Stochastic computing is crucial for the accurate and efficient processing of massive data in advanced artificial intelligence applications. However, performing stochastic computing that balances generalization and efficiency remains a significant challenge. Here, we propose an in situ stochastic synapse (In-SSS) that can achieve in-memory stochastic computing with tunable probability from 0 to 1 within a single device. The device supports an innovative stochastic computing method, enabling in situ stochastic computing within the computation-in-memory architecture. Most importantly, it allows dynamic switching of stochasticity between the forward phase and weight update phase, effectively addressing the conundrum of balancing generalization and efficiency during stochastic computing. With these characteristics, the In-SSS is used to implement a novel stochastic neural network, achieving noticeable improvements in generalization ability and accuracy along with a respectable reduction in computation, compared with the fully connected neural network. The generality and robustness of the proposed method have also been demonstrated. This three-in-one device paradigm, integrating random number generation, weight memorization and efficient computation, provides a cutting-edge strategy for designing next-generation neuromorphic devices toward stochastic computing.

准确和高效的光信号处理对于早期癌症检测至关重要. 本研究开发了一种近红外手性有机突触光电二极管, 通过电调控实现双模式运行, 可同步完成圆偏振光(CPL)检测与神经形态处理. 在负偏压下, 器件高效分离电子与空穴, 实现精准的CPL探测与成像; 而在正偏压下, 界面载流子积累引发历史依赖响应, 从而支持对检测信息的实时神经形态处理. 该双模机制显著增强了对复杂光信号的识别与处理效率. 结合光学卷积神经网络, 器件进一步实现了对CPL信号的高效处理与复杂任务计算. 在癌症检测中, 该系统准确率达83

Brain-inspired neuromorphic computing holds the key to overcoming von Neumann bottlenecks and building intelligent computing systems due to its high energy efficiency and parallel processing capabilities. A significant challenge in the hardware implementation of neuromorphic computing is the separation of synaptic and neural components, which constrains integration density and energy efficiency. Although synaptic-neuronal reconfiguration has been achieved in single memristors or transistors, it relies on compliance current or configuration ports, which suffer from crosstalk, delays, bandwidth bottlenecks owing to electrical signal properties, and increased complexity from additional control modules. This study proposes, for the first time, a UV-regulated transparent memristor with reconfigurable synapse-neuron functions. Its light-controlled mechanism avoids electrical defects, eliminates crosstalk, and enables high-speed, low-power functional switching. The parallel transmission and spatial multiplexing characteristics of light not only achieve multi-channel co-regulation, improve integration density and bandwidth efficiency, but also integrate perception and computation. Furthermore, the transparency of the memristor confers greater advantages in stability and optoelectronic device fields. Leveraging its reconfigurable property, event-driven spiking neural networks (SNNs) were implemented via 1S-1N circuits and integrated with reinforcement learning algorithms to accomplish maze navigation (success rate >90%). This work provides a novel solution for the hardware implementation of neuromorphic computing networks.

Conventional ferroelectric optoelectronic devices suffer from low photocurrent and poor retention due to the physical separation between the ferroelectric and photoelectric conversion process. By incorporating PVN into the ferroelectric P(VDF-TrFE) layer, the composite not only provides a polarization field but also directly participates in light absorption and photoexcitation, significantly suppressing carrier recombination, fundamentally resolving the limitations of conventional ferroelectric photonic synapses. As a result, the fabricated transistor exhibits a large memory window of similar to 58 V and non-volatile retention exceeding 10(4) s. Benefiting from the nonlinear and history-dependent dynamics originating from the ferroelectric photocarrier interaction, an in-sensor reservoir computing system is implemented for UV-excited fingerprint recognition, achieving high training and testing accuracies of 99.5% and 98.5%, respectively. This work unveils a ferroelectric-photosensitive synergy mechanism, offering a viable pathway toward energy-efficient neuromorphic visual hardware.

Artificial synapses that combine transient temporal processing with persistent weight storage are essential for compact neuromorphic hardware, yet these functions are often implemented in separate device modules or coupled within the same stimulation pathway. Here, we report a ferroelectric memtransistive synapse with terminal-decoupled mixed plasticity, establishing a terminal-selective multi-timescale allocation strategy within a single memtransistor. Drain-source stimulation activates volatile synaptic modulation through ferroelectric polarization induced transient responses, providing reservoir-like dynamics for encoding temporal correlations and trajectory features. In contrast, gate-source stimulation enables nonvolatile weight updating through charge trapping, yielding programmable conductance states for feature retention and memory-based readout. Rather than merely demonstrating volatile/nonvolatile coexistence, this strategy assigns volatile dynamics and nonvolatile conductance states to cooperative computational roles, thereby integrating reservoir-readout computing paradigm in one device. The resulting weight-dependent dynamics enable robust perception of time-varying and degraded visual inputs. Mixed-plasticity operation improves dynamic trajectory recognition accuracy by 16% and 55% over volatile-only and nonvolatile-only modes, respectively, and enhances blurred-image recognition by 46% compared to just using non-volatile components. This work establishes terminal-selective plasticity allocation as a compact device-level strategy for unifying temporal computation, persistent memory, and robust multi-timescale perception in a single memtransistor.

Biologically inspired dynamic computing grants machines the capabilities of adaptive preprocessing and high-level recognition, which are crucial for perceiving diverse visual environments and making accurate decisions. However, unifying the full-spectrum dynamics required for dynamic computing within a single device remains a significant challenge. Here we present a single multi-dynamic neuronal-synaptic transistor (MDNST) by craftily customizing its electronic properties, enabling uniquely controlled interface and bulk dynamics by charged particles, including silver ions, oxygen vacancies and holes. The device highly unifies retinal sensitization and desensitization, and cortical neuron and synapse dynamics, enabling both dynamic processing and in-memory computing. We successfully demonstrate the application of the MDNST-based neuromorphic dynamic computing in two typical visual intelligence tasks, including adaptive collision detection and complex scene semantic segmentation. We show that the MDNST can significantly boost the perception and decision-making capabilities of the computing hardware, while notably improve the intersection-over-union and accuracy of semantic segmentation in complex visual environments. This versatile device paradigm that integrates dynamic processing, memory and computation paves the way toward machine vision comparable to biological vision.

Stable, secure, and high-bandwidth information transmission in complex environments is a key requirement for smart sensor networks and unmanned systems. However, traditional optical transmission has problems such as information degradation and encryption dependence on complex hardware. Although the emerging neuromorphic encryption can achieve compact and high-security dynamic encryption with a single device, it is still mainly based on electrical signals, which are limited by its bottlenecks in bandwidth and anti-interference. To address these challenges, an all-optical photo-induced circularly polarized luminescent synapse (PICPLS) is proposed. By combining cholesteric liquid crystals with phosphorescent materials, the physical encryption of circularly polarized light and the temporal encryption of synaptic characteristics are achieved without the need for additional hardware. Furthermore, a dual-mechanism synaptic polarization cryptosystem (DMSPC) was constructed based on this device to achieve all-optical encrypted transmission. Under interference conditions, the image recognition accuracy after DMSPC transmission fluctuates by less than 1%, while the image recognition accuracy of the incorrectly decrypted images drops to approximately 25%. Collectively, this system not only reduces the hardware requirements but also significantly enhances the stability and security of information transmission, thereby providing a highly integrated and reliable solution for more secure all-optical transmission in complex environments.

In the contemporary landscape of accelerating artificial intelligence (AI) development, multi-dimensional information recognition has emerged as a critical enabler for enhancing both data computational efficiency and decision-making precision. However, traditional multi-dimensional recognition architectures exhibit a fundamental reliance on extensive hardware arrays and complex circuit topologies, posing significant challenges to hardware integration and system-level miniaturization for AI-based recognition systems. Here, for the first time, we propose an in situ 4D neuromorphic transistor (I-FNT) and design a 4D spatiotemporal recognition system based on I-FNT. Through dynamic encoding of the input port voltages of I-FNT, programmable switching among three recognition modes (grayscale, depth, and time) is achieved, enabling cross-dimensional information perception. Compared to existing multi-dimensional information recognition systems, our 4D spatiotemporal recognition system significantly simplifies hardware while achieving 100% device integration gain. The I-FNT-integrated convolutional neural network (CNN) harnesses spatial (depth) information to achieve breakthrough performance in object recognition: 122% higher training efficiency and 345% faster training speed relative to conventional architectures, while attaining 94% accuracy. The system simultaneously facilitates object motion trajectory recognition, demonstrating comprehensive spatiotemporal processing capabilities. Therefore, I-FNT provides an efficient and accurate novel solution for multi-dimensional information recognition, representing a significant breakthrough for intelligent sensing and AI-based recognition systems.

Machine vision systems are essential for detection and recognition in autonomous driving and intelligent surveillance. Conventional silicon vision chips suffer from limited dynamic range, fixed photoresponsivity, high energy consumption, and degrading performance in high-contrast scenes. We developed a two-terminal memristive photomodulator with light adaptation. Its behavior arises from light-intensity-dependent transport transitions. Under weak light, trap-state de-occupation increases current; under strong light, space charge screening suppresses drift current. The device operates from 300 to 900 nm, covering the visible to near-infrared region. A 12 × 12 array achieved adaptive image enhancement and contrast regulation over a 135 dB dynamic range. Integrated with a convolutional neural network (CNN), it forms a hybrid neuromorphic visual system that improves recognition accuracy under extreme illumination from ∼45% to ∼97%.

The intrinsic chiroptical properties of organic semiconductors provide a powerful platform for polarization-encoded optoelectronic signal processing. Here, we develop self-powered chiral organic photodiodes (COPs) based on chiral non-fullerene acceptors blended with achiral polymer donors, in which the circularly polarized light (CPL)-dependent responsivity provides the physical foundation for tunable convolutional weighting. Owing to their polarization-dependent photocurrent, mathematically formulated as the product of the CPL amplitude and a sine function of the retarder rotation angle, these COPs act as bias-free, dynamically reconfigurable convolutional filters capable of robust feature extraction under noisy optical propagation conditions. This materials-driven modulation mechanism enables effective contrast enhancement and noise suppression without external bias or additional circuit elements. When incorporated into an optical convolutional framework, CPL-responsive COPs improve the structural similarity index measure of extracted feature maps from 0.15 to 0.80 and increase handwritten-digit classification accuracy from 76% to 87% compared with natural-light-based counterparts. These results establish chiral organic photodiodes as a promising materials platform for low-power, noise-tolerant optical information processing.image

ABSTRACT Bioinspired artificial perception systems (APS) are a crucial technical approach for future embodied intelligence, while mimicking perception with low power consumption, high integration, and matched transmission speed is the key element to achieve real‐time high‐efficiency APS. Recently, tribo‐transistors with a simple architecture have provided an innovative solution by directly converting mechanical stimuli into electrical signals, eliminating the need for external tactile sensors and thus improving integration and efficiency. In this review, we extend the discussion to opto–mechano–electric synergistic interactions, highlighting how triboelectric potentials can optically modulate synaptic transistors through field‐controlled charge transfer, exciton dynamics, and band engineering in low‐dimensional materials. We systematically summarize the working principles of triboelectric nanogenerators (TENGs) and tribo‐transistors, with a particular focus on photosensitive materials, optically transparent/active device architectures, and emerging scenarios such as optoelectronic neuromorphic vision systems and light‐triggered multisensory integration. Additionally, the current development prospects and challenges associated with tribo‐transistors APS are critically discussed, including new photonic functionalities such as triboelectric‐modulated photoluminescence and non‐Hermitian photonic sensing. This review aims to provide strategic insights into designing next‐generation APS and to establish a solid foundation for their future deployment in robotics and other advanced intelligent systems.

Neuromorphic hardware integrates computing and memory with parallelism and ultralow power, offering a promising solution to overcome the von Neumann bottleneck. Artificial synapses are critical components; however, directly observing microscopic electron distribution gradients and their dynamic evolution remains challenging, hindering charge storage understanding and device optimization. This study employed quantum dot (QD) floating-gate synaptic transistors as a model, leveraging high-throughput, high-precision simulations to systematically explore the mutual effects of QD size, QDs' concentration gradients, and carrier concentration fields. The simulations revealed dynamic evolution of charge trapping behavior, transitioning from shallow trap coupling collapse to network-like migration and deep-level localized trapping as the device genome shifted. Simulation-based screening identified high-performance device genome modules, enabling the design of a bilayer different scales QD array device with a specific "gene combination". This device achieved 96.95% MNIST recognition accuracy, a 4.72% improvement over monolayer structures, with optimized charge-trapping depth and retention. These findings demonstrate that high-throughput, high-precision device genome simulations provide an efficient pathway for developing high-performance neuromorphic hardware and systems by establishing gene-performance mapping models.

Intelligent three-dimensional displays require both powerful computing capabilities and high display performance, yet these metrics have long been constrained by the inherent trade-off between efficiency and image quality. Here we introduce circularly polarized bulk-heterojunction memory-computing organic light-emitting diodes featuring an integrated memory-processing-display architecture to break this long-standing limitation. The bulk-heterojunction memory-computing organic light-emitting diode incorporates a chiral bulk-heterojunction active layer formed by an achiral emissive polymer and chiral small molecules. Through precise control of the blend composition, we co-engineer supramolecular ordering and carrier dynamics, enabling the in situ modulation of synaptic weights and multilevel conductance states. This effectively transforms organic light-emitting diodes from passive light emitters into an integrated memory-computing-display system with enhanced computing efficiency and polarized emission. The optimized device achieves a peak luminance of 22,306 cd m−2, a high circularly polarized electroluminescence dissymmetry factor of 0.92, together with ultralow energy consumption of 1.78 pJ per spike. Hardware validation using a 64 × 64 device array demonstrates uniform and distinguishable 5-bit conductance distributions, with an energy consumption of 0.046 nJ per device and 17.472 nJ per pixel per computation. We further integrate our bulk-heterojunction memory-computing organic light-emitting diodes into a neural radiance field framework, enabling stereoscopic greyscale three-dimensional scene reconstruction on 100 × 180 arrays. This memory-computing-display integration strategy provides a promising pathway towards intelligent display technologies. Circularly polarized memory-computing organic light-emitting diodes integrate three-dimensional displays with memory and processing capabilities. Optimized devices achieve luminance of over 22,000 cd m−2, dissymmetry factor of 0.92 and energy consumption down to 1.78 pJ per spike, enabling neural radiance field greyscale three-dimensional scene reconstruction.

Reconfigurable volatile/nonvolatile neuromorphic devices integrate brain-like rapid learning with stable memory, providing a critical pathway toward edge-intelligent systems. The integration of reconfigurable devices and light-emitting functionality transforms the display from a passive output terminal into an integral hub of the intelligent system. However, integrating reconfigurable volatile/nonvolatile memory and light-emitting functionality within a single device remains a major challenge because of the inherent conflicts among multiple functions and the constraints of conventional two-terminal control. Here, for the first time, we present a three-terminal reconfigurable volatile/nonvolatile light-emitting memristor. The three-terminal configuration provides control of the emission region and memory states, enabling seamless transition between volatile, nonvolatile, and light-emission operations. By enabling spatial reconfigurability and temporal multiplexing of emission, the device improves scalability and simplifies system architecture and control, achieving a 69.23% reduction in data-transfer volume and a 40% reduction in system component count compared with separated architectures. In addition, it offers direct visual feedback while facilitating continuous learning in neuromorphic computing systems. Finally, we implement an anomaly-detection visualization system based on the reconfigurable volatile/nonvolatile light-emitting memristor, demonstrating its ability to produce direct visual output while processing information, thus enabling an integrated memory-compute-display architecture for next-generation edge-intelligent display systems.

Real-time edge perception is increasingly constrained by conventional frame-based imaging and von Neumann processing, in which the continuous acquisition of dense image streams incurs excessive bandwidth demands, memory traffic, and energy consumption, ultimately limiting response speed and scalability. Neuromorphic vision offers a compelling alternative by suppressing redundancy and encoding salient visual dynamics into sparse spikes. However, most reported vision neuron systems require peripheral signal conditioning and analog-to-digital conversion, thereby introducing substantial latency and power overhead. Here, we report a reconfigurable multi-mode artificial visual neuron based on Ta2O5/PDVT-10-CsPbBr3 quantum dots that integrates retina-inspired and cortex-inspired spiking behaviors within a single optoelectronic device. The neuron supports bias-free optical Leaky Integrate-and-Fire spiking, remaining effectively quiescent under static scenes while directly converting light-intensity variations into single-spike coding at the source, enabling rapid, energy-efficient event-based representation. In addition, its electrically accessible dynamics provide further programmability for cortex-like processing and learning. The system demonstrates high recognition accuracy, highlighting the potential of Ta2O5-based artificial visual neurons for compact, low-latency, always-on machine vision in edge intelligence.

Bayesian neural networks (BNNs) enable trustworthy edge intelligence by quantifying predictive uncertainty. However, hardware BNN implementations face a bottleneck: digital approaches suffer from high latency, while memristors are limited by the intrinsic coupling between their conductance state (mean) and stochastic noise (variance). Here we report a coupled dual-channel memristor (CDCM) based on an ion gel/ZnO heterostructure that breaks this fundamental trade-off. Utilizing vertical ion-gating to establish two tunable memristive channels, our device defines synaptic weight as the differential conductance between the two channels. This architecture enables the hardware-native orthogonal control over the synaptic weight mean (μ) and standard deviation (σ), allowing for the precise synthesis of decoupled Gaussian weights. We validate our approach with a hardware-calibrated BNN for multimodal human activity recognition, achieving 79.08% accuracy while reliably detecting unseen activities as out-of-distribution anomalies. This work provides a scalable, physics-driven paradigm for energy-efficient, inherently trustworthy probabilistic computing.