The exploration of marine environments is crucial, yet the extreme conditions of the deep-sea, combined with the segregated signal processing in current sensor technologies, lead to bulky systems, high energy consumption, and significant latency, which severely constrains the development of real-time intelligent perception systems underwater. Herein, we developed a neuromorphic floating-gate transistor (NFT) that integrates both electrical and optical memory functionalities, emulating simultaneously visual and auditory synaptic behaviors within a single unit, thus enabling in-memory dual-mode processing of visual-auditory signals. Electrically, it achieves rapid switching (∼14 µs), high on/off ratio (106), and robust endurance (>104 cycles). This enables high-accuracy (88%) classification of seafloor minerals and rocks via sonar echo processing using a convolutional neural network (CNN). Optically, the NFT exhibits tunable synaptic weight modulation from short-term to long-term plasticity under 405-808nm laser pulses. Leveraging the low-attenuation green-light window in seawater, the system, combined with RGB denoising and green-channel enhancement preprocessing, realizes 80% accuracy in marine biological image recognition. This synergistic electro-optical in-memory computing architecture provides an efficient, low-power, and compact hardware solution for intelligent perception in complex underwater environments.

Spiking Neural Networks (SNNs) have emerged as a promising paradigm for brain-inspired edge computing. Leveraging binary spikes and local learning rules, SNNs enable energy-efficient on-chip learning and rapid adaptation to changing environments, which is crucial for edge AI that needs to learn continuously from new data. However, many SNN processors enabling on-chip learning for edge computing confront a trade-off: small-scale task-specific designs offer low power but poor multi-task inference accuracy, while large-scale general-purpose designs achieve high multi-task accuracy at the cost of large memory and poor energy efficiency. To overcome this challenge, this paper presents ANP-R, a 22nm asynchronous SNN-based edge AI processor with coarse-grained reconfigurable architecture enabling one-shot, few-shot, batch and incremental on-chip learning. The processor integrates 64 cores containing 4096 neurons and 0.262 million synapses. Two key features are proposed: 1) An asynchronous coarse-grained reconfigurable architecture that supports various STDP-based SNN topologies. These topologies enable over 95% average accuracy across four sensory tasks; 2) an energy-efficient asynchronous training method incorporating a self-adaptive synaptic weight update mechanism reducing up to 65% redundant updates without accuracy loss, and a trained weights low-bit width coding method reducing up to 50% storage cost with 0.3% accuracy loss. Measurement results demonstrate 92.1% accuracy for hand gesture classification, 93.9% for keyword spotting, 98.6% for object recognition and 99.2% for gas identification. Compared with state-of-the-art SNN-based chips, this work achieves up to 6.02x, 8.61x and 7.1% improvement in energy efficiency, energy per step, and accuracy, respectively.

Driven by the rapid progress of artificial intelligence and robotics, neuromorphic vision systems are gaining significant attention for enabling efficient visual information processing in complex and dynamic environments. In particular, optoelectronic devices that emulate the functionality of the biological retina are essential for achieving efficient neuromorphic visual processing. Here, we report an optoelectronic synaptic memtransistor (OSMT)-based neuromorphic vision system for image processing applications. By integrating photoresponsive indium-gallium-zinc-oxide (IGZO) as the channel material with a hafnium oxide (HfO2) contact-engineered architecture, the OSMT exhibits optically and electrically tunable resistive states, enabling stable and controllable synaptic weight modulation for artificial neural network (ANN) implementation. Benefiting from reliable optoelectronic synaptic characteristics, ANN simulations achieve a handwritten digit recognition accuracy of 92.17%. Furthermore, a 6×6 OSMT array demonstrates neuromorphic image processing capabilities, including contrast enhancement. These results highlight the potential of OSMTs as key building blocks for intelligent machine vision systems, offering new opportunities for advanced robotic platforms and human-machine interfaces.

Energy-efficient and adaptive neuromorphic hardware requires material platforms that can intrinsically integrate transient neural dynamics with stable long-term memory within a single device architecture. Here, a cross-point nanoporous SiO2 memristor that unifies volatile and nonvolatile switching behaviors within a single material platform is reported. The engineered nanoporous framework provides well-defined ion migration pathways, enabling controlled modulation of conductive filaments and reversible transitions between short-term plasticity (STP) and long-term plasticity (LTP) through simple compliance-current tuning. Leveraging this dual-mode functionality, the volatile dynamics of the nanoporous SiO2 memristors are employed directly as a physical reservoir, while the nonvolatile conductance states serve as synaptic weights in the readout layer. Using a conductance-aware training scheme, reservoir computing (RC) is demonstrated on the same device platform, achieving 93.7% accuracy in MNIST handwritten-digit recognition. Beyond standard benchmark datasets, the system further enables ECG temporal biosignal classification, reaching over 88% accuracy in distinguishing normal and abnormal heartbeat patterns. These results establish a single-material, CMOS-compatible neuromorphic platform capable of integrating dynamic processing with persistent memory, offering a scalable and low-power pathway toward compact intelligent edge computing hardware.

To meet the urgent demand for high-performance, low-power hardware devices for near-infrared polarization-sensitive neuromorphic vision in infrared detection, robot perception, and other fields, this work develops a two-dimensional heterojunction-based visual synapse with efficient near-infrared polarization response and flexible regulation. Benefiting from the high carrier mobility of NbSe2 and the strong in-plane optical anisotropy as well as the narrow direct bandgap of 0.36 eV in Ta2NiS5, the two materials synergistically form an atomically sharp Schottky barrier at the NbSe2/Ta2NiS5 interface. This barrier effectively promotes the separation of photogenerated electron-hole pairs and directional interfacial charge transport, endowing the heterojunction with superior photoelectric response properties. Accordingly, the heterojunction exhibits stable polarization-dependent synaptic responses across the 1064-2200 nm near-infrared window, whereby photoinduced synaptic weight can be precisely modulated by three independent parameters: incident optical power, light pulse duration, and applied bias voltage. Benefiting from these features, the heterojunction successfully emulates critical neuroplasticity behaviors, including paired-pulse facilitation, paired-pulse depression, the activity-dependent transition from short-term plasticity to long-term plasticity, and Hebbian learning. At the representative wavelength of 1550 nm, the heterojunction delivers an outstanding optical anisotropy ratio of 16.3. This work provides a promising hardware platform for next-generation infrared polarization vision and adaptive robotic perception systems, exhibiting important theoretical significance and practical application value.

RationaleAlcohol use disorder is defined by drinking alcohol despite knowledge of negative consequences, often referred to as aversion-resistant drinking (ARD). The dorsomedial (DMS) and dorsolateral striatum (DLS) are necessary for goal-directed and habitual action selection, respectively. Leading hypotheses posit that once drug use becomes compulsive, DMS dependence degrades while DLS dependence increases. This shift may be mediated by changes in synaptic weights from glutamatergic inputs.ObjectivesUsing a combination of western-blot, micro-injections, and ex-vivo electrophysiology, we investigated the role of α-Amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptors AMPAR, which drive glutamatergic transmission, during quinine-adulterated alcohol (QuA) drinking in the DMS and DLS across the development of ARD.ResultsWe found that AMPAR subunit composition and function change in the DMS across the development of ARD whereby, calcium permeable (CP) - AMPARs drive behavior. Western blots revealed a negative relationship between DMS GluA1 and QuA drinking in aversion-sensitive mice and positive relationships between DMS or DLS GluA1/A2 ratios and QuA drinking in ARD mice. DMS CP-AMPAR antagonism caused an increase in QuA drinking suggesting that CP-AMPARs in the DMS prevent ARD. Ex-vivo electrophysiology of DMS spiny projection neurons (SPNs) revealed that ARD mice had a greater rectification index than aversion-sensitive mice indicating that SPNs in the DMS express more CP-AMPARs following the development of ARD.ConclusionsThese data provide evidence that repeated alcohol binges alter DMS CP-AMPAR activity, where initial DMS activity acts to prevent ARD but after repeated binges that result in ARD, DMS SPNs recruit CP-AMPARs.

Current memristive circuits for biological decision-making only consider single-dimensional stimulus feedback and synaptic weight regulation without considering the influence of multiple factors interacting. This article designs a biological multibrain collaborative Fogg's behavior model (FBM) neural network circuit based on memristors and considers the influence of emotions on decision-making. The circuit includes the thalamus, dorsolateral prefrontal cortex, ventral tegmental area, hippocampus, amygdala, striatum system, and anterior cingulate cortex module. The circuit adopts a brain-like partitioned modular design, which provides a structural basis for its functional implementation, and it also has the advantages of low power consumption and high integration. The circuit can accurately simulate neural synaptic plasticity and realize learning reinforcement, natural forgetting, and dynamic modulation of synaptic strength under emotional regulation. PSpice simulations show that the integration of more brain region coordination mechanisms exhibits adaptive decision-making characteristics that are closer to the biological brain.

HLSP-LSM: Enhancing image recognition performance of liquid state machines via brain-inspired hybrid long short-term plasticity.

Optogenetics employs light to regulate neuronal activity with exceptional spatiotemporal precision, thereby enabling the direct modulation of learning and memory processes in the human brain. This capability to externally control neuronal signaling via optical stimuli has provided profound insights into brain function and established a versatile strategy for engineering bio-inspired information processing. Herein, a designable van der Waals (vdW) crystal has been demonstrated for device-scale neuronal cell mimicking. The structural similarity between ion-channel in biological membranes and layered vdW lattices is realized with nano-crystallization. As the sequential transition from carrier transport dominance to ion transport dominance reveals the dynamic control over synaptic weight updates. Optoelectric synaptic plasticity in designable vdW crystal (long-term potentiation and depression, paired-pulse facilitation, and a tunable short-term to long-term memory transition) were conclusively observed and correlates with photo-induced carrier trapping and ionic migration. Furthermore, learning-forgetting-relearning cycles achieve 34.7% increased retention efficiency compared to bulk ReSe2. Functional demonstrations in edge detection and CIFAR-10 image recognition confirm the synaptic plasticity into system-level neuromorphic computation, with a recognition accuracy of 96.24%. In conclusion, we envision that designable vdW artificial crystal will provide the versatile advances for 3D stackable neuromorphic vision architectures.

Spiking Neural Networks (SNNs) trained with Spike-Timing-Dependent Plasticity (STDP) are capable of continuous, online learning but critically suffer from unstable network dynamics. Homeostatic mechanisms such as divisive weight normalization improve stability to produce robust SNN models, however current normalization techniques are challenging to implement in neuromorphic architectures due to their computational complexity and memory access requirements. This work presents a novel, linear, self-normalizing STDP rule that drives convergence of each weight vector towards a target norm in an iterative and event-driven manner with O ( 1 ) space complexity. We achieve a mean normalization error of 1.5% for a target average weight of 0.10 on the MNIST and Fashion-MNIST datasets and improve training times on average by 37% compared to models trained with existing normalized STDP rules. The proposed approach leverages multiplicative weight dependence with a single scaling weight maximum term based on the L 0 norm of positive weight changes at each post-synaptic spike event. We also present a variant of the Self-Normalizing rule that learns both positive and negative weights while maintaining a mean of zero. This approach demonstrates impressive noise tolerance on inference tasks and opens avenues of development into the online training of kernels for convolutional networks. By incorporating the process of normalization into the event-driven STDP updates, the proposed Self-Normalizing rule offers a stable and efficient solution for neuromorphic applications where local STDP learning is supported without the need for computationally expensive weight normalization operations. • Synaptic weight normalization stabilizes spiking neural network dynamics. • Novel Self-Normalizing STDP rule steers synaptic weights to converge to a target mean. • Self-normalizing weight updates are local, linear, stable, and event-driven. • Weight mean of zero improves noise tolerance at inference. • Self-normalizing updates are more computationally efficient than existing techniques.

Epilepsy is a prevalent neurological disorder characterized by recurrent seizures, and electroencephalogram (EEG) signals provide a direct measure of brain activity for detection. Although deep learning achieves high accuracy, it often lacks physiological interpretability. We propose the Excitation/Inhibition Dynamic Polynomial Network (E/I-DynPolyNet), a biologically grounded framework for interpretable seizure detection. Specifically, E/I-DynPolyNet introduces a dual excitatory/inhibitory (E/I) pathway with sign-constrained synaptic weights, encouraging the learned activations to reflect latent E/I representations. Furthermore, a differentiable Wilson-Cowan (WC) module is embedded to govern the temporal evolution of E/I interactions, ensuring consistency with neurophysiological principles. A physics-informed optimization strategy integrates supervised learning with dynamical residual constraints and E/I balance regularization, guiding the model to learn physiologically consistent representations. Experimental results on the CHB-MIT and Bonn datasets demonstrate competitive accuracies of 95.81% and 98.5%, respectively. Crucially, E/I-DynPolyNet enables quantitative estimation of E/I imbalance, revealing that E/I ratios increase from 1.01 in the pre-ictal phase to 1.38 during seizures—a finding consistent with clinical observations of ictogenesis. These results indicate that E/I-DynPolyNet not only improves detection performance but also provides a mechanistic description of seizure dynamics, bridging the gap between data-driven learning and neurophysiological interpretation.

Emerging ferroelectric non-volatile memories are revolutionizing von Neumann architectures by providing efficient hardware for both AI training and inference. However, as ferroelectric dimensions scale toward the nanoscale, reliable modulation is hindered by interfacial degradation and phase instability, leading to synaptic weight drift and computational inaccuracies. Here, a high-performance ferroelectric-van der Waals transistor (FeFET) for computing-in-memory by integrating a single-crystalline Bi2O2Se (BOS) layer into a ferroelectric/MoS2 heterostructure is demonstrated. The implementation of an asymmetrical capacitive stack ensures effective polarization-charge compensation during fine-state switching, achieving precise multi-level weight programming with significantly suppressed carrier fluctuations. Fabricated through a low-temperature process, the BOS-based FeFET exhibits exceptional reliability, including 10-year retention at 85°C, endurance exceeding 1011 cycles, stable 32-state analog switching with 0.9% retention variation over 10000s, and ultra-low programming error. Atomically smooth heterointerfaces yield high spatial uniformity (7% variation) across the FeFET array, enabling a hardware neural network that achieves 98.5% accuracy in nonlinear classification. Furthermore, by incorporating intrinsic ferroelectric switching variance into the training phase, it is elucidated how device imperfections can be leveraged to reshape learning dynamics in pixel-wise semantic segmentation. This work establishes a comprehensive co-design methodology bridging advanced ferroelectric materials, device engineering, and algorithmic optimization for next-generation neuromorphic computing.

Previous studies tracking the relationship between manipulations ofC. elegansneurons and the resulting behavioral changes have called for the development of a connectome-constrained neural network model that describes the cascade from neurons to behavior. However, the model using anatomical connectome weights directly did not achieve that. Here, we introduce a framework that jointly optimizes synaptic weights with respect to the relative proportions of anatomical synaptic weights, muscle activity patterns required for locomotion, and the phase relationship between SMD neuron activity and neck-bending angle. As a result, our neural network model generates plausibleC. elegansbehavior mediated by activity changes in forward and backward command-neurons, even without the introduction of pacemaker neurons with intrinsic oscillatory activity. Additionally, we identified necessary neurons for maintaining oscillatory patterns on muscular activity that could serve as clues for the central pattern generator in our neural network model. Finally, we provide 10 optimized synaptic weight sets of C. elegans that reproduce the results of manipulation experiments on the SMD neurons. This study will facilitate future studies for unraveling the multiscale relationship of "from synapse to behavior" in the nervous system.

Magnetic skyrmions, as topologically protected spin textures, hold great potential for energy-efficient neuromorphic systems. While artificial synapses have been demonstrated in magnetic multilayers by electrically controlling skyrmion populations, their probabilistic nucleation severely limits reliability. The recent emergence of 2D van der Waals magnets, with their inherent tunability and novel spintronic phenomena, offers a promising platform to overcome these challenges. Here, we demonstrate an artificial synaptic device in 2D ferromagnet Fe3GaTe2, operating on the fundamentally different principle of a deterministic and collective spin texture transformation from a skyrmion-lattice to a stripe-domain state. This transformation yields a linear, reproducible modulation of the anomalous Hall resistance. The slope of this linear response, defined as the synaptic weight, is effectively tuned by varying the pulse width, thereby enabling multi-weight functionality and multiply-accumulate operations. Projected scaling of the device reduces the single-operation energy consumption to 0.66 pJ, a level comparable to state-of-the-art memristor technologies (e.g., resistive random-access memory and phase-change memory). Furthermore, a hardware-informed quantized neural network based on this synapse achieves a high recognition accuracy (∼96.1%) in handwritten-digit recognition. Our findings establish a robust pathway for creating large-scale and energy-efficient neuromorphic systems based on the collective dynamics of Fe3GaTe2 spin textures at room temperature.

Memristor-based reconfigurable architecture for binarized neural networks: Implementation and robustness analysis.

There is room on the inside. We present an account of neuroplasticity with respect to cell-internal processing pathways and their relation to membrane and synaptic plasticity. We think traditional synapse-centric, weight-based models of memorization are not sufficient or adequate to capture the complexity of neuroplasticity. In standard accounts, we model a network of neurons connected by adaptive transmission links. The adaptation of these transmission links is overly simplified using short-term and long-term potentiation/depression, assuming weight changes according to use of the transmission link. In contrast, we propose a paradigm switch from a synapse-centric model (each synapse learns independently, based on its history of use) to a neuron-centric model (each neuron uses signal selection for intracellular pathways to express plasticity at the membrane). Each neuron has a ‘vertical’ dimension where internal parameters steer the external membrane- and synapse-expressed parameters. A neural model consists of (a) expression of parameters at the membrane, in particular dendritic synapses or spines, and axonal boutons (b) internal parameters in the sub-membrane zone and the cytoplasm with its protein signaling network and (c) core parameters in the nucleus for genetic and epigenetic information. In a neuron-centric model, each node (=neuron) in the horizontal network has its own internal memory. Neural transmission and information storage are separated, not automatically combined by coupling strength. There is filtering and selection of signals for storage. Not every transmission event leaves a trace. This represents an important conceptual advance over synaptic weight models. We present the neuron as a self-programming device, rather than as passively determined by ongoing input. We believe a new approach to neural modeling is necessary, because the experimental evidence is not well captured by traditional synapse-centric models. Ultimately, we are interested in the possibilities of a flexible memory system that processes external signals according to its inherent structure.

Achieving precise and scalable control of neuronal connectivity is a key requirement for next-generation neurotherapeutic and neuroengineering applications. Yet, implementing Hebbian-like plasticity rules and verifying induced changes at a larger scale remain technically challenging. Here, high-density microelectrode arrays (HD-MEAs) are combined with programmable stimulation and analytics to induce and confirm targeted connectivity changes across multiple different in vitro and ex vivo preparations. Conditional Activity Metrics (CAM) are introduced to quantify stimulation-evoked changes in relative timing and spiking density between spike train pairs. CAM changes strongly correlated with synaptic weight changes in simulations. Experimentally, we observed robust synaptic strengthening/weakening in roughly 40% of 279 tested pairs following targeted stimulation. In a subset of pairs, tracked for extended durations, the induced effects persisted for at least 90min. Synaptic modifications were also directly validated by using simultaneous HD-MEA and patch-clamp recordings. This work establishes a foundation for precise, high-throughput neuronal circuit reconfiguration, offering a versatile platform for advancing fundamental neuroscience and enabling novel neurotherapeutic and biohybrid computingstrategies.

In the era of the Internet of Things and edge intelligence, conventional always-on machine vision systems suffer from severe energy bottlenecks because they continuously process massively redundant spatiotemporal data. Inspired by the dual-pathway strategy of the human retina, we propose a bioinspired "Sentinel-Expert" synergetic vision system enabled by polarization-reconfigurable organic ferroelectric phototransistors based on P(VDF-TrFE). By manipulating the ferroelectric polarization states, a single device is reconfigured into three functional modes: (i) a magnocellular pathway-inspired Sentinel mode under negative polarization that exhibits short-term plasticity and fading-memory dynamics for physical reservoir computing and ultralow-power motion event detection; (ii) a parvocellular pathway-inspired Expert mode under positive polarization that provides long-term potentiation-like retention for in-sensor contrast enhancement and high-fidelity static recognition; and (iii) a Programming mode that serves as a unified hardware backend for synaptic weight updates. The system achieves 98.46% recognition accuracy in the Sentinel mode and 97.45% accuracy in the Expert mode. Notably, by exploiting the high spatiotemporal sparsity of valid events in real-world scenarios to activate the Expert mode only upon wake-up events, this event-driven strategy reduces the computational cost by ∼50× at a 1% duty cycle compared with the conventional always-on strategy.

Autonomous driving perception demands low latency, high temporal resolution, and stringent hardware efficiency. While event-based spiking neural networks (SNNs) offer bio-inspired sparse computation, their deployment on edge field-programmable gate arrays (FPGAs) is obstructed by irregular execution patterns and temporal state storage overhead. To address this, we propose HAPQ, a unified hardware-aware pruning and quantization pipeline for compact event-based object detection. Starting from an end-to-end adaptive sampling SNN detector (EAS-SNN), HAPQ conducts hardware-aware configuration search within discrete digital signal processor (DSP) and block RAM (BRAM) budgets, applies single-instruction-multiple-data (SIMD)-aligned structured pruning for computational regularity, and jointly quantizes synaptic weights and membrane potentials via a shift-friendly fixed-point recurrence. Evaluation on the Prophesee Gen1 dataset and an FPGA accelerator shows that HAPQ improves detection accuracy from 0.284 to 0.425 in mean average precision () and achieves 0.722 . Hardware implementation reveals a reduction in lookup table (LUT) usage to 1680, complete DSP elimination, and a maximum operating frequency of 920.81 MHz at 0.630 W. These results confirm that effective temporal SNN deployment requires joint optimization of model architecture, state precision, and hardware-aligned workload organization.

The memtransistor constructed using the emerging two-dimensional tellurene material has demonstrated significant potential for application in artificial synaptic devices and image recognition. However, conventional device fabrication processes, such as dry transfer, restrict tellurene's further application in flexibility and wearable electronics. Here, a fully printed, flexible tellurene field-effect transistor (FET) is demonstrated, consisting of a channel, gate dielectric layer, and contact electrodes, all of which are prepared using functional inks that include tellurene, h-BN, and graphene, respectively. Such a device integrates scalable ink formulations and neuromorphic functionality for advanced electronics, exhibiting relatively stable electrical performance even after 10,000 bending cycles at a curvature radius of 11.05 mm. Applying electric stimulation to the h-BN layer enables the realization of a bioinspired memtransistor, achieving paired-pulse facilitation, reconfigurable short-term plasticity to long-term plasticity transitions, and synaptic weight updates. Moreover, image recognition simulation using an artificial neural network achieves 93.91% accuracy on the Modified National Institute of Standards and Technology database, which can maintain 78.33% accuracy after introducing σ = 0.7 Gaussian noise. These results position printed tellurene FETs as promising, noise-resilient building blocks for scalable, flexible neuromorphic systems.