ruv-FANN
Neural networks and weight-space programming in Rust, JavaScript and WebAssembly. Inspect activations, edit weights, evaluate candidates, train small models, and explicitly promote or roll back revisions.
6 repos · 0 papers · 67 scholar
Showing 73 of 73 items
Neural networks and weight-space programming in Rust, JavaScript and WebAssembly. Inspect activations, edit weights, evaluate candidates, train small models, and explicitly promote or roll back revisions.
AI-assisted PCB design for KiCAD 10: downloads and issue tracker. Commercial, closed source, free during the beta.
Open Source Brain Platform | AI training and inference in Rust
A cross-platform neuromorphic robotics console for intelligent sensor visualisation, real-time robot monitoring, safe control and mission replay.
Neuromorphic computing event-driven memory matrix
Large Language Models (LLMs) are frontier deep learning systems that have achieved remarkable success across a wide range of AI services. However, their substantial computational and memory requirements make them difficult to deploy and run on local hardware. Due to these resource requirements, users often rely on untrusted cloud infrastructure providers to perform model inference. However, outsourcing introduces the challenge of verifying that the returned output is the genuine result of the specified model. In this work, we present DeepProve, the first system to enable efficient end-to-end verification of full LLM inference (i.e., for all generated tokens of a prompt) on untrusted cloud servers using zero-knowledge proofs (ZKPs). In contrast, prior work either provides only a proof-of-concept partial implementation for a single token (zkGPT, USENIX '25), or focuses exclusively on specific components of the inference pipeline, such as Softmax (zkLLM, CCS '24). DeepProve achieves end-to-end verification by certifying the correctness of the output sequence rather than encoding the expensive inference computation in-circuit, an approach that would require either circuit size quadratic in the sequence length or costly in-circuit modelling of RAM operations. The core building blocks of DeepProve are sum-check protocol and lookup arguments, which enable efficient proof of correctness of all operators needed for GPT-2 and Gemma 3, such as multi-head attention and layer normalization for GPT-2, and grouped-query attention, root mean square normalization, and rotary positional embeddings for Gemma 3. Our evaluation shows that DeepProve can prove inference of GPT-2 and Gemma 3 at approximately 174 and 86 tokens per minute, respectively, which is 20-60× faster than the state-of-the-art, without any significant loss in accuracy. Verification takes only 1 to 3.7 seconds. By distributing proof computation across multiple nodes, DeepProve can further improve the prover time while reducing the memory requirements for individual machines. With distributed proving, DeepProve can scale the throughput to 1855 tokens per minute. Our work represents the first full system for end-to-end LLM inference verification, thus paving the way for secure and trustworthy AI services.
This paper presents a comparative analysis and an ASIC-oriented design of a 16-bit Vedic Multiplier based on the Urdhva-Tiryakbhyam algorithm using Kogge-Stone (KSA), Carry Look-Ahead (CLA), and Carry Save (CSA) adders. The architectures are implemented in Verilog HDL and evaluated across 45 nm, 90 nm, and 180 nm technology nodes using the Cadence suite. Pre-layout synthesis metrics (cell count, delay, and power) are analyzed to study performance trade-offs across nodes. Furthermore, physical design placement and routing at 90 nm confirms ASIC standard-cell feasibility, although detailed post-route parasitic extraction is reserved for future hardware prototyping. Comparisons with FPGA-based works demonstrate significant improvements in speed and power efficiency, making the proposed KSA-based design highly suitable for high-speed, low-power VLSI and DSP applications.
Floating-point multiplication is a critical operation in modern digital systems, including signal processing, scientific computing, graphics processing, and high-performance embedded applications. With increasing computational demands, high-speed and area-efficient arithmetic units are essential. This paper presents the design and VLSI implementation of a 32-bit single-precision float- ing-point multiplier optimized for enhanced performance through architectural improvements. The proposed design improves mantissa multiplication using Ra- dix-4 modified Booth encoding to reduce partial products, thereby lowering switching activity and propagation delay. A Wallace tree reduction structure is employed for fast parallel partial-product compression, significantly accelerating accumulation compared to conventional array multipliers. To further enhance throughput, a pipelined architecture is introduced, dividing computation into bal- anced stages to minimize critical path delay and increase operating frequency. Exponent addition with bias adjustment, along with systematic normalization and rounding units, ensures accurate and stable floating-point computation, including proper handling of edge cases. The design is implemented in Verilog HDL and synthesized using standard VLSI tools. Results demonstrate improved timing performance and optimized hardware utilization compared to conventional archi- tectures, achieving an effective balance between speed, area, and complexity.
High-level synthesis (HLS) is becoming increasingly popular because it allows the specification of hardware accelerators in a higher level of abstraction than Register Transfer Level (RTL), for example, in C/C++. Common scenarios include acceleration of machine learning inference and even cryptographic functions. But is HLS secure for such scenarios? Constant-time (CT) programming is a common approach for mitigating timing side channels when executing sensitive software on CPUs. It is unclear whether generic CT rules such as not passing secret-dependent data to memory addresses or branching decisions provide adequate protection when an HLS flow turns CT software into hardware. Furthermore, CPU vendors and instruction set architectures (ISAs) often specify which instructions execute in data-independent time and are safe to use for CT programming, which is currently missing for HLS. To address this gap, we introduce CTurtle, the first C program generator that generates random, yet valid, C programs that follow customizable CT rules, making them CT with respect to certain inputs marked as secret. We further introduce TurboTurtle, an HLS fuzzer that can discover cases where HLS flows turn CT software into hardware that is vulnerable to timing side channels. TurboTurtle detects timing vulnerabilities in the HLS generated Register Transfer Level (RTL) designs, using a miter test bench that can detect when the generated accelerator may take a variable amount of time caused by different secret inputs. Applying TurboTurtle on four different HLS flows, we discover 12 new vulnerabilities where HLS flows produce non-CT hardware from CT software, for which we got two CVEs. Eight of these vulnerabilities establish unsafe operators for a particular HLS flow, and three persist in two HLS flows even after applying generic and tool-specific CT rules.
High-Level Synthesis (HLS) abstracts the hardware accelerator design process, but evaluating the generated RTL often becomes a “black-box” challenge for developers. To address this issue, we propose a systematic methodology to track and analyze the translation of C/C++ algorithms into hardware (RTL). Our approach integrates synthesis via Xilinx Vitis/Vivado 2024.2 with datapath mapping using Intel Quartus and behavioral waveform validation via ModelSim. By applying this methodology to different algorithms, we mapped how the compiler translates constant complexity codes into linear pipelines and iterative logic into multi-stage Finite State Machines (FSMs). To evaluate the architectural results extracted by our method, we conducted performance evaluations on a PYNQ-Z2 board. Experimental results show that while the hardware reduces clock cycles, the absolute speedup is constrained by communication overhead. We conclude that the proposed methodology provides a replicable pathway for developers to understand the HLS-to-RTL mapping.
The management of laboratory assets at the Faculty of Marine and Fisheries Science (FIKP) is still carried out manually using physical documents and Microsoft Excel, which is prone to recording errors, data duplication, and reporting delays. This study aims to develop an integrated web-based laboratory equipment inventory and rental information system. The development method uses the Agile Development approach with stages including requirements, design, development, testing, deployment, and review, with testing conducted using Blackbox Testing. The system was built using PHP with the Laravel framework and MySQL database. Testing results on 19 test scenarios across six main modules (login, equipment management, borrowing, returning, reporting, and user management) showed a 100% success rate with all functions operating as expected. This system integrates inventory, borrowing, rental, payment, and reporting features into a single digital platform, thereby improving efficiency, transparency, and accountability in laboratory asset management
Ferrofluid pumps exhibit significant potential in chip cooling, precision transport, and microfluidic systems owing to their non-contact actuation, self-sealing capability, and controllable microscale transport characteristics. However, existing studies have mainly focused on flow-rate characteristics under zero-backpressure conditions, while the interfacial evolution, dynamic sealing degradation and wall-friction mechanisms of ferrofluids under backpressure remain poorly understood. This study systematically investigated the maximum pressure-bearing capability, flow evolution, and wall-adhesion behavior of ferrofluid pumps under different backpressure conditions, revealing the mechanisms of backpressure-induced interfacial instability and backflow formation. The results show that, at low rotational speeds, ferrofluids maintain a stable slug-like structure and achieve near-ideal positive-displacement transport. As the rotational speed increases, the response of the ferrofluid to the external rotating magnetic field gradually lags behind, degrading the dynamic sealing performance and forming leakage channels under backpressure. Furthermore, backpressure not only suppresses the net flow rate by enhancing the reverse pressure gradient, but also alters the wall adhesion and liquid-film evolution of ferrofluids. Ferrofluids with different concentrations exhibit distinct interfacial response characteristics: low-concentration ferrofluids tend to generate free magnetic-liquid residues and severe wall adhesion, whereas high-concentration ferrofluids produce significant wall deposition due to high viscosity and intensified wall shear effects. Intermediate concentrations yield a relative balance between cohesion and fluidity. In addition, moderate backpressure enhances interfacial confinement of the liquid slug, whereas excessively low or high backpressure induces interfacial instability, exhibiting pronounced nonlinear characteristics. These results provide new physical insights for the design of high-backpressure ferrofluid micropumps and magnetically driven microfluidic systems.
Pressure management is essential for the efficient and sustainable operation of water distribution networks. Its benefits are manifold: it reduces leakage and pipe bursts, lowers pressure-dependent consumption, and decreases energy use. Overall, it enhances the physical, economic, and environmental sustainability of the system. The most commonly applied method to control excess pressure involves dissipating surplus energy using pressure-reducing valves (PRVs). However, in recent decades, an alternative approach has gained traction: recovering part of that energy using pumps operating as turbines (PATs). A third, less frequently considered but potentially more effective strategy is to reduce pressure at the source by adapting the hydraulic design to the terrain’s topography. This paper examines these three strategies from an energy perspective, commonly known as the three Rs—reduce, recover, and remove—with a particular focus on the first, reduce. This approach is aligned with the principles of eco-design and sustainable management of water distribution networks.
The rapid and accurate prediction of vehicle flow field distributions and aerodynamic force coefficients is of great importance in aerodynamic design scenarios. The current deep learning-based surrogate models face three major challenges when handling large-scale irregular meshes: low computational efficiency, the absence of physics priors that are aligned with the dominant flow direction, and the tendency to lose high-frequency geometric details on the surface of the target body. Considering the above issues, the ordered physics Mamba (OP-Mamba) flow field prediction model is proposed in this work based on selective state–space sequence modeling. First, OP-Mamba clusters discrete point clouds into a compact set of physics-aware slices through adaptive soft assignment, thereby reducing the imposed computational burden. A spatial sorting mechanism based on the weighted geometric centroids of these slices is then introduced to reorder them along the principal flow axis. Second, combined with this reordering scheme, bidirectional Mamba scanning not only captures the downstream convective evolution process but also provides a direct backward path for incorporating the nonlocal pressure effects that are associated with the downstream wake structures, thereby efficiently extracting long-range dependencies. Finally, OP-Mamba uses a point cloud-domain residual compensation scheme in which feature mappings and skip connections are implemented in the original discrete space, thus mitigating the high-frequency geometric feature attenuation that is caused by slice-level dimensionality reduction. Extensive experiments conducted on three-dimensional (3D) industrial vehicle datasets (including Baidu CIKM and ShapeNet-Car) and the two-dimensional (2D) airfoil dataset (AirfRANS) demonstrate that OP-Mamba achieves state-of-the-art accuracy in both flow field reconstruction and macroscopic aerodynamic force prediction tasks. Additionally, it maintains competitive performance under physical-field noise augmentation during the training procedure and demonstrates its applicability to both 3D vehicle and 2D airfoil settings.
This study presents the simulation and experimental validation of the Aerodynamic Induction Phenomenon (AIP) in blades for low-power wind turbines. This study describes the physical phenomenon that generates additional aerodynamic torque through the model of Induction Blade (IB). Computational Fluid Dynamics (CFD) simulations in 2D and 2.5D were employed using the k-ω SST turbulence model. The results reveal that the AIP is a jet thrust ( , a reaction force generated by momentum exchange) produced by a double pressure stagnation zone that accelerates the flow between the primary airfoil (A1) and the secondary airfoil (A2). The quantification of the was performed using an integral equation derived from the rocket thrust equation, demonstrating that this component is the predominant driving force under static starting conditions. Experimental validation using an Enhanced Induction Blade (EIB) design on a 50W micro-turbine confirmed that it reduces the start-up wind speed to 3.60 m/s, representing an improvement of 23.8% over the traditional blade and 15.9% compared to the base IB model, without compromising power production. These results validate the 2D sectional analysis for optimizing the geometric parameters of the IB, consolidating aerodynamic induction as an effective strategy for moderate wind speed environments.
Geomagnetic storms disturb the ionosphere through interacting electric fields, particle precipitation, thermospheric heating, and changes in atmospheric composition. Their consequences for satellite services depend on the structure of the disturbed plasma and the design of the receiving system, rather than on magnetic storm intensity alone. This narrative review examines the physical pathways linking solar wind energy input to electron density redistribution, radio propagation errors, scintillation, and communication interruptions. It integrates established ionospheric research, published storm observations, and operational space weather guidance while distinguishing measured findings from analytical illustrations. Particular attention is given to differences between large-scale total electron content variations and small-scale irregularities, because these disturbances require different monitoring and mitigation strategies. An original graph applies the first-order dispersive delay equation to hypothetical electron content values at representative navigation and communication frequencies. The calculation demonstrates frequency sensitivity without claiming to reproduce a particular storm. The review argues that resilience requires coordinated environmental monitoring, suitable receiver tracking, frequency and route diversity, and application-specific integrity checks. Remaining challenges include sparse observations over oceans, imperfect forecasts of irregularities, and limited access to receiver performance records. Improved assessment should connect local plasma conditions to measurable service outcomes and explicitly account for uncertainty.
The rapid advancement of wearable robotic technologies has opened new possibilities for improving healthcare delivery, rehabilitation, and patient mobility. This paper presents the design and development of an AI-assisted wearable robotic system intended to support healthcare applications such as physical rehabilitation, mobility assistance, and injury prevention. The proposed system integrates lightweight wearable hardware with intelligent control algorithms to adaptively assist human motion in real time. Machine learning models are employed to analyze multimodal sensor data, including inertial measurement units, electromyography signals, and force sensors, enabling accurate intention detection and personalized assistance levels. An adaptive control framework combining supervised learning and reinforcement learning is introduced to optimize motion support while ensuring user comfort and safety. Experimental evaluations conducted with healthy participants demonstrate improved motion stability, reduced muscular effort, and enhanced task performance compared to conventional non-intelligent wearable systems. The results highlight the potential of AI-assisted wearable robotics to deliver personalized, efficient, and scalable healthcare solutions, paving the way for next-generation assistive and rehabilitative technologies.
This paper presents the conceptual framework, method of making and cultural context of the work Soft Collapse. This experimental artwork constructs a bridge between the ancient appreciation of Chinese scholar’s rock (gongshi 供石) and an extension of industrialisation and early computation in the form of a 3D-printed textile-like shell. Gongshi are traditional Chinese scholar’s rocks valued for their naturally shaped abstract beauty. The type of gongshi used in this work, is a water-shaped Taihu harvested from the bottom of Lake Tai, formed through discrete acts of erosion. We argue that both the shape of the rock and the woven 3D printed fabric, created line by line through binary decisions, are structurally analogous. They are surfaces produced by long sequences of localized, two-valued events, and both are, by their respective Eastern and Western traditions, read as carriers of meaning exceeding the sum of those events. This creates a hybrid between two information systems separated by millennia. Through structured-light scanning, curvature and geodesic surface analysis, and a custom-designed toolpath-generation, the rock’s topography is rewritten as routing instructions for an additive fabrication system operating as a one-strand-at-a-time loom—a contemporary Jacquard mechanism. Here, the rock functions as a physical dataset, a punch card written by erosion. The resulting three-strand interlocking weave is fabricated as a shell of the scanned surface. The work is contextualised within contemporary concepts of cosmotechnics, a genealogy of the loom as an information machine, and the discourse on digital craft. The resulting hybrid artifact becomes a model of creative practice in which the artist curates networked technologies of scanning, computation, and fabrication, giving meaning through the orchestration of geological morphology and algorithmic instruction into a single object that is at once found and made
This paper presents Ce que l’air retient (What the Air Holds Back), an extended reality (XR) installation that interrogates the environmental monitoring infrastructure surrounding the Fonderie Horne, a copper smelter in Rouyn-Noranda, Québec, Canada. Four three-dimensional objects modeled after real structures in the territory—a chimney, a playground, and two air quality monitoring stations—are anchored at 1:1 scale in the physical exhibition space through passthrough mixed reality, each accompanied by a polyphonic sound design combining my voice, interview excerpts from environmental experts, and ambient field recordings made on site. Drawing on agnotology (the structural production of ignorance), Rob Nixon’s slow violence, and the Realities in Transition White Paper’s call for grounded, non-escapist XR, the paper positions the installation as an artistic practice that casts new light on existing environmental measurement, and documents the artistic process from field research to 3D modeling, shader development, and sound design.
Lithic Reverie is an interactive sound installation and performance that reimagines the traditional Chinese stone chime 磬 (qìng) using concrete fragments retrieved from urban construction sites. Positioning concrete as a material that integrates geological deep time with industrial human processes, the work gives voice to exploitative extraction practices: geological, ecological, cultural, and ontological. Rather than utilizing sonification purely for data representation, this work contributes as a case demonstrating how the sonic translation of physical matter can serve as a form of practice-based ecological critique. The iterative development process highlights the value of practice-based research and leads to the unique dual-mode of this installation. By detailing the conceptual narrative and technical workflow, Lithic Reverie presents a case reframing traditional cultural forms with digital sound technologies while addressing extractivism in industrial processes. This work exists in two modes: a fully automated installation and a hybrid performance in which a human performer replaces selected solenoids, striking fragments with everyday objects and collected minerals. This paper describes the conceptual framework, system design, the audio system architecture, the dual mode development process, and contributions of the work to the fields of sound art and critical making.
Emerging AI accelerators increasingly adopt spatial and dataflow execution models based on distributed compute units, local memories, and explicit data movement. Although primarily designed for machine learning, these architectures are also being explored for scientific computing. This paper presents a tile-based BLAS library for Tenstorrent dataflow architectures, mapping representative Level-1, Level-2, and Level-3 operations onto Tensix cores through tiled layouts, circular buffers, explicit DRAM–SRAM movement, and local data reuse. We describe the main design decisions required by this execution model, including data layout, kernel decomposition, synchronization, data-format constraints, and workload distribution. The library is evaluated on a dual-ASIC Wormhole™ n300 accelerator using device-side profiling in single-chip and dual-chip configurations. AXPY and GER show sublinear multichip scaling, whereas GEMM benefits more strongly from the second ASIC, reaching a maximum measured throughput of 5.79 TFLOP/s and an observed maximum speedup of 2.07×. The results show that tile organization, data movement, arithmetic intensity, and workload distribution strongly influence BLAS performance and multichip scaling on Tenstorrent accelerators.
Production recommendation platforms often grow one query at a time, with each new business requirement creating another pipeline instead of reusing existing ones. We study this in a large retail platform where more than fifty independently scheduled BigQuery/Dataflow pipelines implement only a few recommendation types, raising maintenance cost while cooccurrence counting still lacks user representations. We propose an architecture that repositions the data warehouse as a shared training and feature platform built around a feature store, a two-tower embedding model, and an approximate-nearest-neighbor vector index, so new recommendation variants require only new training-data partitions rather than new pipelines, together with a four-phase migration methodology with governance and rollback support. Experiments on the public RetailRocket dataset compare popularity, co-occurrence, and two-tower retrieval. After validation-set tuning and multi-seed evaluation, the two-tower model achieved higher Recall@10 and Recall@20, and was the only method that recommended low-interaction items (Recall@10 of $0.0053 \pm 0.0021$ vs. 0.0000 for both baselines), although cooccurrence retained a slightly better MRR@10 (paired difference -0.0047, 95% CI [-0.0091, -0.0005]). The gap between initial and tuned results shows why proper tuning is essential before drawing architectural conclusions.
The rapid advancement in Large Language Models (LLM) research has significantly enhanced natural language processing capabilities across diverse applications. However, the massive model parameters and substantial computational demands impose strong requirements for high-throughput hardware to enable efficient edge-side inference. Due to the severe memory imbalance of prefilling and decoding phases, and the trade-off between accuracy and resource cost, it is thirsty for architecture for optimization on memory and computation. The significant memory imbalance between the prefilling and decoding phases, combined with the accuracy-resource trade-off, creates critical demand for architectural optimizations in both memory and computation. We propose an FPGA-based accelerator, ELLME, employing a novel architecture with memory-friendly dataflow scheduling, W4A8/W8A8 low-bit-width matrix computation, and approximate nonlinear computation to address these challenges. We implement the Qwen-2.5 model on our acceleration system and develop a quantization algorithm to restore inference accuracy loss, achieving 99% accuracy on the WNLI dataset. ELLME achieves 307 GOPS at 300 MHz operation frequency, delivering 280 tokens/s in prefilling and 1.9 tokens/s in decoding phases under real task workloads.
Edge AI increasingly requires simultaneous vision-language captioning and real-time object detection, which can exceed the deep learning processing unit (DPU), programmable logic (PL), and double-data-rate (DDR) memory resources of a single field-programmable gate array (FPGA). We present a distributed architecture on Zynq UltraScale+ radio frequency system-on-chip (RFSoC) that repurposes integrated digital-to-analog converter / analog-to-digital converter (DAC/ADC) blocks for hardware emulation of a raw inter-partition channel without equalization. The master partition executes a vision-language model (VLM) pipeline with a DPU-based visual encoder and a high-level synthesis (HLS)-based 8-bit integer (INT8) transformer accelerator, while the slave partition runs YOLOv5 over the unequalized DAC/ADC channel. On a ZU47DR loopback prototype with quadrature phase-shift keying (QPSK) DAC-to-ADC transmission (1.024gigasamples/s, 500MHz) and no Tx/Rx equalization, we achieve 17tokens/s generation throughput, 32ms YOLOv5 DPU latency, and zero observed bit error rate (BER) over 100M pseudorandom binary sequence (PRBS) samples. Preliminary two-board tests show no observed bit errors on nominal links (DAC0→ADC0 and DAC1–DAC6 loopback), while an attenuated DAC7→ADC7 link exhibits visible degradation, supporting scalability and equalization-threshold characterization.
Mixture-of-Experts (MoE) models with FP8 quantization enable efficient large-scale model deployment, but the storage and memory bandwidth overhead of the expert weights remains significant. Existing lossless compression schemes primarily target high-precision weights of large language models (LLMs) using layer-wise contiguous encoding, and thus fail to support the on-demand, low-latency decoding required by dynamic expert access in MoE inference. To address these issues, this paper proposes LPAC-MoE, a lossless and prefetch-aware weight compression framework for FP8 MoE through algorithm–hardware co-design. Specifically, LPAC-MoE uses context-conditioned canonical Huffman coding with globally shared code-length tables to exploit statistical redundancy in FP8 weights, along with a prefetch-compatible compressed organization for selective decoding. We further integrate an efficient hardware decoder with an FP8 systolic array to fuse prefetched decoding and computation, reducing end-to-end latency. Experimental results show that LPAC-MoE reduces storage by 18.79% compared with the original FP8 weights. The proposed architecture yields a lightweight decoder with only 4.78% area overhead compared to FP8 systolic array, while maintaining a high energy efficiency of 8.096 TOPS/W and an amortized end-to-end latency of 9.453ns.
This paper presents a mixed-signal 10T-SRAM Compute-in-Memory (CiM) Processing Element (PE) based systolic array accelerator for Deep Neural Networks (DNNs) implemented in 2 nm GAA CMOS technology. DNN accelerators typically adopt digital or analog CiM macros considering bit precision-energy trade-offs. On the other hand, systolic arrays offer excellent data reuse, but they have rarely been fused with CiM, leaving a critical gap in energy-efficient, data-reuse-aware architectures. To close this gap, we propose a 32 × 32 systolic array in which each PE integrates an SRAM-based CiM core capable of executing four independent 8 additions of 8b × 8b operands in a single iteration. The mixed-signal CiM approach employs a single-cell Analog-to-Digital Converter (ADC) that performs conversion in the same cycle count as a conventional column ADC while occupying 8× less area. Post-layout simulation of our mixed-signal CiM array in predictive 2 nm PDK achieves 5417 TOPS/W for bitwise computations, which is nearly 3.6× better than 3 nm Digital CiM and 1.9× better than 7 nm Analog CiM State of the Art (SOTA). It also shows a bitwise compute density of 1679 TOPS/mm2. System-level evaluations across MobileNet, ResNet-18, and ResNet-50 workloads including off-chip DRAM access demonstrate energy efficiencies of 1.11, 1.87, and 2.03 TOPS/W, respectively, for weight-stationary INT8 dataflow.
The evolution of edge AI workloads from convolutional neural networks (CNNs) to Transformer-based models introduces increasingly heterogeneous computation patterns that challenge conventional embedded processors. In particular, memory access overhead, limited data-level parallelism, and the complexity of quantized arithmetic become major bottlenecks for efficient inference. This paper presents SARA, a profiling-guided RISC-V application-specific instruction-set processor (ASIP) for efficient execution of diverse edge AI workloads. By profiling instruction-level behavior on MLPerf Tiny and MobileViT benchmarks, we identify dominant workload characteristics and use them to guide coordinated hardware–software co-design, including ISA extensions, microarchitectural customization, and software-programmable dataflow optimization. The resulting design integrates vector-style execution for parallel low-precision multiply-accumulate operations, fusion-style instructions for quantized scaling and clamping, and vectorized memory access with post-modify addressing to improve both compute throughput and data movement efficiency. Implemented in 28nm CMOS, SARA achieves 64GOPS at 1GHz with an energy efficiency of 0.41GOPS/mW. Compared with a baseline RISC-V processor, SARA reduces dynamic instruction count by up to 66.8% and achieves 3.0×–13.5× speedup across representative TinyML workloads, including MobileNet, ResNet, and MobileViT.
Chiplet-based architectures for deploying large Transformers suffer from significant communication latency. This is mainly due to limited Network-on-Package (NoP) bandwidth, especially in Weight-Gathered dataflows. Optimizing this is difficult because non-linear normalization prevents pre-communication compression, and fixed-rank methods ignore layer-wise variations. To address this, we propose Projection-Invariant Normalization and Energy-Aware Compression (PIECE). Architecturally, PIECE enables pre-communication compression on activations by decoupling Normalization statistics with an energy-ratio correction factor. On the other hand, PIECE dynamically learns optimal per-layer compression ranks by differentiable soft-gating and knowledge distillation. Evaluations of vision and language models show that the proposed PIECE outperforms fixed-rank baselines. It can reduce end-to-end inference latency by up to 18.17% in the CV task and 20.31% in the NLP task. Furthermore, compared to the baseline, PIECE improves the top-1 accuracy of ViT-Base and AIMv2-lit by 3.63% (at 59% retention) and 1.09% (at 70% retention), respectively, while reducing the perplexity of OPT-1.3B by 0.38 at 68% retention.
Integrated sensing and communications (ISAC) enables spectrum reuse for RF sensing of airborne threats such as unexplained aerial phenomena (UAP). In full-duplex ISAC transceivers, strong transmitter-to-receiver self-interference (SI) masks weak target reflections, limiting sensing dynamic range. This paper presents the design, implementation, and experimental validation of a real-time wideband digital SI cancellation architecture for a full-duplex ISAC system operating at 5.8 GHz. The key contributions are: (1) a real-time in-system impulse-response-based SI channel estimation method requiring no external calibration equipment; (2) a massively parallel 256-tap polyphase systolic matrix-vector multiplier (MVM) implemented on an AMD-Xilinx RFSoC ZCU111, enabling wideband digital cancellation at 1 GHz instantaneous bandwidth; and (3) experimental validation of SI cancellation and cross-correlation-based range estimation across BPSK, QPSK, 16-QAM, and 64-QAM waveforms. The system achieves approximately 20 dB average SI suppression across 1 GHz bandwidth, and distance estimation results closely match theoretical delays for all four modulation formats. The single-element Tx/Rx prototype establishes the perchannel cancellation building block required for scalable multielement ISAC arrays targeting UAP sensing applications.
Residential buildings are responsible for a large and rising share of global and national energy consumption, and a significant part of this demand is fixed at the design stage through decisions about room arrangement, orientation, compactness, and zoning. Traditional layout design relies on the designer's experience and iterative manual trial-and-error, a process that cannot practically explore more than a small fraction of the possible design space and therefore routinely leaves substantial energy savings unrealized. Artificial intelligence (AI) - encompassing machine learning (ML) predictive models, evolutionary and metaheuristic optimization algorithms, and, more recently, generative deep-learning models such as generative adversarial networks (GANs) and diffusion models - offers a way to search this design space rapidly and systematically, coupling fast energy-performance prediction with automated layout generation and refinement. This paper reviews and synthesizes recent research on AI-based optimization of residential building layouts for energy efficiency, describes the principal computational pipelines used (machine-learning-based energy prediction, evolutionary/genetic-algorithm optimization, and generative layout models), and develops a representative framework and case-study workflow suited to Indian residential construction. The review shows that machine-learning surrogate models (random forest, gradient boosting, deep neural networks) can predict building energy consumption with R² values above 0.99 in some studies, several orders of magnitude faster than physics-based simulation, allowing them to be embedded inside evolutionary optimization loops; genetic-algorithm and multi-objective evolutionary approaches (e.g., NSGA-II) coupled with such surrogates have achieved reported energy reductions of roughly 5- 25% relative to conventional designs; and diffusion-based generative models combined with energy-aware genetic refinement have achieved up to 17.5% energy reduction relative to baseline generative layouts while preserving architectural plausibility and spatial constraints. Based on this synthesis, the paper proposes a three-stage AI-based layout optimization framework - datadriven energy prediction, generative layout synthesis, and evolutionary refinement - and discusses its applicability, limitations, and future development for Indian residential design practice, including integration with the National Building Code, Eco-Niwas Samhita, and locally available climate and construction-cost data
BACKGROUND Growing interest suggests that the broad application of artificial intelligence (AI) in acute myocardial infarction (AMI) has advanced diagnosis, prognosis, and treatment. AI methods applied to AMI research have recently yielded significant predictive outcomes. However, few studies have systematically analyzed this field using bibliometric approaches. We employ CiteSpace to examine the visual relationships of publications, identify key research trends and hotspots, and stimulate broader participation in AI-AMI studies. METHODS Data for this bibliometric analysis were retrieved from the Web of Science Core Collection. A systematic search strategy was developed using Medical Subject Headings to identify publications pertaining to AI in AMI from January 2016 to December 2025. The initial query returned 2089 records, which were subsequently refined through manual screening and algorithmic deduplication, yielding a final corpus of 2088 eligible publications. Bibliometric analyses were performed using CiteSpace (version 6.2.R6) with the following configuration: annual time slices of 1 year, a selection threshold of the top 25 most frequent items per slice (Top N = 25), and network pruning via the Pathfinder algorithm applied to both sliced and merged networks. Multidimensional analyses were conducted across six node types - country/region, institution, author, cited author, cited journal, and keyword - employing co-occurrence mapping, co-citation analysis, cluster detection, and burst term identification. These approaches collectively delineated the intellectual landscape, collaborative architectures, and thematic evolution characterizing the AI-AMI research domain. RESULTS We analyzed publications, citations, and networks. Post-2016, publications surged from 24 (2016) to 517 (2025) with dramatically increasing citations. Global collaboration centered on the China (705 publications), USA (553 publications), and India (195 publications), with Harvard-linked institutions forming the dominant cluster (e.g., Harvard University: 100 publications, centrality = 0.12). Author networks identified key bridges Saba, Luca (13 publications, centrality = 0.05) and Pontone, Gianluca (centrality = 0.06), contrasting with high-output but low-centrality authors like Acharya, U Rajendra (21 publications). High-centrality cited authors (e.g., ACHARYA UR: centrality = 0.09) were pivotal nexuses. Journal analysis showed CIRCULATION (1375 cites) and J AM COLL CARDIOL (1120 cites) as citation leaders. Keywords revealed evolving focus: prevalent terms ("myocardial infarction": 698 freq., "machine learning": 603 freq.) were highly productive, whereas central terms ("prediction": centrality = 0.36, "acute myocardial infarction": centrality = 0.24) bridged domains. Keywords bursts shifted from computational themes pre-2020 (e.g., "neural network," strength = 8.31) to the practical application of AI post-2020 (e.g., "system," strength = 5.43 and "prevalence," strength = 4.61). CONCLUSION AI-AMI research exhibits globally distributed multilevel collaboration structures anchored by major nations/institutions and critical author bridges. Knowledge dissemination hinges on domain-bridging journals like CIRCULATION. Thematic emphasis has evolved from computational techniques towards clinical applications, interconnected by high-centrality concepts.
The increasing demand for highly customised products challenges the ability of manufacturing systems to sustain high levels of automation, particularly in high-mix low-volume (HMLV) environments. Although decision support has become a major research area in response to these challenges, existing reviews either focus narrowly on specific methods such as reinforcement learning for scheduling or address HMLV manufacturing without systematically analysing decision support across manufacturing levels. This paper addresses that gap through a survey of 205 publications on decision support in HMLV manufacturing. The survey provides a structured synthesis of how decision support approaches are applied across four hierarchical manufacturing levels: process, line and cell, factory, and enterprise and supply chain. The reviewed studies are analysed through four research questions addressing: (1) the key drivers motivating decision support in HMLV manufacturing, (2) the optimisation objectives pursued, (3) the technologies and methods employed, and (4) the evaluation practices and limitations reported. The analysis reveals a rapid growth in research activity over the past decade, with a strong methodological concentration around scheduling problems, which account for 69.8% of the reviewed literature, while areas such as logistics and supply chain (6.3%) and quality and maintenance (2.9%) remain comparatively underexplored. Reinforcement learning, graph neural networks, digital twins, and metaheuristic optimisation emerge as dominant technologies and methods, within dynamic scheduling contexts. The findings show that current research is concentrated on simulation-based scheduling optimisation, with comparatively limited attention to cross-level decision support, industrial validation, human-centred decision-making, and integrated decision-support architectures spanning multiple manufacturing levels. Based on these findings, the survey identifies key research gaps and outlines future directions for decision support in HMLV manufacturing. These findings should be considered alongside the search and screening process, which relies primarily on broad HMLV, manufacturing, and decision-support terminology and may therefore under-represent research using specialised, industry-specific terminology.
Objective. The purpose of this research is to search for and analyze solutions in the field of semantic segmentation of land cover on high-resolution satellite images, which provide high quality and reduce dependence on the expensive process of pixel-by-pixel data annotation. Method. The study uses three methods: classical supervised learning based on the U-Net architecture with a ResNet101 encoder, the unsupervised STEGO method that uses features extracted by a self-learning transformer, and the state-of-the-art third-generation DINOv3 model, which is pre-trained on the large-scale satellite dataset SAT-493M and adapted to the segmentation task by adding a segmentation head and a knowledge distillation procedure to the convolutional network. The experimental evaluation is performed on the open DeepGlobe dataset using the IoU metric for land cover classes. Result. The results demonstrate that a baseline supervised approach yields strong performance when a sufficient amount of labeled data is available, while STEGO confirms the feasibility of generating semantically consistent regions without the direct use of annotations. High metric scores were achieved by transferring DINOv3 features into a convolutional architecture via distillation, which improved segmentation quality when training on a limited subset of the data. Conclusion. Self-supervised methods achieve high segmentation accuracy even when labeled data is scarce, which justifies their use and further development within the field of computer vision as applied to Earth remote sensing data.
Context quote extraction (CQE), which selects the most relevant verbatim span from a source document given an external context, sits at the intersection of information retrieval and retrieval-augmented generation (RAG). As LLM-based systems increasingly require grounded, attributable outputs, the ability to extract the correct span has become a practical necessity. To the authors knowledge, no prior systematic review has mapped CQEs methodological development. Six databases were searched following the PRISMA 2020 guidelines, supplemented by citation snowballing from five anchor studies. Of the 388 records identified, 142 were removed before screening (98 duplicates, 44 non-English), leaving 246 screened by title and abstract and 146 assessed in full text. The 57 retained studies were appraised using a five-criterion quality rubric (mean: 6.4/10; range: 4–10). Four method clusters were identified, tracing an arc from lexical/feature-based approaches (2012–2017) through encoder-based neural models (2016–2022) and generative LLM prompting (2022-present) to multitasking RAG pipelines (2024–present). Among the benchmarks where systems were directly evaluated, RA-MTR reports the strongest results (Precision@1:71.8%; BoW F1:61.4% end-to-end). Two structural gaps persist: the journalistic domain accounts for approximately 5% of the included studies despite representing the highest volume of real-world applications, and only two publicly available CQE-native benchmarks were identified. The field has advanced rapidly in architectural sophistication while remaining constrained by the scarcity of benchmarks, metric fragmentation, and unresolved subjectivity. Six research priorities are proposed, with news-domain CQE and hallucination-resistant extraction being the most important. This review provides what is, to the authors knowledge, the first comprehensive evidence base for directing future CQE research.
To address the limited predictive accuracy of existing models and the challenges in optimizing stacking models, this study proposes an interpretable stacking learning model that is automatically optimized using a genetic algorithm (GA) to predict the shear capacity of reinforced concrete (RC) beams with and without stirrups. A diverse candidate model pool was constructed using 11 classical regression algorithms spanning five major domains: ensemble learning, neural networks, Bayesian theory, kernel methods, and nonlinear fitting. A Tree-structured Parzen Estimator (TPE)-based optimization algorithm was then employed to automatically and finely tune the hyperparameters of each individual model, and GA was introduced to perform global automated optimization and dynamic updating of the base model and meta-model combinations within the stacking architecture. The results demonstrate that the proposed GA-Stacking exhibits excellent predictive accuracy and generalization compared to R-Stacking and code-based models, while achieving search efficiency far superior to random search. In addition, both global and local Shapley additive explanations (SHAP) analyses indicate a positive correlation of section geometry (b, h0) and material strength (fc, fsv) to the shear capacity. For λ < 2.80, the SHAP predictions exhibit a pronounced enhancement in shear capacity, followed by a subsequent degradation trend as the shear span-to-depth ratio increases. These findings provide valuable methods and insights for studying and applying RC beam shear capacity prediction.
Graph representation learning has become a powerful paradigm for modeling the relational structure of visual data in computer vision tasks. Unlike traditional grid-based representations, graph representation learning offers a flexible framework for modeling non-Euclidean structures, which can capture long-range dependencies while facilitating structured reasoning. While several recent surveys review general Graph neural networks (GNNs) architectures and applications across domains, existing literature reviews in computer vision are typically organized around tasks or data modalities. As a result, there remains a lack of a unified perspective for categorizing graph representations by the structural roles they play in visual reasoning. Following standard literature search guidelines, we collected studies published between 2017 and 2025 from major Computer Vision/Machine Learning conferences and journals using targeted keywords related to graph representation learning. From this corpus , we identified representative works covering seven graph schemas and extracted their graph designs, learning mechanisms, and task-level findings to build our taxonomy. Our analysis identifies seven fundamental graph schemas: pixel, point, skeleton, object, region, image, and label graphs, which can form a hierarchy ranging from low-level spatial detail to high-level semantic abstraction. For each schema, we review representative models, discuss graph construction strategies, and compare learning mechanisms, including graph convolutional networks (GCNs), graph attention networks (GATs), spatio-temporal (ST)-GNNs, and graph transformers. Common challenges across schemas include reliance on predefined topologies, uniform neighbor aggregation, difficulty modeling long-range dependencies, and scalability limitations in large graphs. Emerging solutions include graph structure learning, attention-based and adaptive message passing, hierarchical and multiscale graph designs, and integration of graph reasoning into transformer architectures. This survey presents a unified taxonomy of graph representation learning in computer vision, synthesizing methodological trends across diverse graph schemas. By clarifying the structural roles of graphs from pixels to labels, we provide a coherent framework for understanding how graph-based reasoning advances visual recognition, geometric understanding, and scene-level interpretation. Traditional challenges, such as adaptive graph construction, efficient reasoning on large graphs, and standardized evaluation, highlight important directions for future research. The insights presented here aim to guide the development of more scalable, interpretable, and context-aware graph-based solutions.
Introduction. This paper proposes the design of an electroencephalogram (EEG) signal analysis system based on a Fuzzy Adaptive Resonance Theory Mapping (Fuzzy ARTMAP) neural network for clinical signal classification. Materials and Methods. To accelerate the signal clustering process, a parallelized Fuzzy ART architecture is introduced. The system parallelizes the evaluation of choice functions and vigilance condition verification, replacing iterative sequential search with a unified parallel process. Standard Winner-Takes-All (WTA) units are replaced with rank-order filter (ROF)-based WTA neural circuits to determine the maximum choice function values. The architecture is structured for deployment on hybrid analog/digital hardware platforms to ensure real-time operation. Results. Parallelizing category choice and resonance verification significantly reduces computational latency compared to traditional sequential Fuzzy ARTMAP iterations. The integration of rank-order filter hardware circuits facilitates rapid identification of winning categories while maintaining clustering stability during continuous EEG processing. Conclusions. The proposed parallel Fuzzy ARTMAP model eliminates sequential search bottlenecks in neural category selection, enabling high-speed processing of EEG signals. Its compatibility with low-cost analog/digital hardware provides a foundation for compact, energy-efficient, real-time biomedical monitoring devices.
Chemists do not draw molecules; they draw agreements about molecules. A hexagon with three alternating lines is not benzene, and, read literally, it is not even an adequate model, since it localizes bonds that experiment treats as equivalent. Chemists repair this discrepancy automatically and never write the repair down; a machine receives the convention, plus whatever further theory its software compiles into it. This Perspective asks what our models actually see, across structural formulas, SMILES, fingerprints, graphs, conformers, simulated ensembles and the token sequences on which chemical language models are trained. We argue that both members of the training pair are representations─one of a chemical entity, the other of an assay-conditioned observation─and interpret tautomeric ambiguity, stereochemical and solid-form identity, activity cliffs, heterogeneous bioactivity labels and generative failure as consequences of this double reduction. We propose accountability, not realism, as the operative standard: explicit entity definitions, complementary representations, assay provenance, calibrated uncertainty, prospective testing. We then make the standard operational, setting out the controlled comparisons that separate representational loss from model limitation, sparse training coverage and experimental heterogeneity, and specifying for each perturbation test the control that makes it interpretable, the criterion for an actionable failure, and the corrective action it implies.
An Embedded Real- Time System (ERTS) must com- ply with two key non- functional constraints: reaction time and energy consumption. We consider this bi- objective minimization problem for ERTSs designed with a synchronous programming language, compiled into a Timed Control Flow Graph (TCFG), and deployed over a single- core processor equipped with Dynamic Voltage and Frequency Scaling (DVFS). Due to the criticality of ERTSs, we consider the worst- case versions of these objectives, giving the bi- objective minimization problem 〈Worst- Case Reaction Time (WCRT), Worst- Case Energy Consumption (WCEC) These objectives being antagonistic, we wish to produce the Pareto front of all the possible solutions in the 〈WCRT, WCEC〉 space. To solve this problem, we rely on a geometric formal reasoning that reduces the number of candidates from exponential to polynomial, each evaluated by a WCRT/WCEC analysis. This comes at the price of optimality. We propose a DVFS algorithm that generates a non- dominated solution by assigning one of the available frequencies to each block of the TCFG. We produce the Pareto front by varying this DVFS assignment. Consisting of real- life and synthetic benchmarks, our experiments show that our DVFS algorithm ApproxTEPF is able to produce excellent Pareto fronts for each of them, containing only strong Pareto optima (but not all the possible ones).
Personal financial wellness is shaped less by raw arithmetic than by behavioural biases: loss aversion, mental accounting, and status-quo bias systematically push individuals away from welfare-maximising allocations. Existing robo-advisory and reinforcement-learning (RL) approaches model the optimisation side well but treat user behaviour as exogenous noise rather than as a structured failure mode an agent can detect and counteract. We introduce PFM-Agent, a research scaffold and design that combines (i) LangGraph state-graph orchestration with typed shared state and bounded cyclic edges, (ii) a recurrent RL policy (RecurrentPPO) that allocates a user’s budget across five canonical buckets while preserving the temporal structure of cash flows, and (iii) three behavioural-economics critic agents whose natural-language verdicts are aggregated into a consensus score that gates revisions at inference time and shapes the reward at training time. Operating over the public IBM TabFormer dataset (24M synthetic transactions, 2k users), we formalise the problem as a partially observable MDP with five-simplex actions, report scaffold-validation results establishing correct end-to-end execution of the environment, reward, and compiled graph, and report a preliminary PPO/RecurrentPPO training probe whose main outcome is a reward-calibration finding rather than a hypothesis test. The full empirical study is deferred, and we preregister three hypotheses for it covering recurrence, critic-shaped reward, and interpretability.
We introduce astrograd, a Python astrodynamics library that combines automatic differenti-ation, hardware-accelerated parallel execution and embedded code generation with modelsspecific to Guidance, Navigation and Control applications. The library is built around asingle idea: assemble a differentiable simulator once from reusable modules, then run itunmodified across many numerical computing backends. Simulators built with astrogradcan be coupled with deep-learning components and executed with massive parallelismon machine-learning frameworks such as JAX, PyTorch or TensorFlow. The same modelswork directly with CasADi for optimal control and estimation, or can be compiled to em-bedded code for flight processors. astrograd ships with validated modules for the orbitalenvironment, rigid body dynamics, ephemerides, frame and time conversions, numericalintegration and finite-state machines needed to build sophisticated spacecraft simulators forcontrol system development and verification. astrograd is released under a permissive open-source license for the aerospace community, together with representative examples andtutorials covering hardware accelerated robust satellite control design, frequency-domainanalysis, re-entry corridor analysis, uncertainty propagation, trajectory optimization andembedded code generation.
Large Language Models (LLMs) have revolutionized software engineering, yet applying them to High-Performance Computing (HPC) kernel generation remains a distinct challenge due to the stringent requirements for functional correctness and execution efficiency. This paper presents a comprehensive survey of LLM-driven kernel generation and optimization methodologies emerging from 2024 to 2026. We categorize existing approaches into two paradigms: Training-Free methods that leverage agentic workflows, in-context learning, and compiler/profiler feedback, and Training-Based methods utilizing Supervised Fine-Tuning (SFT) and Deep Reinforcement Learning (DRL) for domain adaptation. Furthermore, we conduct a data-driven meta-analysis across diverse benchmarks (e.g., KernelBench, Triton-Bench) and hardware platforms (NVIDIA GPUs, Ascend NPUs).
Trojaning machine learning models is a well-known threat that introduces unwanted behaviors beyond the expectation, e.g., making incorrect predictions when an adversary-chosen trigger is present in the input. Recently, this threat has been extended to privacy, where adversaries trojan models to leak information about training data rather than manipulate predictions. The research question we study falls within this emerging line of privacy trojaning, but targets a fundamentally different objective: Private Input Reconstruction, i.e., whether an adversary can trojan an image-to-image (I2I) generative model so that a user's private input image can be reconstructed purely from the output, i.e., the generated image. To our knowledge, this is a new attack surface that no one has studied before, and cannot be solved by existing approaches. On one hand, although there are many training data reconstruction attacks, test-time input data reconstruction without trojaning is inherently challenging, i.e., no prior works have succeeded in doing so with high quality. On the other hand, while steganography is a promising approach to embed data into the generated images, the threat model between Private Input Reconstruction and steganography is fundamentally different. Steganography assumes that the sender and the receiver share full knowledge, whereas our adversary in the trojan attack operates with zero shared knowledge with the model user and should recover the input from the output alone. Furthermore, how to embed steganography pipelines into the generative model is also challenging. In this paper, we design the first trojaning framework, called PIRecon, which compromises the Variational Autoencoder (VAE) of a diffusion-based I2I pipeline to covertly encode each private input image into its generative output. The trojaned pipeline operates normally during inference, generating synthetic outputs that are perceptually indistinguishable from those of a clean model, while carrying private input information. An adversary-controlled decoder then reconstructs the private input from the collected synthetic image. Our key insight is to first fine-tune the VAE to embed and recover a bitstream representation of each query and then enforce visual realism to preserve the perceptual fidelity of the outputs. Despite the threat model differences, PIRecon still outperforms all steganographic baselines---assuming that they also embed the input data---in terms of reconstruction quality and fidelity on three datasets. Notably, an Institutional Review Board (IRB)-approved clinician study on a private clinical dermatology dataset further confirms the visual utility of the trojaned outputs and reconstructed inputs.
Flood Early Warning Systems (FEWSs) increasingly combine heterogeneous observation streams, computational models, and interoperable information infrastructures. This survey reviews how Internet of Things (IoT)-based sensing, communication protocols, data pre-processing, machine learning, data assimilation, and operational platforms contribute to flood monitoring and early warning. Rather than presenting a new forecasting model or claiming end-to-end experimental validation, the paper synthesizes cross-layer integration patterns between observation networks, computational methods, uncertainty-aware decision-support, and standards-based middleware. Particular attention is given to the interface between in situ IoT observations and data-driven or physics-informed modelling approaches, including data assimilation and hyper-local flood simulation. A FIWARE/NGSI-LD hydrometeorological monitoring deployment is included as an infrastructure and interoperability case study, demonstrating the continuous collection and ingestion of sensor observations into the platform, semantic normalization, temporal persistence, and standards-based access to monitoring data.
The MEDUSA project (MEDiterraneo: Uso Sostenibile dell'Ambiente), funded by the Italian Space Agency (ASI) under the PNRR framework, develops a multi-scale dynamic Digital Twin (DT) of the Mediterranean coastal ecosystem to support sustainable decision-making in vulnerable areas, such as the Amalfi Coast. The platform leverages a hybrid cloud architecture combining a high-performance local NAS with cloud services (ArcGIS Online), balancing data security and national jurisdiction with remote processing scalability. Workflows integrate and homogenize multi-source data—including satellite imagery (COSMO-SkyMed, PRISMA, Sentinel-2/3), UAV LiDAR and photogrammetry, bathymetry, regional geodatabases, and real-time IoT feeds. These datasets serve both to feed the five dynamic models implemented within the DT and to construct the baseline cartography and thematic layers of the study area. The infrastructure powers five dynamic monitoring applications: structural vulnerability assessment via PS-InSAR, water quality estimation from PRISMA hyperspectral data, coastline extraction from SAR time series, atmospheric corrosion mapping, and IoT-enabled microclimate monitoring for heritage sites. Validation across the Amalfi Coast test site confirms the high fidelity of the implemented pipelines, with sub-pixel coregistration accuracy, millimeter-scale displacement measurements, and strong machine-learning predictive performance for water quality parameters. Beyond its technological architecture, the platform connects real-time environmental monitoring to decision-making for coastal protection and is designed to provide public administrations with a unified access point to validated environmental indicators and predictive scenarios for climate adaptation and risk mitigation. The MEDUSA DT thereby also serves as a Third Mission initiative dedicated to knowledge transfer for public administration and society.
Hardware fault injection attacks threaten machine learning systems, yet their impact on reinforcement learning (RL) has received little attention despite the use of RL in safety-critical applications. This paper presents a security analysis of Q-Learning implementations under voltage and clock glitching attacks on a ChipWhisperer platform, evaluating five variants: Standard, Robust, Resilient, Linear, and Deep Q-Learning. In standard Q-Table implementations, faults corrupt reward computation, action selection, and Q-value updates during learning, and change the selected action during inference. Robust Q-Learning does not prevent these attacks: NaN and infinite values injected during learning persist in the Q-table, and its greedy policy uses the same code as standard Q-Learning. This paper also evaluates a Resilient Q-Learning variant that applies triple modular redundancy with majority voting. It corrects corruption of a single Q-table copy, but inputs shared by the copies, such as the reward, and the voting code itself remain outside its protection. The analysis extends to Linear and Deep Q-Learning, where faults corrupt stored weights during learning and change actions during inference. Moving beyond benchmark environments, this work evaluates an online traffic light controller in a microscopic traffic simulator, where a single injected infinite value degrades accumulated vehicle waiting time by 148x while corrupting under 4% of the Q-table, and where triple modular redundancy corrected the faults injected in our tests, although glitching could still halt the device. Altogether, this work provides a basis for understanding and securing RL deployments against fault injection attacks.
The increasing complexity of Internet of Things (IoT) environments necessitates efficient and reliable anomaly detection. This study experimentally evaluates established tree-based machine-learning models instead of proposing a new detection algorithm. Decision Tree (DT), Random Forest (RF), and XGBoost are compared using CPU- and GPU-based implementations on the highly imbalanced DS2OS dataset by considering training time, overall classification performance, and minority-class detection. GPU experiments were conducted using an NVIDIA T4 in a cloud-based environment; the results do not establish performance on physical IoT or edge devices. The recorded training times yielded CPU-to-GPU time ratios ranging from 2.12 to 2.94 across the evaluated implementations. However, the original CPU hardware configuration could not be verified, preventing these differences from being attributed exclusively to GPU acceleration and limiting their reproducibility under equivalent conditions. In the initial evaluation, Random Forest configurations failed to detect minority attack classes despite high overall accuracy, whereas supplementary evaluations demonstrated substantially improved performance with alternative feature encoding. XGBoost achieved near-perfect classification in the initial test partition and reduced-training configurations. Although supplementary evaluations initially showed lower and more variable performance with the complete-training partition under automatic initialisation, a controlled ten-fold evaluation demonstrated that fixing the initial prediction score eliminated this deterioration on both the CPU and GPU within the evaluated DS2OS configurations. Supplementary experiments further demonstrated that the GPU-based Decision Tree proxy achieved classification performance comparable to the CPU Decision Tree following adjustments to its model configuration, indicating that the original performance difference was primarily configuration-dependent rather than an inherent limitation of GPU hardware. Energy efficiency and embedded-device performances were not experimentally established.
The rapid growth of digital crime through cyber intrusions, ransomware, social engineering, and cloud exploitation has substantially increased the volume, heterogeneity, and complexity of digital evidence. Modern investigations span mobile devices, operating systems, volatile memory, networks, cloud infrastructures, and emerging environments such as the Internet of Things (IoT) and blockchain, challenging the scalability of conventional forensic workflows. Artificial Intelligence (AI) and Machine Learning (ML) offer promising capabilities to automate evidence analysis, identify complex patterns, and accelerate investigative processes. This systematic review provides a comprehensive and critical synthesis of AI/ML applications across major digital forensic domains, including file, memory, network, cloud, IoT, blockchain, mobile, social media, multimedia, biometric, vehicular, and drone forensics, as well as cross-cutting investigative tasks. A structured literature search and eligibility assessment identified 76 studies for qualitative synthesis. The review synthesizes the literature with attention to dataset availability, validation practices, reproducibility, explainability, adversarial robustness, and forensic admissibility. Beyond descriptive synthesis, this study introduces a six-dimensional task maturity framework based on observability, labelability, stationarity, time-to-insight, explainability, and adversarial robustness. An evidence audit links individual ratings to supporting references and qualifications. The framework assigns literature-informed ordinal scores using explicit dimensional criteria. Supporting references, rationales, and scope limitations are documented for all 96 task-dimension ratings. Sensitivity analysis examines the same profiles under alternative weights, thresholds, and individual rating changes. The reference mapping does not independently establish the validity of each numerical level. Contextual and interpretive judgment remains constrained by the need for human situational and legal reasoning. The review further identifies persistent gaps in benchmark datasets, real-world validation, explainability, adversarial testing, and legal defensibility. These findings motivate a research agenda centered on standardized evaluation protocols, robust and explainable forensic AI, improved benchmark resources, and accountable human–AI collaboration. Overall, this study provides a structured assessment of the current capabilities and limitations of AI/ML in digital forensics and identifies priorities for developing reliable, transparent, and legally defensible intelligent forensic systems.
In recent years, tourism and hospitality has been transformed by the rapid development of artificial intelligence (AI), the Internet of Things (IoT), and increasingly accurate machine learning tools (Law et al [...]
We describe a simple yet powerful wearable device containing an embedded system composed of proprietary designed hardware and firmware which is able to generate random seeds. Each seed generated after the hardware is powered on is further used by a classical pseudo-random number generator routine. The resulted pseudo-random number generator (PRNG) is stored on a laptop for testing purpose using a serial to universal serial bus (USB) interface and also truncated and displayed as dice symbols on seven ultra-low power LEDs. The random generator quality has been computed using various testing methods by acquiring a relevant number of samples of random buckets generated by hardware immediately after powering on. Statistical National Institute of Standards and Technology (NIST) and Diehard suite tests reveal the feasibility of the random seed and pseudo-random numbers generation. The electrical performance of 25 manufactured devices was also measured in ON and OFF state. The hardware has a maximum leakage current of 220 nA and draws an average of 0.8 mA at 3 V supply during displaying of ten successive dice symbols in 8 seconds, running with the same CR2032 battery for more than one year. The method used for variable seed generation does not require the use of any type of volatile or non-volatile memory. Our device and seed generation method can be used in any scientific application or toy industry requiring true random numbers generation at ultra-low power supply.
Static Random Access Memory (SRAM) takes on a vital role in digital technology especially for applications that demand high speed and low power consumption. However, there are top priority with leakage power, reliability and stability owing to continuous technological advancement and the demand for energy efficient systems. Thorough research of low power SRAM design approaches is presented with an emphasis on cutting edge cell topologies and innovative device technologies. Power efficiency, read/write stability and area overhead are examined for various SRAM designs including 6T, 7T, 8T, 9T and 10T cells. Special attention is focused on SRAM designs using Carbon Nanotube Field Effect Transistor (CNTFET) which provide excellent electrical characteristics and less leakage than conventional silicon enabled methods. Subthreshold operation and assist methods that enable ultra low voltage capability are also included in the article. To further emphasize the need of low power SRAM designs, growing application domains like Internet of Things (IoT), biomedical systems and Compute-In-Memory architectures are investigated. Major performance deals and design trends are identified through a comparative analysis of this research. In order to facilitate growth of next generation low power SRAM systems, important challenges and future research directions are finally specified.
The enhanced communication model (ECM) integrates multi-baseline delta encoding, FlagBag encoding of Boolean fields, and partial deserialization in an Apache Avro-based communication pipeline. ECM and regular Avro communication (RAC) were compared in 15 paired, constructed scenarios with 100 runs per mode and scenario. Paired ECM and RAC runs used the same set of 10,000 logical messages. ECM materialized a fixed 80% of application fields, whereas RAC fully deserialized them. The mean size of the complete transmitted container decreased by 14.98–31.73% across scenarios (mean scenario reduction: 24.41%). The duration from the first send to processing the last message of the 10,000-message batch decreased by 14.86–29.30% (mean: 22.42%). Both reductions exceeded a conditional 10% target in all 15 scenarios. The byte reduction characterizes the combined FlagBag and MDE implementation; the time reduction compares complete procedures with different amounts of deserialization. The experiment does not quantify the individual contribution of each method. Per-scenario field counts, Boolean proportions, message correlation, CPU use, and memory use were not measured in the available records; the findings cannot be generalized beyond the studied conditions without further testing.
Quantum processors are often compared by the number of physical qubits they contain. That number, however, does not by itself indicate whether a processor can execute a useful computation reliably. Qubit coherence, gate and measurement fidelity, connectivity, control infrastructure, fabrication yield, and the overhead of quantum error correction all influence practical capability. This paper presents a structured comparative review of six major quantum-computing modalities: superconducting circuits, trapped ions, photonic systems, silicon-spin qubits, neutral atoms, and topological approaches. It examines their physical implementations, operational strengths, engineering constraints, and prospects for scaling. Particular attention is given to surface-code error correction and the 105-qubit Willow processor result reported by Google Quantum AI, which demonstrated below-threshold surface-code memory and a logical-error suppression factor of 2.14 ± 0.02 when code distance increased by two. The paper also considers cryogenic control, wiring, modular interconnects, fabrication variability, real-time decoding, and the Indian National Quantum Mission. The review finds that progress towards fault tolerance is best assessed through logical performance and system-level engineering rather than physical-qubit count alone. No modality currently removes the need to address scaling overhead, reliable operations, and application-level advantage
Graph Neural Networks (GNNs) offer a powerful paradigm for financial fraud detection by exploiting the relational structure of transaction networks to identify suspicious behaviour and coordinated fraud. Translating them into a viable production system, however, remains hard under strict memory, compute and latency budget – costs that span the entire model lifecycle, not merely GNN training. A real transaction graph of hundreds of millions of accounts is overwhelmingly a graph of benign nodes, so much of that lifecycle budget is spent processing nodes that carry little discriminative signal. The same benign majority drives a second challenge, fraud-class homophily collapse: because fraudsters deliberately interact with legitimate entities to camouflage and evade detection, the proportion of same-class edges incident on fraud nodes drops and the fraud class is driven into high heterophily, so neighbourhood aggregation averages the fraud signal away while imbalance biases the classifier toward the benign majority. We propose CGGS (Confidence-Guided Graph Sampling), a two-stage framework that turns both liabilities into a single opportunity at no additional compute cost. Stage 1 trains a lightweight tabular scorer – gradient-boosted trees (XGBoost) in our instantiation – on the node features, ranking every candidate entity by fraud probability without touching graph topology. Stage 2 prunes the confidently-benign entities before the graph is assembled, so the GNN trains only on the compact, higher-homophily subgraph of uncertain, high-risk, and informative nodes. We prove that confidence-guided pruning increases the expected fraud-class homophily and reduces the effective imbalance ratio under a mild ranking assumption. Experiments on a production-scale proprietary transaction graph and two public benchmarks show that CGGS reduces graph size and lifecycle computational cost while outperforming full-graph training and alternative pruning strategies, with ablations confirming robustness across pruning thresholds.
Chemical reactions offer unique opportunities for energy-efficient and adaptive computing. We integrate fluidic networks filled with the oscillatory Belousov-Zhabotinsky (BZ) reaction in silicon microchips to investigate the potential of these networks for on-chip chemical computation. Using nanofluidic confinement and networks of thousands of microreactors (5-30 μm) connected by nanochannels, we demonstrate stable oscillations under 300 nm vertical confinement and estimate an intrinsic chemical energy expenditure of ≈17 fJ per cycle at nanoscale reaction volumes. Reaction-diffusion waves propagate signals with diffusion lengths of ≈40 μm and speeds up to 123 μm s-1, while light-sensitive modulation provides dynamic control over the reaction dynamics, offering a route toward programmable architectures. These features support large fan-out, weighted connections, and reconfigurable pathways, laying the foundation for chemical logic, neuromorphic networks, and adaptive systems. Our results position oscillatory chemistry as a promising route toward ultra-low-power computing and bio-inspired architectures.
Organic electrolyte-gated transistors (OEGTs) are promising platforms for bioelectronics and neuromorphic systems owing to their low-voltage operation and strong ionic-electronic coupling. However, in vertical OEGTs, which offer higher current densities and faster switching, the semiconductor channel is often shielded from direct electrolyte contact, limiting ionic accessibility and resulting in inefficient gate coupling and slow switching dynamics. Here, we introduce a photolithography-free ion-access strategy for vertical OEGTs, combining femtosecond laser micropatterning with inkjet-printed solid-state electrolytes. Laser-defined pores penetrating the top electrode and semiconductor layer create direct ionic pathways to the channel region, significantly enhancing ionic-electronic coupling. As a result, the devices exhibit more than two orders of magnitude enhancement in transconductance (up to 9.17 mS), over 300-fold enhancement in current density reaching 51.8 A cm-2, and an on/off ratio approaching 107, together with accelerated switching dynamics compared to devices without pore structures (>6-fold reduction in turn-on time constant). This ion-access engineering strategy is compatible with multiple organic semiconductor materials and provides an effective structural strategy for overcoming ionic transport limitations in vertical OEGTs. Moreover, array-level demonstrations and a proof-of-concept NARMA2 reservoir computing task (NMSE = 0.107) further illustrate the scalability and neuromorphic application potential of this platform.