Abstract Imaging flow cytometry (IFC) provides spatial resolution and multi-parametric measurements to characterise cells of interest. It enables profound insights into cell signalling, co-localisation, cell-to-cell interaction, and DNA studies. However, the use of traditional frame-based sensors (FBS) introduces the classic trade-off between speed, resolution, and sensitivity, where improving one often degrades the others. This limitation has hindered both technological progress and widespread adoption. The large data volumes generated by IFC further complicate real-time applications. The spatial richness of IFC data, however, makes it highly compatible with machine learning (ML), enabling automated analysis, gating, and the discovery of rare events. To overcome FBS constraints, neuromorphic vision sensors (NVS) have recently been introduced. Their event-driven design, characterised by data sparsity, provides greater data-processing efficiency, temporal resolution, and fluorescence sensitivity compared to conventional imaging methods. This review examines advances in flow cytometry (FC) and IFC, evaluating their respective strengths and drawbacks. It highlights innovations in NVS that address long-standing limitations, explores recent advances in neuromorphic imaging cytometry (NIC) across different modalities and ML approaches, and discusses current challenges and future directions shaping this emerging field. NIC represents a pivotal step toward real-time, data-efficient cytometry platforms that integrate event-driven sensing with spiking neural computation.
Biological nervous systems encode information through transient, event-driven spikes, motivating neuromorphic hardware that can reproduce such dynamics efficiently. Diffusive memristors are strong candidates because their volatile, threshold-driven responses generate spike-like signals and support temporal processing. However, their switching relies on ionic motion and transient metal clusters that are highly sensitive to interfacial chemistry, leaving the influence of environmental factors unresolved. Here, the chemical role of moisture in enabling volatile threshold switching is systematically established in symmetric Pt/Ag/SiO _2 /Ag/Pt diffusive memristors using a rigorously controlled fabrication and measurement framework. Systematic comparisons show that interfacial water facilitates Ag oxidation and hydrated transport pathways, whereas its depletion suppresses switching or increases variability. These findings demonstrate that moisture is not a minor perturbation but a fundamental chemical requirement for stable operation in such device systems. Controlling interfacial hydration emerges as a key design principle for achieving reliable and commercially scalable diffusive memristors for neuromorphic hardware.
Neural population activity typically evolves on low-dimensional manifolds and can be described as trajectories through quasi-stable assembly states. Here we develop a unified definition of clustered neural networks with local excitatory–inhibitory balance in which enhanced within-cluster effective coupling is implemented through connection probability (structural clustering), synaptic efficacy (weight clustering), or any mixture of both. We introduce a single mixing parameter $\kappa \in [0,1]$ that redistributes a defined cluster contrast between connection probabilities and synaptic efficacies while preserving the mean input of the underlying balanced random network. Mean-field theory and binary-network simulations show that metastable dynamics are supported across the full $\kappa$ continuum. Varying $\kappa$ changes higher-order input structure, reshaping multistable regimes, correlation structure, and the balance between single- and multi-cluster episodes. Because real nervous systems jointly organize topology and synaptic strength, our approach provides a biologically interpretable parametrization of clustered assembly models and a basis for future models combining structural and functional plasticity. We further demonstrate metastable switching in spiking leaky integrate-and-fire networks and in a BrainScaleS-2 neuromorphic implementation for the cases of structural, weight, and combined clustering. The $\kappa$ -framework offers a controlled translation axis for neuromorphic and other constrained substrates, exposing trade-offs between routed synapse count, fan-in, synaptic weight resolution, and calibration when implementing attractor-based computational primitives.
Backpropagation (BP) remains the dominant and most successful method for training parameters of deep neural network models. However, BP relies on two computationally distinct phases, does not provide a satisfactory explanation of biological learning, and can be challenging to apply for training of networks with discontinuities or noisy node dynamics. By comparison, node perturbation (NP), also known as activity-perturbed forward gradients, proposes learning by the injection of noise into network activations, and subsequent measurement of the induced loss change. NP relies on two forward (inference) passes, does not make use of network derivatives, and has been proposed as a model for learning in biological systems. However, standard NP is highly data inefficient and can be unstable due to its unguided noise-based search process. In this work, we develop a modern perspective on NP by relating it to the directional derivative and incorporating input decorrelation. We find that a closer alignment with directional derivatives together with input decorrelation at every layer theoretically and practically enhances performance of NP learning with large improvements in parameter convergence and much higher performance on the test data, approaching that of BP. Furthermore, our novel formulation allows for application to noisy systems in which the noise process itself is inaccessible, which is of particular interest for on-chip learning in neuromorphic systems.
Abstract The advent of the big data and the Internet of Things (IoT) has created an urgent demand for novel devices capable of simultaneously realizing in-sensor and in-memory computing. Although the introduction of van der Waals (vdW) materials has provided corresponding solutions, their applications in in-sensor computing operating at the infrared (IR) band remains less explored. In this work, we developed an infrared-sensing memory device (ISMD) based on Se0.3Te0.7/CuInP2S6 (CIPS) vdW heterostructure, which combines Se0.3Te0.7 semiconductor with considerable infrared (IR) photoresponse as the channel and CIPS as the ferroelectric layer. Leveraging plasma-induced interfacial charge trapping as the dominant mechanism, which is further stabilized by the ferroelectric polarization effect of CIPS, the device exhibits multi-bit non-volatile storage characteristics under electrical control. In addition, benefiting from the superior IR photoresponse of Se0.3Te0.7, the device demonstrates an instantaneous response to 1550 nm infrared light, with the responsivity controllably modulated by electrically tuning the channel conductance state. We further simulated a convolutional neural network (CNN) system based on ISMD. The frontend of the architecture utilizes the optoelectronic properties of ISMD as convolution kernel parameters, enabling selective responses to edges in different directions. The backend is a fully connected classifier constrained by actual conductance values. Using this system, we perform handwritten digit classification on the Modified National Institute of Standards and Technology (MNIST) dataset, achieving an accuracy of over 90%. The ISMD provides new materials and structural options for preparing memory devices capable of sensing IR bands and offers a new paradigm for realizing integrated in-sensor and in-memory computing.
Conceptors are a powerful extension of reservoir computing (RC) that enable the selective recall and stabilization of internal dynamics. However, their application to physical RC remains challenging because measured reservoir states are inevitably affected by noise and physical perturbations. In this work, we propose a noise-robust cross-trial-correlation-based (CTC) method for computing conceptors in noisy reservoir systems. By exploiting the consistency of the reservoir response across repeated trials, the method suppresses noise contributions that are uncorrelated between measurements. Numerical simulations of leaky echo state networks under additive state noise and parameter drift show that CTC-based conceptors preserve the relevant internal dynamics and extend the operational range of the reservoir compared with standard conceptors and unconstrained reservoirs. In an autonomous-generation task, the CTC-based conceptor maintains predictive capability beyond 50% noise, where the other configurations fail to sustain autonomous generation. Under combined noise and parameter drift, the CTC approach maintains normalized root mean square error values below 0.3, while the alternative approaches considered exceed this error level in the tested conditions. These numerical results establish a concrete step towards adapting conceptor-based control to physical RC platforms.
This paper presents a novel neuromorphic tactile sensing system inspired by human skin–neuron–cortex pathways, developed to address the energy and privacy challenges in modern vehicular surveillance. The proposed system integrates a heterojunction piezoelectric transistor as a tactile sensory node based on stacked up structure of ZnO/NiO/ZnO semiconductors with a neuromorphic matrix computing chip designed for on-chip classification of tactile events. The sensor exhibits an enhanced signal-to-noise ratio and high sensitivity via internal current and voltage amplification, while the neuromorphic circuit implements core neural network functions with parallel low-power processing. By extracting frequency-domain features to classify physical events such as impact, scraping, rainfall, and hail in parking scenarios, our system demonstrates real-time recognition with 93.1% accuracy on hardware and 99.2% in software, all while significantly reducing power consumption compared to conventional vision-based surveillance systems and avoiding visual privacy concerns. This work offers a promising solution for energy-efficient, event-driven smart surveillance in edge environments.
We report environmentally robust optoelectronic memristive devices based on an ITO/MoWS _2 /HfO _x /Pt heterostructure for neuromorphic computing applications. Individual dot-point devices exhibit stable bipolar resistive switching with an ON/OFF ratio of ∼1k and retention exceeding 10k s, alongside broadband, photo-tunable responsivity across the visible spectrum (405–785 nm), enabling wavelength-dependent synaptic modulation. To address scalability and device performance, 10 µ m ^2 crossbar arrays were developed, demonstrating endurance beyond 100k cycles and reliable operation under harsh conditions, including temperatures up to 200 °C and aqueous environments. Notably, the proposed crossbar platform combines high-temperature operation, direct water-exposure resilience, broadband optoelectronic functionality, and neuromorphic behavior within a single heterojunction architecture. The crossbar array devices show intrinsic time-dependent photoresponse decay that follows a single-exponential relaxation, providing a hardware-level analogue of synaptic forgetting. Based on these characteristics, a binary neural network simulation for image-diminishing tasks achieves a classification accuracy of 91.76%. These results establish the MoWS _2 /HfO _x heterostructure as a promising platform for resilient, multifunctional, and time-adaptive neuromorphic hardware.
We present a fully digital Ising solver for maximum-cut implemented on a low-cost Artix-7 field-programmable gate array (FPGA), where tabu-inspired inhibitory dynamics are realized as a compact, Block RAM-resident short-term memory. The solver operates in deterministic fixed-point arithmetic and scales up to the on-chip limit of $N = 850$ spins across multiple graph sizes and sparsity regimes. We benchmark the proposed Tabu-based dynamics against (i) a parallel Hopfield-network update rule implemented on the same FPGA, and (ii) a CPU-based quantum approximate optimization algorithm (QAOA) reference used as a fixed variational baseline for solution quality under a prescribed budget. Across the tested graph families, the Tabu-enhanced solver reaches higher cut values with reduced run-to-run dispersion than purely deterministic descent. Quantitatively, field-aligned warm start improves the median normalized cut of TS by 0.014 and reduces solve time by $0.18$ ms at $N = 100$ , $\rho = 20\%$ . At the largest tested size ( $N = 850$ ), TS improves the median normalized cut over the fixed-budget CPU-QAOA reference by 0.014–0.016, with median CPU–FPGA solve-time differences of 640–720 ms under the reported protocol. The $N = 850$ FPGA implementation closes timing at 100 MHz, demonstrating that short-term inhibitory memory can be embedded in a fully digital Ising network without relying on stochasticity, analog variability, or annealing schedules.
Spiking neural networks offer a promising route toward low-power sequence computation on neuromorphic hardware, but they continue to lag behind attention-based artificial neural networks on long-context tasks. A central open question is whether this gap reflects only implementation and optimization limitations, or whether architectural features of spiking computation impose unfavorable learnability constraints as sequence length increases. Here, we address this question using a covering-number analysis of feedforward non-leaky integrate-and-fire (nLIF) networks in the probably approximately correct framework. Building on causal-piece decompositions and local Lipschitz continuity, we derive a global sensitivity bound for feedforward nLIF networks and extend it from single-token inputs to multi-token spike sequences. For fixed architectures under stated boundedness and margin assumptions, the resulting sufficient worst-case sample requirement has leading quadratic dependence on sequence length. This dependence arises from cumulative causal participation across time and depth, which increases global sensitivity along active spike paths. We then test the mechanistic implications of this theory using finite-sample cue-recall and teacher–student benchmarks across spiking, recurrent, and attention-based model classes. In cue-recall, an early cue must be retained across distractors and reported at a final query token; in teacher–student, labels are generated by a fixed nLIF teacher, separating representability from finite-sample learnability. Unconstrained feedforward spiking models show sequence-length sensitivity, elevated hidden spike-participation density, and increased samples-to-threshold burden. Post-spike refractoriness, leak-mediated forgetting, learned lateral inhibition, and activity-constrained winner-take-all competition reduce hidden participation and improve empirical robustness in task- and regime-dependent ways. Together, these results identify diffuse causal-set growth as a fundamental architectural bottleneck for baseline feedforward spiking sequence models and suggest that scalable neuromorphic sequence architectures will require circuit mechanisms that explicitly constrain temporal accumulation and effective spike participation.
Automotive radar signal processing is supported by highly specialized hardware and software, yet rapidly increasing sensor bandwidths and antenna counts are driving growth in data rates and energy consumption that challenge existing architectures. Neuromorphic processors offer an alternative computational paradigm based on sparse, event-driven dynamics, but their suitability for real radar sensor setups remains largely unexplored. Here we present an experimental neuromorphic radar processing pipeline operating on real radar sensor data and executing on Intel’s Loihi 2. To our knowledge, this is the first Loihi 2-based radar processing pipeline demonstrated in a vehicle-mounted automotive radar setup. We implement key stages of the conventional radar pipeline, including Fourier transforms, non-coherent integration, and constant false-alarm rate detection, using spiking neural networks operating directly on streaming sensor data. By adapting the underlying radar algorithms to the constraints of neuromorphic hardware, we demonstrate real-time operation for selected on-chip configurations and characterize latency, throughput, and dynamic power across different system scales and data transfer modes. Our results show that low-level radar processing can be realized on neuromorphic hardware under realistic sensor conditions, while also identifying important bottlenecks in host transfers, routing, and digital implementation efficiency. At the same time, the experiments demonstrate that neuromorphic architectures can sustain low-latency on-chip processing and scalable multi-channel execution. These findings provide a concrete reference point for future work on more tightly integrated neuromorphic radar systems and on higher-level radar processing stages where event-driven computation may offer additional advantages.
Event-based vision is a neuromorphic sensing approach that enables low-latency, energy-aware perception by encoding brightness changes as sparse, asynchronous events rather than dense image frames. Event cameras provide microsecond-level timestamps and high dynamic range, supporting perception during fast motion and challenging illumination. In autonomy, event streams are often fused with complementary sensors such as inertial measurement units, color cameras, light-detection-and-ranging sensors, or radar to improve scale estimation, robustness, and redundancy. Event-centric fusion is challenging because sensor clocks and noise models differ across modalities, event rates are activity-dependent, event representations trade timing fidelity for computational convenience, and hardware cost is often dominated by communication and memory movement. Robustness to missing or degraded modalities is common in deployment but remains inconsistently evaluated. This topical review synthesizes event-based vision fusion around three coupled design decisions: fusion stage, temporal coupling, and event representation. We show how these choices determine whether systems preserve event-driven efficiency or revert to dense processing. As a case study, we review event-based depth estimation, consolidate benchmark results, and analyze missing-modality behavior. The evidence suggests that many gains are reported primarily as accuracy improvements, with limited analysis of modality dropout, event-rate shifts, timing mismatch, preprocessing cost, and efficiency. We relate recurring patterns to optical flow and semantic perception, connect them to neuromorphic hardware and hardware-algorithm co-design, and conclude with benchmarking recommendations and open problems in robustness, scalability, and training of spiking and hybrid models.
Recent advancements in brain-inspired complementary vision chips (CVS) with intensity, multi-bit temporal difference (TD) and spatial difference (SD) sensing capabilities offer a promising solution to the limitations of traditional image sensors by enabling high-speed, high-precision sensing with reduced bandwidth consumption. However, noise characterizations and denoising strategies for these sensors remain underexplored. In this study, leveraging a recently developed state-of-the-art CVS, Tianmouc, we present a theoretical analysis of its noise characteristics, revealing the main challenge for denoising: a distinctive distribution that varies with local illumination. Building on this analysis, we develop a suite of novel denoising algorithms named locally adaptive direction-aware filter (LADF). LADF implements multi-stage denoising algorithms consisting of preprocessing followed by an adaptive threshold filter that adjusts parameters locally to mitigate noise variability. Additionally, considering the distinct characteristics of SD, we develop a multi-directional and polarity-aware separation strategy, while for TD, we exploit the inherent time-space correlation between TD and SD to suppress noise further. To enable rigorous evaluation, we construct a large-scale paired dataset through a novel synthetic-real approach that combines accurately labeled synthetic images with real-world captured data. Experimental results demonstrate that LADF achieves an average signal-to-noise ratio (SNR) of 10.11 in SD, outperforming two baseline methods by factors of 1.54 & times; and 2.73 & times;, respectively, and an average SNR of 4.52 in TD, surpassing the baselines by 1.47 & times; and 3.57 & times;, respectively. Furthermore, our method reduces errors in motion estimation by 28.9%, and enhances the peak SNR in reconstruction by 3.35 dB, demonstrating its effectiveness in downstream tasks. Our algorithm establishes a new benchmark for CVS denoising and demonstrates significant potential to enhance the performance of application systems utilizing CVS.
Taking inspiration from the brain to design efficient computational systems remains challenging due to its complexity. This work investigates a key biological feature observed across mammals: a hierarchy of time scales in cortical areas. Experimental evidence shows that intrinsic neural dynamics slow down across the cortical hierarchy, forming a temporal hierarchy. We examine whether this property benefits artificial systems by introducing temporal hierarchy into spiking neural networks (SNNs), which inherently process information over time. We implement hierarchical time scales across neuronal, synaptic, and recurrent dynamics, and evaluate their effect under two settings: (1) as an inductive bias, and (2) as an emergent property through optimization. On temporal benchmarks such as multi-timescale-XOR and keyword spotting, hierarchical SNNs consistently outperform non-hierarchical ones, achieving 2%-6% higher accuracy and up to 5 & times; parameter reduction under iso-accuracy conditions. Moreover, when trained freely, temporal hierarchy emerges spontaneously through gradient descent. Finally, our theoretical analysis shows that hierarchical time constants enable processing of multi-frequency temporal signals with only log N layers-compared to N layers for non-hierarchical systems-highlighting hierarchy as a key organizational principle for efficient temporal computation.
This work presents a comprehensive study on amorphous LiNbO3 (LNO)-based memristors and their application in nonvolatile memory and neuromorphic computing. We demonstrate the fabrication of Ti/LNO/Au memristors exhibiting robust bipolar analog switching, excellent cycle-to-cycle uniformity, and long-term retention. These devices retain full switching functionality up to 400 degrees C, emphasizing the intrinsic thermal resilience of the amorphous LNO matrix and the stability of oxygen-vacancy-mediated filament dynamics. The devices further display a range of synaptic behaviors such as long-term potentiation, long-term depression, paired-pulse facilitation, and spike-timing-dependent plasticity, which are essential for biologically plausible learning rules in neuromorphic systems. The history-dependent conductance evolution of the memristors enables metaplasticity, which adaptively regulates learning rates and mitigates catastrophic forgetting in sequential MNIST tasks while maintaining high accuracy even at 200 degrees C. These findings position amorphous LNO as a promising material for next-generation memristive hardware, uniquely combining thermal endurance, analog plasticity, and metaplastic learning regulation.
The leaky integrate-and-fire (LIF) model is one of the most widely used models in characterizing neuronal dynamics. The simplicity and brevity of LIF models allow for analytical solutions that could quickly navigate the parameter space and provide insights of the neuronal computational properties. In this work, we developed analytical methods for computing neuronal dynamics of an LIF model receiving non-normal inputs, which have not been effectively characterized in existing works. We first characterize the distribution of a general input using the kernel density estimation (KDE), which accurately tracks varying non-normal features. Based on the KDE-distribution, we analytically derive the steady-state firing rates and inter-spike-interval distributions of the LIF model. As an example, we specify the input as consisting of two coupled stochastic Ornstein-Uhlenbeck processes and a sinusoidal drive; such total input is significantly non-normal in general. The analytical firing rate is highly consistent with direct model simulations, in response to wide ranges of different parameters. The analytical method developed in this work could be potentially implemented to general input current like experimental recordings, and help in facilitating neuromorphic computation of spiking network models, navigating parameter spaces for model fitting, increasing neuronal mechanistic understandings in terms of computational properties.
Forming a memory involves linking the different aspects that form an experience. Single-neuron recordings in humans suggest that, in the hippocampus, memories are encoded by partially overlapping engrams, with partial overlaps representing associations. Furthermore, theoretical work with static attractor networks has demonstrated that partial overlaps between engrams can support associative recall within a limited range, beyond which memories remain independent or merge into one. However, how overlaps emerge through plasticity in an attractor network remains unknown. Here we developed a modelling approach in order to explain how partial overlaps encoding associations can emerge as a function of repeated co-stimulation. We built on a previously validated dynamic attractor network model to which we introduced heterogeneous baseline firing rates as an additional stabilizing mechanism. We found that repeated co-stimulation of initially orthogonal engrams could induce the formation of shared representations, with the overlap size scaling with the fraction of paired to individual stimulations and modulated by neuronal excitability. These findings provide a mechanistic link between experimental observations of overlapping engrams and theoretical predictions about their functional role, offering new insights into the hippocampal coding of associative memory in humans, also informing the design of flexible neuromorphic memory systems.
Although neuromorphic computing promises energy-efficient and biologically inspired machine learning architectures, the implementation of neural dynamics into practical hardware design remains a major challenge. This work presents a unified analytical framework that integrates equilibrium propagation (EP) with oscillatory neural networks (ONNs) through a phase-deviation formulation. By exploiting the synchronization properties of oscillatory systems, we establish a mathematical link between circuit dynamics, learning behavior, and the underlying energy landscape derived from the phase-deviation equations. This formulation reveals how ONNs naturally operate as an analog associative memory, where equilibrium points correspond to minima of the energy function governing collective phase dynamics. We derive the full phase-deviation formalism for weakly coupled oscillatory networks and adapt EP to their dynamics. Numerical simulations on pattern-recognition tasks are provided as a supporting example, illustrating how the proposed analytical framework can reproduce learning and retrieval behavior in a phase-based ONN setting. These results support the use of the proposed formalism as a principled framework for analyzing equilibrium structure, learning dynamics, and possible failure modes in oscillatory neuromorphic systems.
Artificial intelligence (AI) and neuromorphic computing demand hardware platforms that combine energy efficiency with physically informed data processing. Photonics offers unique advantages in this context, but optical neuromorphic systems in which memory and information processing arise from intrinsic material dynamics remain scarce. We demonstrate a photonic layer in which history-dependent material photophysics implements physically embedded information processing before digital learning, using luminescent phosphors. The photonic layer exhibits dual fluorescence and phosphorescence together with excitation-history-dependent photoactivation dynamics that emulate synaptic functionalities, including short-term memory, long-term memory, and synaptic potentiation. Quantitative analysis and modelling reveal efficient nanoscale interlamellar energy transfer consistent with the lamellar material morphology, establishing a link between structure and function. When integrated as an active optical front-end within a hybrid photonic-digital AI architecture, the photonic layer performs a material-based transformation of input data before digital learning. Using a laboratory-based optical readout, classification accuracy comparable to state-of-the-art hybrid neuromorphic computing (approximately 94%) is achieved while requiring fewer training epochs. A mobile-compatible implementation based on smartphone optical readout yields similar accuracy, demonstrating robustness to non-specialized optical hardware. These results outline a general strategy for neuromorphic photonic systems in which functional materials act as adaptive computational primitives, enabling energy-efficient computing architectures.
Spiking neural networks (SNNs) are a promising, energy-efficient alternative to standard artificial neural networks (ANNs) and are particularly well-suited to spatio-temporal tasks such as keyword spotting and video classification. However, SNNs have a much lower arithmetic intensity than ANNs and are therefore not well-matched to standard accelerators like Graphical Processing Units and Tensor Processing Units (TPUs). Field Programmable Gate Arrays (FPGAs) are designed for such memory-bound workloads, and here we present a novel, system-on-chip design using a RISC-V softcore with a vector co-processor (FeNN-DMA), tailored to simulating SNNs on modern UltraScale+ FPGAs. We show that FeNN-DMA has comparable resource usage and energy requirements to state-of-the-art fixed-function SNN accelerators, yet it supports more complex neuron models and network topologies, and can simulate up to 16 thousand neurons and 256 million synapses per core. Using this functionality, we demonstrate state-of-the-art classification accuracy on the Spiking Heidelberg Digits, Neuromorphic MNIST and Braille tactile classification tasks.
