-based low-latency machine vision. The main objective is to connect DVS sensor-level characteristics, event representations, task-dependent processing flows, and measured response times on mobile application processors. We first analyze latency requirements for three representative edge AI applications (i.e., person detection, gesture recognition, and Simultaneous Localization and Mapping (SLAM)), which correspond to different latency regimes and processing structures. We then describe the DVS operating principle, pixel-level event latency, and readout latency, showing how asynchronous event generation reduces sensing delay and suppresses redundant static background information before algorithmic processing. In contrast to prior event camera studies that mainly optimize a single task or a specific event representation, this work evaluates three task-specific event processing systems on mobile processors. Person detection achieves 92 ms processing latency on Exynos 7570, gesture recognition based on event-driven 4-DoF motion estimation achieves 20 ms latency on Exynos 5422, and SLAM achieves 15.9 ms latency on Snapdragon 845. These results satisfy the practical latency targets of the corresponding applications and demonstrate that DVS-based sensing can provide not only sensor-level speed advantages but also system-level latency benefits for AIoT, mobile, robotics, and AR/VR machine vision systems. Full article

Conventional photodetectors and image sensors deliver high-fidelity digital outputs but face a growing data-movement bottleneck: the energy and latency cost of transferring raw pixel streams to off-chip memory and processors increasingly dominates over both sensing and computation in modern machine-vision pipelines. An emerging response is the neuromorphic photodetector, a class of optoelectronic device that converts incident light into an electrical signal while simultaneously storing, modulating, and pre-processing that signal in a manner inspired by biological synapses and retinas. Over the past decade, demonstrations have spanned at least eight material platforms—organic semiconductors, organic–carbon-nanotube hybrids, perovskite and perovskite hybrids, metal oxides (including ultra-wide-bandgap and printable variants), wide-bandgap III-nitrides and 4H-SiC, two-dimensional materials, photo-memristors, and silicon CMOS in-sensor compute architectures—and have been realised through four distinct architectural families: phototransistor synapses, photo-memristors, heterojunction in-sensor compute, and linear photovoltaic neural networks. Here, we provide a quantitative cross-platform benchmark across forty in-scope articles, identify persistent photoconductivity as a near-universal device-physical substrate underlying synaptic functionality, characterise the responsivity–speed–energy trade-off structure observed across platforms, and present a critical assessment of energy-reporting practice in the field. We further identify three best-practice exemplars from three independent material platforms that converge on operating biases of 0.01–0.1 V and energies of 0.07–0.8 fJ per event, and we propose a unified reporting framework to enable meaningful cross-platform benchmarking of next-generation neuromorphic photodetectors. Full article

Spiking neural networks (SNNs) have emerged as a promising paradigm for energy-efficient neuromorphic computing, particularly when processing asynchronous event streams from dynamic vision sensors (DVSs). However, SNNs often suffer from limited representational capacity and suboptimal feature recalibration compared to their artificial counterparts. To address these challenges, we propose SE-SNN, a novel architecture that integrates Squeeze-and-Excitation (SE) blocks into deep residual SNNs, enabling channel-wise attention without spike generation. Furthermore, we introduce a Robust Parametric Leaky Integrate-and-Fire (RobustPLIF) neuron model with learnable membrane time constant ($\tau$) and firing threshold ($v_{\mathrm{th}}$), allowing adaptive temporal dynamics in each layer. Our model is trained on the CIFAR10-DVS dataset.The experimental results demonstrate that SE-SNN achieves an accuracy of $78.8$% on CIFAR10-DVS with 16 time steps, outperforming baseline SNNs while maintaining biological plausibility and hardware efficiency. Ablation studies confirm the individual contributions of the SE blocks and learnable neuron parameters to the performance gains. Full article

Memristors, as transformative electronic devices designed to transcend the von Neumann architecture, enable the physical unification of information storage and computation, thereby offering a foundational hardware pathway toward energy-efficient, brain-inspired computing. Their intrinsic analog resistive switching, non-volatility, and history-dependent learning capabilities allow them to natively implement in-memory computing and emulate synaptic plasticity, addressing the critical bottlenecks of energy and speed in conventional systems. Notably, the evolution from electrically controlled memristors to optoelectronic memristors marks a paradigm shift from pure computing to integrated sensing-processing, opening new dimensions for high-speed, parallel, and adaptive signal processing. In recent years, significant progress has been made in the development of memristor-based neuromorphic vision and tactile systems, on-chip signal processors, and dynamic trajectory trackers, demonstrating their potential in edge intelligence, adaptive robotics, and real-time perceptual tasks. This review systematically summarizes the latest advances in memristor technology, providing a comprehensive analysis of their operating mechanisms, material and structural innovations, and cutting-edge applications in neuromorphic perception and computing. Furthermore, it discusses the key challenges and future directions for the development and integration of memristor-based systems in the post-von Neumann era. Full article

Flexible electrolyte-gated synaptic field-effect transistors (EGFETs) have emerged as a promising platform for neuromorphic visual systems, owing to their low-voltage operation, diverse synaptic plasticity, and exceptional mechanical flexibility. In particular, gel-based electrolytes, including hydrogels and ion gels, play a pivotal role as functional gate dielectrics, enabling efficient ion transport and strong ion–electron coupling through electric double-layer (EDL) formation. By leveraging these unique properties at the semiconductor/gel interface, EGFETs can effectively emulate essential biological synaptic behaviors, including short-term and long-term plasticity under optical stimulation. The inherent compatibility of EGFETs with a broad range of semiconductor channels, gel electrolytes, and flexible substrates enables the development of wearable and conformable neuromorphic platforms that seamlessly integrate sensing, memory, and signal processing within a single device architecture. Recent advances in gel material engineering, such as polymer network design, ionic modulation, and nanofiller incorporation, have significantly improved ion transport dynamics, interfacial stability, and device performance. Despite remaining challenges related to ion migration stability, multi-physical field coupling, and large-area device uniformity, these developments have substantially advanced the practical potential of gel-based systems. This review provides a comprehensive overview of the operating mechanisms, gel-based material systems, synaptic functionalities, mechanical reliability, and future prospects of flexible electrolyte-gated visual synapses, highlighting their considerable potential for next-generation intelligent perception and artificial vision technologies. Full article

Spiking Neural Networks (SNNs) executed on neuromorphic hardware promise energyefficient, low-latency inference well-suited to edge deployment in size, weight, and powerconstrained environments such as autonomous vehicles, wearable devices, and unmanned aerial platforms. However, a coherent research pathway to deployment of neuromorphic devices remains elusive. This paper presents a structured review and position on the state of SNN-based vision across four interconnected dimensions: network architectures, training methodologies, event-based datasets and simulation techniques, and neuromorphic computing hardware. We survey the evolution from shallow convolutional SNNs to spiking Transformers and hybrid designs which leverage the advantages of SNNs and conventional artificial neural networks. We also examine surrogate gradient training and ANN-to-SNN conversion approaches, catalogue real-world and simulated event-based datasets, and assess the landscape of neuromorphic platforms ranging from rigid mixed-signal architectures to fully-configurable digital systems. Our analysis reveals that while each area has matured considerably in isolation, critical integration challenges persist. In particular, event-based datasets remain scarce and lack standardisation, training methodologies introduce systematic gaps relative to deployment hardware, and access to neuromorphic platforms is restricted by proprietary toolchains and limited development kit availability. We conclude that bridging these integration gaps, rather than advancing individual components alone, represents the most important and least addressed work required to realise the potential of SNN-based vision at the edge. Full article

Event-based vision sensors (EVSs) provide unique frequency analysis opportunities due to their event data output and high temporal resolution. Anomaly detection methods used in hyperspectral analysis can be used on the event frequency spectra to detect targets. However, the introduction of a strong, flickering interfering source can reduce the EVS sensitivity and obscure targets of interest. Previous work presented a method showing that targets could still be detected through an overwhelming source using frequency analysis, background suppression, and statistical filtering. This paper extends that research and compares the ability of five different eigenanalysis anomaly detection methods (principal component background suppression (PCBS) with peak threshold detection, Mahalanobis distance (MD) detector, complementary subspace detector (CSD), Reed–Xiaoli (RX) detector, and subspace Reed–Xiaoli (SSRX) detector) to detect targets in a high noise environment. The PCBS, MD, and CSD detectors performed well and were able to detect the targets through the overwhelming source. The PCBS detector had the best performance at low false-alarm rates (a > 400% detection probability increase at a false-alarm probability of 10⁻⁵). While the MD and CSD detectors had the best detection at higher false-alarm probabilities (approximately 7 × 10⁻²), the MD detector had a sub-second execution time. Depending on the application, the PCBS or MD detector are the best choice out of these five methods to detect targets in this type of high noise environment. Full article

From falcons spotting prey to humans recognizing faces, the ability to rapidly process visual information depends on a foveated retinal organization that provides high-acuity central vision while preserving low-resolution peripheral vision. This organization is conserved along early visual pathways, yet remains under-explored in machine learning. Here, we examine the impact of embedding a foveated retinotopic transformation as a preprocessing layer on convolutional neural networks (CNNs) for image classification. By applying a log-polar mapping to off-the-shelf models and retraining them, we achieve comparable accuracy while improving robustness to scale and rotation. We demonstrate that this architecture is highly sensitive to shifts in the fixation point and that this sensitivity provides an effective proxy for defining saliency maps that facilitate object localization. Our results demonstrate that foveated retinotopy encodes prior geometric knowledge, providing a solution for visual searches and a meaningful classification robustness and localization trade-off. These findings provides a proof of concept in order to connect principles of biological vision with artificial networks, suggesting new, robust and efficient approaches for computer vision systems. Full article

In event-based neuromorphic processing, computer vision finds an efficient alternative capable of optimizing computational and energy resources, inspired by the dynamics of biological neural systems. In the development of real-time processing systems, it is crucial to visually represent the information captured by sensors and to explore its content with precision. Thus, machine learning models are implemented with the capability of being deployed on hardware devices with limited capabilities, depending on the intended purpose, ensuring savings in computational resources. The aim of this work was to evaluate the limits of the implemented neuron model, leaky-integrate and fire (LIF), for fitting convolutional layers of a neural network. To this end, the characteristics of the LIF neuron model used are summarized, as well as the details of its implementation in a hardware design, using configurable parameters. The experimental phase considered two convolution approaches to compare performance, Matlab R2022a software and a spiking convolutional processor for an FPGA, using sample recordings from the MNIST-DVS dataset and Sobel kernels for edge detection. The results reflect that the number of spikes generated by both approaches is very similar and their distribution by frame addresses is directly proportional. Full article

Dynamic Vision Sensor (DVS) is an event-based imaging technology inspired by biological photoreceptors, which holds great promise for edge computing. The event streams produced by DVS are often contaminated by Background Activity (BA) noise and hot-pixel noise, which degrade downstream processing. Existing filters typically use fixed parameters, resulting in poor adaptability to changing illumination. In this paper, we propose a lightweight Adaptive Event-based Filtering Spiking Neural Network (AEFSNN) to address these limitations. Inspired by homeostatic plasticity, AEFSNN dynamically adjusts neuronal thresholds by monitoring the input-to-output spike ratio, allowing the network to autonomously converge to an optimal operating point across different lighting conditions. Furthermore, we introduce a novel neuronal wake-up mechanism that inhibits processing neurons until triggered by valid input, which effectively suppresses redundant events generated by neighboring activity. Experiments show that AEFSNN is more robust under varying illumination. Compared with current filters, our method increases the Signal-to-Noise Ratio (SNR) of the output data by 1.42–2.33 dB. Additionally, the filtered data improves classification accuracy on downstream tasks, validating its practical value for neuromorphic vision systems. Full article