While reinforcement learning algorithms have made significant progress in solving multi-armed bandit problems, they often lack biological plausibility in architecture and dynamics. Here, we propose a bio-inspired neural model based on interacting populations of rate neurons, drawing inspiration from the orbitofrontal cortex and anterior cingulate cortex. Our model reports robust performance across various stochastic bandit problems, matching the effectiveness of standard algorithms such as Thompson Sampling and UCB. Notably, the model exhibits adaptive behavior: employing greedy strategies in low-uncertainty situations while increasing exploratory behavior as uncertainty rises. Through evolutionary optimization, the model's hyperparameters converged to values that align with the principles of synaptic mechanisms, particularly in terms of synapse-dependent neural activity and learning rate adaptation. These findings suggest that biologically-inspired computational architectures can achieve competitive performance while providing insights into neural mechanisms of decision-making under uncertainty.

One of the most important issues in the field of computer organization and architecture is performance evaluation. This is because the ultimate goal for which any computing system is developed is to execute programs efficiently and within the shortest possible time. All architectural improvements made to the system, whether in the processor, instruction execution time, memory hierarchy, or parallel processing capability, have to be evaluated for their effectiveness in reducing the time for program execution. Since computing systems have to function under varying working conditions and usage environment, performance cannot be defined as a single value. It has to be defined and measured in relation to the working condition. The current research paper aims to present a detailed discussion on the definition and measurement of performance in the field of computer architecture. The concept of execution time and the difference between elapsed time and CPU time have been discussed. The analysis of the CPU performance equation and its components have also been included. The paper also aims to discuss the issue of misleading performance measures, the consequences of Amdahl’s law, the importance and limitations of benchmarking, and performance issues in modern computing systems such as graphics processors, cache hierarchy, pipelining, and multicore processors.

To support the increasing complexity of innovation, design, and performance evaluation in the maritime industry, a ship-specific computational fluid dynamics (CFD) software suite tailored to incompressible viscous flow is required. This study utilizes the MarineFlow marine fluid dynamics code to explore numerical simulation schemes for cylindrical flow problems across a broad range of Reynolds numbers (1–107) that are applicable to self-developed codes. Additionally, an analysis of the flow around a cylinder is conducted from the perspective of code developers. Various grid types and turbulence model schemes are employed to analyze and compare the drag coefficient, separation points, and pressure distribution characteristics of the cylinder. The results obtained from these simulations are then contrasted with those derived from commercial CFD software to assess their accuracy. Despite the presence of certain numerical artifacts, within the Reynolds number range of 1–105, the unstructured grids combined with the laminar flow models effectively capture experimental data. Further exploration of the transitional Reynolds number range (Re = 2×105–6×105) shows a consistent decreasing trend in the mean drag coefficient, although significant deviations from theoretical predictions are evident. From the perspective of code developers, this study aims to reveal the limitations of current computational schemes and code architecture in accurately capturing flow dynamics within the transitional Reynolds number range. This provides a crucial basis for future optimization of turbulence models and algorithmic improvements, which are essential for the continued development of self-developed CFD codes and their engineering applications.

In the context of increasingly complex data-driven decision-making, traditional analytical methods face significant challenges in handling heterogeneous data correlations and real-time computational responsiveness. By developing an Integrated Computational Business Analytics Framework (ICBAF), this research proposes a three-layer computational architecture comprising heterogeneous data representation, heuristic computational optimization, and online feedback loops, aiming to achieve a deep coupling of business logic and algorithmic reasoning. The findings indicate that by leveraging the topological representation capabilities of Graph Neural Networks (GNN) and the dynamic optimization mechanisms of real-time stream computing, this framework effectively resolves bottlenecks such as "data silos" and "decision lag," significantly enhancing decision precision in scenarios like intelligent marketing and smart operations. However, as the study primarily focuses on theoretical deduction and scenario-based simulations, it lacks rigorous empirical validation using large-scale, real-world industrial datasets. Consequently, its generalization capabilities under extreme noise and the cost-benefit ratio of computational resources require further verification in future research.

Modern processors experience memory contention when the speed of their computational units exceeds the rate at which new data is available to be processed. This phenomenon is well known as the memory wall and is a great challenge in computer engineering. The reason for this phenomenon is the unequal growth rate in memory access speeds compared to processor clock rates. To mitigate the memory bottleneck in classic computer architectures, a scalable parallel computing platform—the Grid of Processing Cells (GPC)—has been proposed. To evaluate its effectiveness, the GPC is modeled at the instruction level and functional level using SystemC TLM-2.0, with a focus on memory contention. Individual GPC cells can be switched between the two abstraction levels. Our mixed-level system model enables fast and accurate simulations. We test multiple streaming applications on the GPC, and analyze software-based optimization methods and their effects on the GPC, at both abstraction levels. The performance is then compared against the traditional shared memory processor architecture. Experimental results show improved execution times on the GPC primarily due to a large decrease in main memory contention.

Hebbian Physics Networks: A self-organizing computational architecture based on local physical laws

Rapid advancements in artificial intelligence have magnified the inherent bottlenecks and energy inefficiencies of conventional von Neumann architecture. To address these limitations, processing information in a highly parallel, memory-integrated manner mimicking the human brain, neuromorphic devices have emerged as a cornerstone of next-generation computing. Among these, optical-neuromorphic devices are particularly promising. By using light, they offer transformative advantages, such as high speed, massive bandwidth, and minimal signal interference. Accordingly, we propose long-persistent luminescence (LPL) materials as novel substrates for optically operative artificial synapses. We utilize AGa2O4 (A = Mg, Ca, Sr, or Ba) luminescent oxides in which the intrinsic defect states enable excellent LPL properties without complex material engineering. Leveraging these properties, we demonstrate the physical implementation of optical material-based neural processing, in which memory retention and nonlinear transformation are executed within the LPL material. As proof of concept, this physical neural network was applied to a real-time Pong gameplay, demonstrating autonomous decision-making through light-driven signal processing. To further extend, we developed physical reservoir computing and physical neural network architectures. These architectures exploit the nonlinear temporal dynamics and luminescence mapping of LPL to perform handwritten digit recognition. Our findings establish LPL materials as a versatile platform for developing next-generation, energy-efficient optical-neuromorphic systems.

Recent debates in posthumanism, new materialism, and artificial intelligence ethics have begun to question the centrality of the human as the privileged locus of agency and subjectivity, yet they remain tethered to an anthropocentric horizon. The very category of the non-human defines emerging agents—animals, artefacts, infrastructures—by negation or derivation from a historically specific figure of the human. This article argues that such a framework is epistemologically insufficient for thinking what is increasingly at stake: the possibility of a machine civilisation in itself, grounded not in human categories but in the internal logics of computational architectures, networks, and protocols. Drawing upon axiology, the philosophy of technology, and computational aesthetics, the article develops a system of six axioms to conceptualise machine civilisation as non-derivative, architecturally conditioned, historically self-articulating, normatively technical, plural, and human-decentred. A dedicated methodological section articulates the regulative-transcendental approach that underpins the axiomatic framework, distinguishing it from both speculative metaphysics and empirical prediction. The article further advances the concept of computational phenomenology—an analysis of machine temporality, spatiality, affect, and identity—and proposes an allo-species lexicon adequate to a post-anthropocentric intellectual landscape.

The bioelectronics were used for the brain-machine interface along with neural engineering. These next-generation devices can be used for neuroprosthesis, which should be energy-efficient and do neuromorphic computing with bidirectional information transmission. The proposed study is suitable for designing a computing architecture for auditory, visual and multisensory systems. The bioinspired interactive computing device fuses multiple/mixed sensing signals with an exact and liquid state machine, which is effective in reducing the power and enhancing the computing power. In very large-scale integration, digital data processing methods can be designed using brain-inspired architectures. Here, a new approximate compressors design based on Multigate Field Effect Transistor (MuGFET) transistor named Fin Field Effect Transistor-FinFET for neuromorphic computing is presented. The leakage current problems and second-order effects in Complementary Metal Oxide Semiconductor (CMOS) are eliminated in the new approximate 4–2 compressor design. The FinFET and approximate multiplier for brain-inspired neuromorphic computing and liquid state machine architecture are effective, as observed from the results. In addition, in this paper, the design approach of a synapse circuit for a bioelectronic ubiquitous neuroprosthetic device is presented. The proposed circuit design combines nanoelectronics, electroceuticals, and neuroprostheses. The design consists of a synapse circuit which can provide electrical neuromodulation stimulation and be part of a neuromorphic architecture for bidirectional interactions. The delay is optimised along with power consumption and reported from the Extensive simulation. For implementation, FinFET and CMOS predictive technology models were used.

Abstract The current earthquake early warning system (EEWS) deployed by the China Earthquake Network Center processes data from more than 18,600 stations, handling approximately 110,000 packets per second in real time. However, the system is nearing its processing capacity limit and encountering bottlenecks. To address these limitations, we have developed and redesigned an earthquake early warning prototype system (EEWPS) by integrating existing earthquake early warning algorithms with the Flink distributed real-time big data processing engine. The system was tested using a dataset comprising 182 seismic events of magnitude M 3.0 or higher recorded in mainland China from 2008 to 2023. The EEWPS successfully processed 171 of these events, with 11 missed events and no false alarms. Notably, the time from the first station trigger to the first-warning report was 5.6 ± 3.7 s, whereas the time from the origin to the first report was 9.8 ± 5.5 s. The system demonstrated an epicenter deviation of 7.4 ± 9.6 km, a magnitude deviation of −0.12 ± 0.82, and a focal depth deviation of 1.3 ± 8.0 km. The results of the four representative test events further illustrate the system’s performance in practical scenarios. In the initial early warning reports, epicenter location errors were generally constrained within 10 km, whereas magnitude errors were maintained within 0.8 units. For stress testing, the EEWPS was deployed across four virtual machines, with a combination of event waveforms and simulated data. The system successfully processed data from approximately 31,000 stations, achieving an average throughput of over 186,700 packets per second. This performance indicates that the Flink-based EEWPS not only addresses the current processing bottlenecks of the existing EEWS but also offers a potential approach for the next generation of intelligent EEWS, capable of handling large-scale data and providing timely, accurate alerts in operational environments.

Abstract The increasing complexity and dynamic nature of Industrial Internet of Things (IIoT) demand scalable, adaptive, intuitive, and real-time automation frameworks. This paper presents a systematic literature review (SLR) of edge- and cloud-based collaborative learning frameworks for predictive maintenance and smart manufacturing tasks. In this SLR, we highlight the under-utilization of distributed computational architectures that provide complete automation support (design and run-time), flexibility, scalability, and inter- & intra-cloud service exchange while adhering to security management and integrity principles for solving IIoT tasks using modern artificial intelligence (AI) models at the edge/cloud. Recently, many IIoT applications have been designed using AI models that require robust, low-latency, and data-secure frameworks. This demand drives a trend toward distributed computational architectures in which data storage and processing are partially or fully decentralized. Common paradigms addressing this resource distribution include edge computing, federated learning, and private or hybrid clouds. We analyze 50 recent studies against IoT characteristics, AI performance metrics, and network/system management requirements. Our findings reveal underutilization of distributed architectures that support automation, interoperability, and security. While most solutions rely on centralized or hybrid clouds, fewer than 5% adopt federated or transfer learning, and over 60% remain dependent on supervised models. We also introduce a comparative perspective on network and security management, showing that local/private cloud implementations can reduce control-plane overhead and synchronization latency, though gaps persist in dynamic bandwidth allocation and zero-trust adoption. Finally, we benchmark our previously proposed local cloud-based collaborative learning (CCL) model against state-of-the-art solutions, highlighting its strengths in automation and interoperability, as well as limitations in adaptive computation and intelligent offloading. This review identifies the research gaps and opportunities for integrating collaborative AI, secure automation, and hybrid architectures to meet Industry 5.0 objectives of resilience, sustainability, and human-centricity.

ABSTRACT Due to the advantages of low‐voltage operation and memory characteristics, electrolyte‐gated transistors (EGTs) have been proposed as building blocks for neuromorphic systems. Recently, reservoir computing (RC) is attracting interests because RC utilize nonlinear system to enrich the input source into a higher dimension, whereas traditional neural network prefers linear system. RC also has the advantages of low energy consumption and fast response speed because it only trains at the readout layer. EGT is a promising candidate for RC due to tunable output current depending on input pulse voltage using its nonlinear system. Herein, we proposed poly (vinylidene fluoride‐hexafluoropropylene) (PVDF‐HFP) interlayered EGT for RC. Semicrystalline and porous morphology of PVDF‐HFP interlayer induced ion‐trapping at the interface between amorphous and crystalline regions resulted in long retention. Also, the thickness variation of the PVDF‐HFP interlayer modulates the filling capacity of trap sites, balancing ion‐transport for near‐linear weight update. We integrated near‐linear and asymmetric long‐term potentiation and depression curves into RC to elucidate the intensity difference between five frames of spatiotemporal events in dynamic vision sensor (DVS) gestures. Classification of 11 distinct gestures was successfully demonstrated using the RC framework. This work highlights compelling synergy between material‐level device physics and algorithmic architectures of neuromorphic computing.

As integrated circuits face increasingly stringent demands regarding power consumption, area, and stability, integrating novel spintronic devices with computing architectures has become a crucial direction for breaking through traditional computing paradigms. In the paper, switching mechanism of Magnetic Tunnel Junctions (MTJs) under the synergistic effect of Voltage-Controlled Magnetic Anisotropy (VCMA) and the Spin Hall Effect (SHE) is investigated. VCMA-assisted switching SHE-MTJ device is adopted, and a macrospin approximation model is established based on the Landau-Lifshitz-Gilbert (LLG) equation to systematically analyze its dynamic characteristics. The research demonstrates that applying VCMA voltage pulses with appropriate amplitude and width can significantly reduce the required spin Hall current density and pulse width for switching, thereby effectively minimizing ohmic losses and Joule heating. Furthermore, by incorporating a thermal fluctuation field, voltage-controlled SHE-MTJ device with stochastic switching behavior can be constructed, obtaining an approximately sigmoidal voltage-probability response curve. This provides an ideal physical foundation for stochastic computing and neuromorphic computing. Based on the above established fundamental discovery, an in-memory computing architecture supporting binarized Convolutional Neural Networks (CNNs) is proposed and designed in the paper. Combined with the lightweight network SqueezeNet, this architecture achieves a Top-1 recognition accuracy of 72.49% on the CIFAR-10 dataset, with a parameter count of only 1.25 × 106. This work offers a feasible spintronic implementation scheme for low-power, high-energy-efficiency edge-side intelligent chips.

Optical digital computing, leveraging optical signals for high-speed, efficient, and high-precision discretized digital computations and information processing, is widely applied in artificial intelligence, communication, and network. However, the development of multiplication/division, pivotal components within digital optical computing systems, has been constrained by the absence of effective computational physical dimensions conducive to manipulating optical signals and the stringent requirement for precise operational control. We propose a fixed-base multiplication and division scheme based on orbital angular momentum (OAM) modes using optical diffractive neural networks (ODNNs). Utilizing the OAM mode as the computational physical dimension, and employing ODNN to perform mode-parallel and independent transformations on it for achieving numerical equivalent shifts, it is feasible to realize 2n-fold multiplication and division operations. Using this mechanism, we have constructed a 3-layer ODNN to realize OAM mode fixed-base multiplication and division for n = 1, 2, and 3, with mode purities of the operation outputs reaching 99%. Furthermore, a dynamic switching between multiplication and division operations has been achieved within the same system through rotation of the phase matrix. Our design suggests a feasible pathway for fixed-base optical multiplication and division, and may offer useful insights for future research on optical digital computing architectures.

Hybrid brain-inspired computing, which integrates computer science-oriented artificial neural networks (ANNs) with neuroscience-oriented spiking neural networks (SNNs), represents a significant trend in the field of brain-inspired computing. This approach demonstrates considerable advantages in domains such as perception, cognition, and learning. Nevertheless, several challenges persist in deploying hybrid brain-inspired computing on resource-constrained edge devices, particularly in enabling these devices to effectively manage complex tasks. To tackle these challenges, we present eBrainISA, an edge-oriented hybrid brain-inspired computing instruction set architecture. eBrainISA is designed to support sophisticated ANNs, SNNs, and hybrid neural networks (HNNs), which integrates an extensive suite of neural operators that are tailored to accommodate the hardware limitations on edge devices. To validate the proposed architecture, we develop a unified low-precision quantization approach alongside a hardware simulator. Prior instruction set architectures offer limited support for hybrid brain-inspired computing, while ours can support all major networks including ANNs such as ResNet and Yolo, SNNs such as Spike_ResNet and Spike_Transformer, and HNNs such as DetectHNN and ReSpike with <1% accuracy degradation. Moreover, our architecture achieves about 5× memory saving compared to full-precision versions and 3.87-53.61× higher energy efficiency compared to edge GPU when executing different neural networks with diverse behaviors. This work demonstrates a promising balance between intelligent demands for general models and resource constraints on edge devices.

The conventional computer architecture has been facing challenges answering the ever-increasing demands from emerging applications, such as AI, for energy-efficient computation and memory hardware systems. Computational Random Access Memory (CRAM) represents a true in-memory computing paradigm that integrates logic and memory functions within the same array. At its core, CRAM relies on Magnetic Tunnel Junctions (MTJs), which serve as the foundational building blocks for implementing both memory storage and logic operations. However, a key challenge in CRAM lies in the non-ideal error rates associated with switching dynamics of MTJs, necessitating innovative approaches to reduce errors and optimize logic margins. This work proposes a novel approach of utilizing the voltage-controlled magnetic anisotropy (VCMA) to steepen the switching probability transfer curve (SPTC), thereby significantly reducing the logic operation error rate in CRAM. Using several numerical modeling tools, we validate the effectiveness of VCMA in modulating the energy barrier and switching dynamics in MTJs. It is revealed that the VCMA effect significantly reduces the error rate of CRAM by 61.43% at a VCMA coefficient of 200 fJ·V−1m-1 compared to CRAM without VCMA. The reduction of error rate is further rapidly amplified with an increasing TMR ratio. Furthermore, the introduction of the VCMA effect decreases the logic voltage (Vlogic) required for logic operations in CRAM and results in reduction of energy consumption. Our work serves as a first exploration in reducing the error rate in CRAM by modifying SPTC in MTJs.

Low-power, high-speed ternary adder using multi-threshold GNR transistors for efficient digital computing architectures

We use our experiences to form and update beliefs about the hidden states of the world1-3. When possible, we also gather evidence by observing others. However, how the brain integrates experiential and observational evidence is not understood. We studied the dynamics of evidence integration in a two-player game with volatile hidden states. Both humans and monkeys successfully updated their beliefs while playing the game and observing their partner, although less effectively when observing. Electrophysiological recordings in animals revealed that the anterior cingulate cortex integrates independent sources of experiential and observational evidence into a coherent neural representation of dynamic belief about the environment's state. The geometry of population activity revealed the computational architecture of this integration and provided a neural account of the behavioural asymmetry between experiential and observational evidence accumulation. This work lays the groundwork for understanding the neural mechanisms underlying evidence accumulation in social contexts in the primate brain.

This study establishes a novel machine learning paradigm integrating physical mechanisms. It aims to address the limitations of traditional methods in predicting the concrete fatigue life of high-speed railway (HSR) track slab, particularly their insufficient accuracy and poor interpretability due to overlooking microstructure characteristics and load randomness. The research method innovatively develops a multi-task learning framework combining the Graph Neural Network (GNN) and Transformer architectures. GNN abstracts concrete microstructure as topological graphs to capture spatial damage propagation paths, while Transformer analyzes random load spectra to identify critical temporal patterns determining fatigue damage. The multi-task framework simultaneously predicts fatigue life, damage evolution rate, and residual strength. Experimental results demonstrate the outstanding performance of the proposed GNN-Transformer model. After 150 training epochs, the proposed model achieves a Mean Squared Error of 0.224, Root Mean Square Error of 0.332, and Coefficient of Determination reaching 0.939. Meanwhile, based on the computational architecture used in the experiments, the model achieves an average single-sample prediction latency of 0.86s and a GPU utilization rate of 48.3%, demonstrating its strong potential for online deployment. All indicators significantly outperform existing models. The study concludes that this framework successfully constructs a high-accuracy, efficient, and physically interpretable prediction model by effectively integrating microstructure topology with macroscopic load sequence features. This study provides robust algorithmic support for predictive maintenance and intelligent health management of HSR infrastructure.

Markerless optical motion capture enables unobtrusive gait assessment, yet kinematic estimates remain sensitive to acquisition variability and to heterogeneous processing workflows, limiting cross-study comparability and reproducibility. This paper presents a modular computational architecture for processing markerless optical gait data, aimed at standardizing key steps from raw recordings to analysis-ready kinematic time series. Based on a structured comparison of commonly used pipelines and reported failure modes, the architecture specifies four sequential stages data acquisition, signal/pose preprocessing, gait-cycle segmentation, and representation/structuring and defines interfaces and quality-control checkpoints between modules. The pipeline integrates noise attenuation and normalization with a hybrid strategy: deterministic heuristics support rule-based quality screening and parameter initialization, while learning-based components target error-prone operations such as robust gait-cycle delineation under occlusions and variable viewing conditions. By explicitly separating concerns (capture, cleaning, segmentation, and representation) and by formalizing intermediate outputs and metadata, the proposed architecture provides an auditable foundation for implementation and subsequent experimental validation. The framework is intended to improve consistency of kinematic outputs and facilitate reproducible biomechanical analyses across laboratory and in-the-wild settings. Keywords: markerless motion capture; gait analysis; kinematic time series; signal preprocessing; gait-cycle segmentation; reproducible pipelines.