2026-11-09

5 repos · 25 papers · 52 scholar

Showing 82 of 82 items

GitHubframeworks

BiSeNet

Add bisenetv2. My implementation of BiSeNet

Python1,643335
GitHuboptimization

iptv-proxy

IPTV reverse proxy for M3U/M3U8 playlists and Xtream Codes accounts, with HLS. One binary or Docker container.

Go749211
GitHubmodel-compression

llama-modchip

🐋 Run 284B DeepSeek-V4-Flash (<59GiB RAM, 44.4 tok/s pp, 9.1-14.7 tok/s) & 30B MoEs (25-111 tok/s) on a pure 64GB CPU at 256K ctx (0 GPUs). SwiGLU-H2, Stockfish-PVS L3 & 13ms Claude Code restore.

Python67436
GitHubai-hardware

NMC

I'm trying to recreate the NMC architecture discussed in the Overflow-free compute memories for edge AI acceleration paper

SystemVerilog00
GitHubai-hardware

floating-gate

Aspirational floating-gate arm of memristive-crossbar (02): can in-fabric learning run on a real analog floating-gate array (tunneling/injection writes)? Verdict-gated, H5-first, provisional.

Python00
arXivhardware-design

Memory as a Temporal Resource: Lifetime Contracts for Multi-Retention Systems

Computers move and rewrite data whose useful lifetimes range from inner-loop iterations to years, yet memory systems usually hide retention as a fixed physical property. We argue that required lifetime should become a first-class resource across algorithms, programming systems, controllers, and devices. Building on prior lifetime-aware placement, we propose an event-based validity contract that links application completion events to retention-class selection and renewal of protection. Applications would be enabled to express semantic contracts such as ``until this tile completes'' and the system will map data to homogeneous arrays engineered for ephemeral, epoch, or durable retention and preserve correctness through margins, ECC, refresh, migration, or recomputation. We present a trace-based simulator and FPGA prototype that exercises class migration.

Adrian JacksonPDF
arXivaccelerators

Atik: A Shared-Mesh Transformer Accelerator for RISC-V

Transformer inference repeatedly invokes dense matrix multiplication and attention, including its data-dependent softmax normalization. We present Atik, an open-source RoCC accelerator for RocketChip that executes matrix multiplication and dense attention through one shared fixed-point MAC mesh. Atik accepts and returns BF16 tensors, uses a streaming online-softmax schedule that never materializes the full score matrix, and is invoked through a small descriptor-based ABI. We evaluate three meshes cycle-by-cycle on FireSim and report standalone Sky130 physical-design area. Across 137 logged FireSim test records, all tests pass their benchmark tolerance with zero reported output mismatches. Relative to the same scalar Rocket BF16 reference, the 8x8 configuration reaches 31.5x peak matmul speedup and 225.8x peak attention speedup. Hybrid projections for the single-layer workload-derived C models, using directly measured accelerator-kernel cycles, give aggregate speedups of 44.2x, 77.1x, and 32.3x for ViT, TinyBERT, and GPT-2 prefill, respectively; these are not full-model hardware replays.

Ahmed Zeer +1PDF
arXivmodel-compression

BitDecoding-FP4: Native FP4 KV Cache Decoding via Algorithm-System Co-Design

The key-value (KV) cache dominates memory capacity and bandwidth in long-context LLM serving. While 4-bit microscaling formats (NVFP4, MXFP4) shrink the cache and run attention on FP4 Tensor Cores, a fundamental gap remains between algorithmic quantization and system-aware acceleration: existing FP4 KV systems use vanilla quantization that causes severe accuracy drops, whereas accuracy-oriented 4-bit algorithms rely on vector codebooks whose lookup overhead degrades decoding throughput. We present BitDecoding-FP4, an algorithm--system co-design for accurate and efficient native FP4 KV decoding. We mitigate three error sources: Dual-Residual Query Quantization splits queries into two FP4 tensors, cutting query logit error from 60% to 1%; Underflow-Aware Probability Scaling prevents E4M3 scales from rounding long-context probabilities to zero; and Outlier-Aware Key Quantization combines static smoothing, Hadamard rotation, and scale search to suppress key outliers under token-wise scaling. At the system level, we remove the softmax bottleneck via transposed softmax with fused probability quantization and stacked query MMA on paged KV caches. Across dense and MoE models with GQA and MLA, BitDecoding-FP4 stays within 0.5 points of BF16 KV accuracy with NVFP4 (1 point with MXFP4), and tracks BF16 KV in RL training rollouts. In end-to-end serving on NVIDIA B300 with Qwen3-4B and Qwen3-30B-A3B, it delivers 2.4-3.0$\times$ higher peak serving throughput than BF16 KV. The implementation is available at https://github.com/OpenBitSys/BitDecoding-FP4.

Dayou Du +5PDF
arXivother

Link-Level Performance Comparison of 5G Beamforming: Reciprocity-Based SVD vs. Codebook-Based Precoding

Downlink (DL) precoding in mid-band MIMO relies on uplink (UL) sounding reference signals (SRS) with reciprocity-based singular value decomposition (SVD) or channel state information reference signals (CSI-RS) with codebook-based precoding matrix indicator (PMI) feedback. Although SVD precoding offers unquantized beamforming, it depends on a channel estimate from SRS transmitted at the limited user equipment (UE) power, whereas Type-I PMI is quantized but derived from a high-power CSI-RS based channel estimate. In this paper, we compare SRS-SVD and CSIRS-PMI for single-user MIMO through an end-to-end, 3GPP-compliant link-level simulation at 3.75 GHz with 100 MHz bandwidth, using a 192 antenna element gNB mapped into 32 ports, the 3GPP Release 19 UE antenna model, and ray-traced (RT) channels in Manhattan and Chicago. We show that neither scheme is uniformly superior: CSIRS-PMI serves weak and shadowed users better and reduces outage, whereas SRS-SVD achieves higher spectral efficiency (SE) and enables more multi-layer transmission for users with strong channels. The results on the Clustered Delay Line (CDL)-C statistical channel also confirm this crossover. We further show that by utilizing a simple logistic regression based on the received power of the standard CSI reference signal (CSI-RSRP) and the channel quality indicator (CQI), the gNB can select the optimal transmission scheme with around 90% accuracy, achieving an SE that closely approaches an ideal per-user selection.

S. M. Haider Ali Shuvo +2PDF
arXivaccelerators

FOSCAR: Accelerating Decode Attention over 2- and 4-Bit KV Caches through Affine Folding

Large-language-model decode attention streams the key-value (KV) cache for every generated token. Packing that cache into 2-bit or 4-bit quantized integers reduces memory traffic but does not automatically translate into latency saving. Existing 2-bit fused kernels reach under 31%/15% of peak HBM throughput on H100/B200. Converting each value and applying its scale and offset before feeding Tensor Cores is a major bottleneck because instruction-issue throughput is scarce on modern GPUs. Instead of applying scale and offset to each KV value, FOSCAR uses affine folding to apply them to the other operand or the output of decode-attention matrix products, depending on the quantization scheme (token-wise or channel-wise). This reduces scale-and-offset operations by up to two orders of magnitude while preserving the dequantized attention output in exact arithmetic. FOSCAR also uses bit manipulation to feed KV operands to Tensor Cores efficiently, requiring only 1 and 0.19 ALU instructions per INT2 element on H100 and B200, respectively. Finally, FOSCAR builds warp-specialized GPU pipelines for Hopper (H100) and Blackwell (B200) using wgmma and tcgen05. On H100, FOSCAR INT2 and INT4 kernels achieve geo-mean speedups of $3.36\times$ and $2.83\times$ over FlashInfer BF16, and $2.03\times$ and $1.71\times$ over TensorRT-LLM FP8. On B200, the corresponding INT2 and INT4 speedups are $2.67\times$ and $2.30\times$ over FlashInfer BF16, $1.48\times$ and $1.28\times$ over TensorRT-LLM FP8, and $1.16\times$ and $1.00\times$ over TensorRT-LLM NVFP4. We integrate FOSCAR into SGLang to benchmark serving efficiency. On long-context serving workloads, FOSCAR increases peak aggregate throughput by up to 170% and 341% over 2-bit and 16-bit baselines, respectively. We open-source FOSCAR as a kernel library at https://github.com/FutureMLS-Lab/FOSCAR.

Donglin Zhuang +4PDF
arXivother

Concurrent Use of FP32 and FP64 Computing Units on GPUs and Mixed-Precision Placement in WENO Reconstruction

Scientific computing often requires FP64 arithmetic, although FP64 performance is limited on NVIDIA GPUs. Double-float (DF) emulation, which represents a value using two FP32 values with about 48 bits of significand, outperforms native FP64 where FP64 units are limited. On the other hand, mixed-precision strategies offload to FP32 the operations that do not determine accuracy. In both approaches, the native FP64 units remain idle during FP32 instruction execution. We investigated whether this idle resource can be recovered on A100, GH200, GB200, and RTX PRO 6000 Blackwell Max-Q. A microbenchmark shows that co-resident FP32 and native-FP64 kernels exceed the throughput of the faster kernel alone on A100 but not on GH200 and GB200; on RTX 6000 Blackwell, they do so against FFMA but, against DF emulation, only when the issue density of the DF kernel is reduced. We then tested whether this gain carries over to computing the WENO-Z reconstruction with the nonlinear weights in FP32 and the candidate polynomials in native FP64 (W32P64). We observe substantial overlap on A100, whereas GH200, GB200, and RTX 6000 Blackwell show little overlap. Importantly, the fastest placement does not always preserve the design accuracy of WENO-Z: for a one-dimensional entropy wave problem, W32P64 fails to achieve the design order of accuracy for WENO7 and WENO9. In contrast, DF emulation of the entire reconstruction comes close to it on the grids tested for WENO5, WENO7, and WENO9 while delivering a $2.51-3.94 \times$ speedup over FP64 on RTX 6000 Blackwell. For WENO5 on the two-dimensional Euler vortex problem the W32P64 error at $n=256$ is $1.26 \times$ that of FP64 and 1/104 of that of the all-FP32 reconstruction.

Jun Hatayama +4PDF
arXivmodel-compression

StockASkill: A Local-First AgentSkill for Intelligent Stock Selection

Quantitative equity selection and portfolio construction in emerging and cross-border markets frequently suffer from computational bottlenecks, remote API fragility, survivorship bias, and rigid execution constraints such as discrete board lots and investor-mandated drawdown limits. In this paper, we present StockASkill, a local-first, full-data-driven AgentSkill engineered as a specialized Quant Co-Processor for autonomous AI agents and quantitative researchers. Tailored for agile portfolios (10^5 ~ 10^7 RMB) where discrete lot execution and capital protection are mandatory, StockASkill embeds a columnar DuckDB OLAP engine over 42.44 million historical price records and 186,432 financial disclosures spanning 26 years across China A-shares, Hong Kong, US equities, and ETFs. By eliminating ad-hoc network dependencies during analysis, StockASkill reduces cross-sectional screening latency across 5,400+ equities from over 25 seconds to sub-0.5 seconds (a 95%+ reduction). We formalize an end-to-end mathematical pipeline integrating a 7+1 factor alpha model, an SLSQP convex optimizer under smooth CVaR drawdown surrogates with exact path projection, dynamic 100-share board lot integer quantization, and an autonomous thesis lifecycle finite-state machine. Furthermore, we provide a zero-disk, strongly-typed JSON streaming interface designed for high-throughput LLM tool-calling. Open-sourced at https://github.com/axjing/stockaskill, comprehensive empirical benchmarks from 2018 to 2024 demonstrate that the architecture achieves annualized Sharpe ratios exceeding 1.48 and Calmar ratios up to 2.17, while reducing max drawdown by up to 85% relative to benchmark indices.

Xiangjing An +2PDF
arXivmodel-compression

e-means: outperforming k-means (in one-tenth the epochs)

Lloyd's k-means algorithm is a long-running standard for clustering, with applications in data analysis, retrieval, quantization and data curation. In this paper, we introduce an exponential moving average optimization procedure for k-means with mini-batch updates, powered by modern deep learning machinery and new insights. We also establish scaling laws on how decay rate and batch size impact final performance. This gives us a formulaic rule to fix them for multiple problem sizes and embedding datasets, without the need for any hyperparameter sweep. The resulting clustering algorithm, coined e-means, is much faster than Lloyd k-means and converges to lower losses, frequently even beating k-means++. On a wide array of embeddings datasets, it beats the loss of Lloyd k-means at 25 epochs in 1.4 to 2.6 epochs, which translates to 10-45x GPU wall-clock speedups compared to state-of-the-art k-means implementations. As an example, e-means partitions DINO10B, a dataset of 10 billion DINO embeddings, into 16 million clusters in only 2.5 epochs, reaching a loss below Lloyd's at 25 epochs.

Swann Bessa +6PDF
arXivmodel-compression

MINE: Entropy-coded Memory-efficient Inference for Large Language Models

We present MINE (Memory-efficient INference Engine), a lossless entropy coding approach that accelerates single-stream inference of Large Language Models (LLMs), where existing lossless compression methods yield speedups only for specific data types such as \texttt{BF16} and under additional assumptions. Single-stream decoding is strongly memory-bound, since every weight must be streamed from memory once per generated token and requests cannot be batched in local or low-latency deployments. At the core of MINE is dtANS-LLM, a GPU entropy coding scheme adapted from the prior dtANS to the needs of LLM inference. It exploits the non-uniform distribution of weights, which persists even after quantization, and exposes hyper-parameters that trade compression ratio against decoding cost. We auto-tune these parameters for each tensor and GPU to obtain the Pareto front between model size and decoding latency. MINE integrates into vLLM as a plugin and selects the best kernel configuration for a given memory budget. On Llama~3.1~8B~Instruct with a 2.1-bit EntQuant checkpoint, MINE achieves speedups of 2.52$\times$ and 1.48$\times$ over the fastest baseline on an RTX~4060~Ti and an RTX~5090, at effective bitrates of 2.24 and 3.43~bits per weight. MINE is available as open source.

Emil Schätzle +1PDF
arXivmodel-compression

HiMEM: Hippocampus-Inspired Memory Consolidation, Pruning, and Composition for Vision-Language-Action Models in 3D Scenes

Memory plays a crucial role in long-horizon task execution and cross-task generalization for Vision-Language-Action (VLA) models operating in complex 3D scenes. However, existing VLA memory mechanisms suffer from memory noise, particularly in complex 3D scenes, as stored episodic memories contain extensive redundant details about scene layouts, object configurations, and robot-specific states, which obscure generalizable manipulation principles and hinder cross-task transfer. To address this challenge, we first introduce a hippocampus-inspired memory consolidation mechanism that synthesizes extensive episodic experiences into a compact set of essential semantic memories, enabling VLA models to derive generalizable principles from practical experience. Second, we propose an episodic memory pruning mechanism that removes redundant episodic memories whose contents are captured by consolidated semantic memories, thereby increasing effective memory capacity and improving long-horizon task execution. Finally, to handle complex task scenarios, we develop a flexible memory composition mechanism that jointly leverages multiple episodic and semantic memories to enhance cross-scenario generalization. Experiments on LIBERO, LIBERO-Plus demonstrate the effectiveness of HiMEM across different VLA backbones, improving the average success rate of pi_0 from 91.88% to 98.20% on LIBERO.

Boran Zhao +6PDF
arXivaccelerators

Robust Deep Prototype Classifiers via Lipschitz Training

Robustness and interpretability are interrelated and desirable properties of machine learning models. A model that changes its prediction under imperceptible perturbations behaves inconsistently with human-recognizable features and, hence, is not interpretable. Prototype-based classifiers such as Generalized Learning Vector Quantization address both at once. They are interpretable by construction because they make the decision by comparing the input to the prototypical example; and their hypothesis margin yields a closed form robustness certificate. However, on high-dimensional data, these shallow models need deep embedding networks to achieve high accuracy, and this voids the robustness certificate. The margin is a statement about distances in the embedding space, and an unconstrained network can map an arbitrarily small input perturbation to arbitrarily large attacks in the embedding space. We close this gap with a theorem: if the embedding network is locally L-Lipschitz around a point, and the classifier is ε-robust around its embedding, then the full model is ε/L-robust in the input space. Based on this, we propose a Lipschitz training scheme to achieve L-Lipschitz properties for embedding networks. We evaluate three classification heads (linear, linear Support Vector Machine, and GLVQ) crossed with two embedding networks (standard and Lipschitz-constrained) on MNIST, CIFAR-10, and Imagenette-160, under FGS, PGD and C\&W attacks. The evaluation results show that combining a prototype-based head with a Lipschitz-constrained embedding provides the strongest robustness. This suggests that the Lipschitz constraint provides an additional robustness benefit in addition to the advantage provided by the prototype-based head. Since the Lipschitz constraint is enforced by sampling rather than by construction, the reported robustness is empirical rather than certified.

Adia Khalid Khan +2PDF
arXivaccelerators

An Empirical Study of Issues in GPU Kernel Libraries for LLM Systems

Behind the expanding use of large language models (LLMs) across application domains lies a growing dependence on specialized GPU kernel libraries. These libraries underpin the correctness and efficiency of training and inference across diverse GPU platforms, implementing attention, GEMM, quantization, and low-precision operators while managing tensor shapes, memory layouts, code generation, and hardware-specific dispatch. Problems at this layer can silently corrupt outputs, hang GPUs, prevent compilation, introduce deployment incompatibilities, or degrade performance, thereby reducing developer productivity and wasting expensive GPU resources. We present the first comprehensive empirical study of issues in GPU kernel libraries for LLM systems. We manually analyzed 300 high-quality GitHub issue reports, their discussions, and available fixes from seven representative libraries: FlashInfer, FlashAttention, Triton, CUTLASS, DeepGEMM, xFormers, and bitsandbytes. The analysis characterizes symptoms, root causes, fixes, and testing and debugging practices. Our main findings are: (1) 60.33% of the sampled issues originate in the Compiler and Build System or Runtime Environment dimensions, so validation must cover the deployment and execution stack, not just kernel code. (2) Kernel specialization imposes implicit correctness constraints on tensor shapes, metadata layouts, memory aliasing, and hardware resources. Violations can produce incorrect outputs or launch failures even for supported operations, highlighting the need for explicit contract validation and boundary-case testing. (3) Current testing and debugging practices are often insufficient for identifying low-precision numerical problems and hardware-specific execution failures. We outline research opportunities and tooling directions to support future LLM GPU kernel development.

Yanjie Gao +3PDF
arXivaccelerators

Fused-Softmax for efficient quantized attention

Low-precision matrix throughput on modern GPUs is growing far faster than vector-arithmetic throughput. Attention therefore cannot realize the available GEMM performance because softmax has become the pipeline bottleneck. We introduce Fused-Softmax Attention, a training-free algorithm that reformulates and approximates softmax to fold its dominant operations into the attention GEMMs, moving work from vector units to matrix cores. Implemented for the AMD CDNA 4 architecture, our MXFP8, MXFP6 and MXFP4 kernels reach approximately 3.0, 4.0 and 4.6 PFLOP/s on an AMD Instinct MI355X GPU, yielding speedups of $2.1\times$, $2.8\times$ and $3.2\times$ over the BF16 attention. Importantly, the additional error introduced by Fused-Softmax remains below the error floor set by quantizing the GEMM operands. On Wan2.2-T2V-A14B, a hybrid MXFP6/4 schedule delivers up to $2.05\times$ end-to-end speedup while preserving the visual quality of the outputs. The kernels ship in AMD's open-source AITER library, and they powered AMD's MLPerf Inference v6.1 Wan2.2 submission, which ran $1.7\times$ faster than AMD's previous-round result on the same hardware. Our results show that redesigning softmax to utilize the matrix pipeline enables attention to capitalize on continued growth in low-precision matrix throughput.

Tianxing Wu +2PDF
arXivoptimization

MIRA: Mitigating Pole Geometry Bias in Recurrent Language Models

We show that the Transformer's recurrent successors --- such as selective state-space models (SSMs), Receptance Weighted Key Value (RWKV), and Retentive Network (RetNet) --- can be characterized as infinite-impulse-response (IIR) systems. The zero--pole geometries of these IIR systems characterize the temporal modes they can represent. However, to trade parallelism, the algebraic structures of the parameterizations of these recurrent systems often give rise to poles constrained to the real axis, thereby inducing an architectural bias in their recurrent dynamics. To unify these recurrent architectures and eliminate their pole geometry bias, inspired by the lattice--ladder realization of general IIR systems, we present \textsc{Mira} (\textbf{M}odal \textbf{I}IR \textbf{R}ecurrent \textbf{A}rchitecture), a unified family of linear-time recurrent architectures for building large language models (LLMs) with a highly parallel algebraic structure. Each \textsc{Mira} block realizes a higher-order IIR system that represents temporal modes from token streams through a lattice--ladder structure, followed by a feed-forward network (FFN). This realization enables the model to represent the temporal modes associated with both real and complex-conjugate poles. We pretrain \textsc{Mira-S}, a 12-layer small-scale LLM with 0.14B parameters composed entirely of \textsc{Mira} blocks. Throughout the pretraining, \textsc{Mira-S} achieves lower perplexity than strong recurrent baselines, exhibits stronger context-window generalization beyond the training length, and attains higher throughput with a dedicated CUDA kernel.

Róisín Luo +4PDF
arXivaccelerators

More with Less: Efficient Parallel Scientific Code Translation via an Agent-Skill Workflow

Translating parallel scientific codes between HPC programming frameworks is a critical but labor-intensive task. Existing LLM-based approaches rely on fixed sequential pipelines with undifferentiated repair strategies and correctness criteria that do not match scientific computing practice. We present an agent-skill framework that separates translation, compilation diagnosis, and execution validation into three autonomous agents coordinated by a verdict-driven orchestrator, routing each failure to a targeted repair skill rather than retrying blindly. An automated output check drives the self-correction loop, complemented by an offline LLM-based verification that exposes false positives escaping compiler and runtime checks. We evaluate the framework on 30 HeCBench kernels translated bidirectionally among CUDA, OpenMP, and SYCL with three open-source LLMs. Our framework attains the higher verified correctness rate in all 18 model and direction combinations over a state-of-the-art baseline while consuming up to 52.4% fewer tokens, remaining consistent across models and directions without architectural modification.

Jianuo Shen +2PDF
arXivframeworks

Transformer-FSQ: Discrete Latents for Extreme KV-Cache Compression

Private, personalised language models must run on local consumer devices, where large, frontier models cannot fit, and retaining long contexts is non-trivial. Sequential memory---the key-value (KV) cache---grows linearly with context length, quickly dwarfing the model itself: for Qwen3-1.7B-Base, 200,000 retained entries require 21 GiB, 6.6 times more than the model weights. Yet the KV-cache compression methods we measured fail beyond moderate compression ratios. We introduce Transformer-FSQ, addressing extreme KV-cache compression by replacing standard key and value projections with a finite scalar quantisation bottleneck. Our cache stores a packed vector of discrete integer codes per token, with no per-token quantisation metadata. Converting a pretrained transformer trains only these projections on 0.25 tokens per original parameter. We evaluate Transformer-FSQ against five baseline families on models from 0.6B to 8B parameters. On long-context retrieval (RULER), from 12$\times$ to 97.5$\times$ Transformer-FSQ is the most accurate method we evaluate, across scales. A fused Triton kernel decodes directly from the integer codes; with baselines given their authors' kernels where available, it decodes at $1.01\times$ (95% interval $0.98$--$1.05$) the throughput of native SDPA attention on the uncompressed Qwen3-8B-Base at $32\times$ and $1.54\times$ at $97.5\times$. Extreme KV-cache compression is therefore a practical way to extend how much context a fixed device can hold.

Andras Szecsenyi +4PDF
arXivaccelerators

HMMF-Net: A Lightweight Hierarchical Multi-Frame Multimodal Fusion Network for Accurate Fish Feeding Intensity Assessment

Automated fish feeding intensity assessment (FFIA) is important for precision aquaculture and feed cost optimization. However, unimodal methods are vulnerable to acoustic noise and visual disturbances, while existing multimodal frameworks often rely on computationally intensive 3D convolutions or Transformers, limiting their use on edge devices. To address these limitations, we propose HMMF-Net, a lightweight multimodal architecture that combines early multi-frame visual encoding and multi-axis acoustic modeling through a hierarchical bidirectional messenger cross-modal fusion (BMCF) mechanism. The visual branch captures temporal variations in feeding behavior by stacking temporally distant frames as a multi-channel input to an adapted 2D EfficientNetB0 backbone. The acoustic branch employs adaptive multi-axis residual units (AMARU) to extract temporal, frequency, and local time-frequency representations from log-mel spectrograms. BMCF progressively exchanges information between the two modalities at intermediate representation levels through compact messenger tokens, reducing the computational overhead of dense cross-modal attention. Evaluated on the AV-FFIA dataset using five-fold cross-validation, HMMF-Net achieves 96.77% accuracy and 96.88% F1-score on the held-out test folds, with 3.987M parameters and 1.002 GFLOPs. Compared with the best-performing visual configuration (95.49% accuracy), HMMF-Net reduces the classification error by 28.38%. Deployment on an NVIDIA Jetson Orin Nano using TensorRT with FP32 precision maintains comparable accuracy, with an inference time of 11.54 ms per 2-second sample and a model size of 18.81 MB. These results demonstrate that HMMF-Net combines effective multimodal representation learning with computational efficiency, enabling accurate and practical FFIA on resource-constrained edge platforms.

Phi-Hung Hoang +3PDF
arXivframeworks

ISER-MoE: Dynamic In-Switch Replication of Hot Experts for Load Balancing in Expert-Parallel

Expert-parallel mixture-of-experts (MoE) layers run at the pace of the most loaded GPU, and a few hot experts carry most of the excess load. Replicating them at every MoE layer balances the load only if their weights arrive within the layer, and broadcasting from the owner is the bottleneck: point-to-point copies overload it, and baseline broadcasts pay latency or bandwidth overheads. We present ISER-MoE (In-Switch Expert Replication for MoE), built in TensorRT-LLM, which multicasts hot experts through the NVLink switch to fixed GPU groups, so each byte leaves its owner once. A GPU scheduler assigns hot experts to groups per layer, and a multi-port broadcast spreads the injection over the group's ports, both overlapped with the shared expert. In trace replay, the scheduler removed 98-99% of the excess load; in prefill serving of DeepSeek-V4-Pro, throughput rose 17.4% on a math benchmark and 7.2% on agentic prompts.

Jintao Peng +7PDF
arXivaccelerators

UniExit: A Unified Latency-Aware Early-Exit Runtime for Heterogeneous Edge AI

Edge applications increasingly combine CNN-based perception with edge-deployable Transformer-based language models, but deploying such heterogeneous workloads on resource-constrained devices is challenging because inference latency varies with both input difficulty and runtime hardware state. Early-exit inference reduces computation by terminating at intermediate stages, yet existing methods are often model-specific and rely on static latency profiles that become stale under dynamic edge conditions. This paper presents UniExit, a unified latency-aware early-exit runtime for heterogeneous edge AI. UniExit abstracts CNN blocks and Transformer layers as ordered stages and applies a common latency-aware routing policy while retaining model-specific quality mechanisms. It predicts the cost of reaching the next exit from runtime state and inference progress to make budget-aware continue-or-exit decisions. UniExit combines runtime-state-aware latency prediction, cache-based prediction reuse, trend-aware fallback, and CNN-specific exit stabilization to provide low-overhead and robust online control. We implement UniExit on multiple edge and edge-cloud platforms and evaluate it with representative CNN and Transformer-based language-model workloads. Across single-workload, concurrent heterogeneous, and edge-cloud settings, UniExit consistently improves deadline satisfaction and latency-accuracy trade-offs over existing progressive inference baselines, while reducing redundant exit-module computation and task-level energy consumption.

Yuhang Guan +3PDF
arXivai-hardware

A Novel Algorithm to Simulate Conductance-based Neuronal Networks

Conductance-based models are central to neurophysiology and neuromorphic engineering, yet their simulation scales poorly to large networks. We study and analyze an iterative solver that operates in signal space. Each iteration reduces to $N$ decoupled scalar linear time-varying differential equations, one per neuron, which are solved in parallel and in closed form. The solver admits two complementary interpretations: a variable metric forward-backward splitting method, rooted in operator theory, and a waveform relaxation quasi-Newton method, rooted in circuit simulation. We prove uniform and superlinear convergence, with a bound that reflects the conductance structure of the dynamics. The algorithm is illustrated on the classical Hodgkin--Huxley neuron and on networks of up to 5000 neurons.

Amir Shahhosseini +1PDF
arXivother

RPEG: Training Reasoning Language Models Scalably and Cost-Effectively

Reinforcement learning (RL) is now central to training Reasoning Language Models (RLMs), but RLM pipelines are difficult to scale and costly to run because each RL step couples multiple large models, heterogeneous stages, cross-step state, and complex asynchronous dependencies. Existing systems typically optimize only parts of this design space. We address this complexity by jointly optimizing the pipeline structure as well as its execution using a novel data-centric graph intermediate representation called RLM Pipeline Execution Graph (RPEG). RPEG exposes structural choices such as bounded asynchrony and actor-critic sharing together with execution optimizations such as stage fusion, fine-grained scheduling, and multi-dimensional model parallelism. It is coupled with a comprehensive analytical performance model for RLM training, jointly capturing compute, communication, memory, and hierarchical rack-, node-, and GPU-level effects. Across systems with up to 256 GH200 GPUs, our distributed runtime engine implementing RPEG improves performance by up to 2x over ReaL and reduces step latency by up to 2.48x over verl, while increasing hardware utilization by up to 30%. Overall, RPEG provides a graph-driven foundation for next-generation training pipelines for frontier reasoning LLMs.

Maciej Besta +5PDF
arXivai-hardware

Circuit Theory-Informed AI for Condition Monitoring in Power Electronics

Artificial intelligence for power electronics is evolving toward a deeper comprehension of system intrinsic behaviors and dynamics. As a novel integration of neural computation and electric circuits, spike-based neuromorphic computing offers promising potential in high energy efficiency and adaptability for data-driven applications. Building on the latency-coding principles, this article proposes a framework for data-light condition monitoring that relies solely on three-phase AC current measurements. The AC current signals are encoded as membrane potentials of RC-based spiking neurons, which are inherently compatible with spiking neural networks that can effectively facilitate the estimation of device parameters. Case studies are presented to validate the performances, while constraints and further optimizations are also evaluated and discussed, and future extensions of the methodology are provided.

Yubo Song +2PDF
arXivhardware-design

SEU Hardened Multi-Context Spintronic FPGA

Due to susceptibility to radiation-induced soft errors and resource constraints implementing large and complex designs on field programmable gate arrays (FPGAs) in harsh environments is challenging. Therefore, to mitigate these issues, this paper proposes an innovative approach through the introduction of a single event upset (SEU)-hardened spintronic FPGA architecture that incorporates multi-context capabilities. Leveraging the non-volatile properties of magnetic tunnel junction (MTJ), the proposed design can store two/four distinct configurations internally, eliminating the dependence on external non-volatile memory which significantly boosts the system performance. Additionally, through post-layout simulation, we demonstrate our design immunity to SEU, which enhances its reliability and operational stability in environments susceptible to radiation-induced soft errors. The efficiency of the proposed design was also evaluated using the ROSSETA benchmark, which serves as a representative model for real-world FPGA applications.

Abdolah Amirany +3PDF
arXivhardware-design

An Educated Guess: Deriving Statistically Aligned Gradient Estimates for Zeroth-Order Optimisation

As novel hardware architectures are explored for the purposes of machine learning, including neuromorphic chips, asynchronous systems, and ASICs, effective learning algorithms are highly desirable. While backpropagation is the standard for differentiation on standard hardware, its requirement for global synchrony and exact differentiability limits the exploration of exotic algorithms on non-traditional substrates and devices. Zeroth-order (ZO) optimisation methods, such as SPSA and weight perturbation, offer a compelling alternative by estimating gradients through inference-only passes; however, these methods have historically suffered due to the `curse of dimensionality', where random perturbations become increasingly misaligned with the true gradient in high-dimensional spaces. In this work, we move beyond blind random noise by deriving zeroth-order search directions from activation geometry available during the forward pass. This leads to a family of Spectral Zeroth-Order methods, SZO-$κ$, whose directions solve local response-based gradient-guessing objectives under explicit stochastic assumptions. These methods explore feature-supported directions that are more likely to produce informative loss responses. Using multilayer and convolutional networks, we analyse gradient alignment, bias-variance structure, and training stability of these methods relative to standard perturbation baselines and backpropagation.

Justus Huebotter +3PDF
arXivaccelerators

Enabling Retraining-Free Accurate Inference in Analog Compute in Memory Accelerators

Analog compute-in-memory (CIM) accelerators compute matrix-vector products directly inside memory arrays, eliminating weight movement and promising order-of-magnitude gains in inference efficiency. However, converter and device errors make pretrained models unusable on these arrays, and hardware-aware retraining is costly. We enable retraining-free, accurate inference of pretrained models on simple analog arrays with low-cost, shared calibration. Our key observation is that analog error depends on how each layer is represented on the array, and that this representation can be changed offline without changing the computation. We center the inputs and apply Hadamard rotations designed for analog arrays, which keep shared reference errors spread and confine input mixing to each array. Across CNNs, vision transformers and LLMs up to 14B parameters on PCM and ReRAM, the method brings vision models and larger LLMs within a few points of floating-point accuracy and matches or approaches published hardware-aware training without training any weights. On ReRAM with 6-bit ADCs, it restores ViT-B/16 from 37.4% to 82.9% and Phi-4 from 27.0% to 72.2%, both within 2.2 points of floating-point accuracy.

George Karfakis +1PDF
Scholarmodel-compression

Anomaly Detection in EVCS Using Federated Learning With Auto-Optimized Edge Models

This paper presents FedNAS, a novel federated learning (FL) framework designed for robust and personalized models to detect anomalies at distributed electric vehicle charging stations (EVCSs), tailored to real-world industrial deployment. The proposed framework enables collaborative anomaly detection in spatially distributed EVCS without sharing raw data, thus preserving user privacy and complying with data governance requirements. A lightweight neural architecture search (NAS)-optimized Simple Neural Network (SimpleNN) is proposed for each EVCS node to ensure computational efficiency within edge resource constraints. In contrast to traditional methods that depend on fixed structures and simplistic aggregation, FedNAS employs an ensemble-based meta-aggregation strategy, facilitating superior knowledge integration among heterogeneous clients compared to standard Federated Averaging (FedAvg). The framework is evaluated using the IEEE 123-bus test system with multiple EVCS nodes subjected to a cyberattack. The results demonstrate that the FL framework achieves a detection accuracy of up to 98.25%, with significantly reduced communication overhead and improved resilience to adversarial threats. Comparative analysis with centralized learning and federated variants validates the robustness and scalability of the proposed approach outperforms FedAvg by 6.25% in average precision. Furthermore, real-time simulation data is extracted using the OPAL-RT platform to emulate realistic industrial grid conditions, validating the framework’s effectiveness in cyberattack detection within dynamic EVCS environments. These findings highlight the potential of federated learning to enable secure, privacy-preserving, and scalable cyberattack detection for next-generation EVCS infrastructure.

Arif Hussain +23IEEE transactions on industry applications
Scholarhardware-design

Evidence-Calibrated Hierarchical Repair of Combinational RTL under Bounded Verification Budgets

Verification-guided repair prioritizes candidate edits using evidence from failed executions, but its effectiveness under bounded verification budgets depends on both the fault model and the search schedule. We investigate evidence-calibrated hierarchical repair (ECHR), combining backward output-cone localization, four implemented semantic counterexample tags, operator-family priors, feedback-based reranking, guarded two-edit branches, and exact finite-domain acceptance. The tag classifier, operator priors and cone localization adapt the counterexample-signature machinery of the VGMA verification-guided multi-agent RTL framework. Across ten deterministic synthetic-corpus seeds, we execute 2,520 detectable faults on 180 small combinational reference circuits, including 1,080 two-site faults. Every candidate is evaluated on 64 witness vectors, and accepted repairs must agree with the reference for all 512 three-input, three-bit assignments. With a fixed 30-query budget, ECHR repairs 820/2,520 cases (32.54%), compared with 400/2,520 (15.87%) for random search, 762/2,520 (30.24%) for static combined ranking and 771/2,520 (30.60%) for the strongest static ranking (tag-only). The ECHR–combined-ranking difference is positive in all ten seeded replications (mean +2.30 percentage points; 95% seed-bootstrap interval +1.67 to +2.94). Ablations show complementary effects of feedback reranking and two-edit expansion, although two-site repair remains limited to 53/1,080 cases. Results characterize a restricted, synthetic bit-vector repair problem rather than sequential or industrial Verilog.

Ryan Nguyen +30
Scholarhardware-design

Coupling Leakage in Theory and Practice - Unveiling (Post-PnR) Security Flaws in Masked FPGA-Mapped Designs

With the widespread adoption of Field Programmable Gate Arrays (FPGAs) in security-critical industries such as defense and telecommunications, ensuring the confidentiality of sensitive data processed by these devices has become paramount. Side-Channel Analysis (SCA) poses a significant threat, necessitating the protection of cryptographic primitives through effective and efficient countermeasures. Within the framework of well-established formal adversary models, Boolean masking offers provable resistance to SCA by randomizing sensitive intermediate computations using Shamir’s secret sharing. However, the secure and efficient implementation of Boolean masking in hardware presents a complex and error-prone challenge. Physical effects such as glitches, transitions, and couplings can undermine essential security assumptions, potentially weakening the effectiveness of masking. Although robust masking schemes are designed to maintain their security in the presence of glitches and transitions, and their secure implementation can be verified at the gate level using a wide range of automated verification tools – albeit mostly for Application-Specific Integrated Circuits (ASICs) and not FPGAs – leakages caused by coupling effects remain undetectable by such tools, as they occur at a lower abstraction level.In this work, we focus on detecting (low-level) vulnerabilities in masked FPGA designs through experimental and tool-assisted evaluation. The flaws that we target are not detectable at the Register Transfer Level (RTL) level, and in some cases, not even at the gate level, as they arise from optimizations introduced during synthesis or implementation, or from the specific outcomes of the Place and Route (PnR) process. We demonstrate that these flaws are not only theoretically concerning, but can lead to observable leakages in practical experiments. Furthermore, we show how to formally abstract, unveil, and mitigate such leakages, thereby enabling a security-aware FPGA design flow that spans from the behavioral to the physical level. As an example, we implement all the evaluation steps presented as an extension of PROLEAD and validate its accuracy and effectiveness through practical case studies.

Nicolai Müller +30IACR Cryptology ePrint Archive
Scholarhardware-design

SVAN-FSM: Static Analysis of Finite State Machines in SystemVerilog

Finite state machines (FSMs) are a widely used design pattern in hardware description languages (HDLs) for modeling control logic. Despite their simplicity, FSM implementations are prone to various classes of design errors that are difficult to detect using conventional simulation-based verification techniques. This paper presents SVAN-FSM, a tool for automatic extraction and static analysis of FSMs from SystemVerilog designs. The tool operates on RTL netlists, builds graph-based representations, and detects unreachable states, dead transitions, deadlocks, and livelocks. It also identifies interacting FSMs to expose cross-module control dependencies and handles parameterized designs by analyzing each elaborated instance independently. Analysis results are mapped back to the source code to improve interpretability. The tool has been evaluated on multiple open-source RISC-V designs and security-critical projects, demonstrating its effectiveness on industrial-scale designs. A comparison with a commercial static analysis tool shows that the proposed approach identifies significantly more FSM instances, particularly in designs where the transition logic extends beyond canonical coding patterns.

Andrei Viazovtsev +50International Test Conference in Asia
Scholarhardware-design

A Multiplier Free Logarithmic AGC for FPGA Front Ends with Hardware Validated Sub 60 dBc Distortion at 200 MHz

A distortion constrained digital logarithmic amplifier is described in this paper. It is VHDL for FPGA signal processing front ends with complete hardware verification. Conventional linear domain AGC loops suffer from amplitude dependent dynamics and harmonic accumulation at the gain stage. The proposed architecture closes the loop completely in the log domain. It achieves predictable pipeline latency, −60 dBc THD, and zero DSP48 usage. The design resolves three conflicting constraints without compromising any of them. The way three bricks come together is what makes the contribution. First, a leading zero count log estimator. Then, a shift scaled integrating loop filter. Finally, a shift augmented antilog reconstruction stage. Everything runs on a 16bit fixed point. Loop dynamics do not change in amplitude when input swings are greater than 40 dB. The normalization objective can be programmed and is determined by a static register. Validation was done using a Xilinx Zynq 7020 SoC.The AD9361 transceiver was used to route sinusoidal stimuli. At 200 MHz on the Artix 7 fabric, the design achieves −62.1 dBc THD, 78.4 dB SINAD, and 1.28 μs settling to 1% for a 20 dB amplitude step. It occupies 347 LUTs with four cycle deterministic latency.

Jabran Safdar +10International Bhurban Conference on Applied Sciences and Technology
Scholarhardware-design

Scalable Uvm-Based Verification and Fpga Implementation of a Parameterized High-Speed Barrel Shifter

This paper presents a scalable UVM-based verification framework and FPGA implementation of a parameterized high-speed barrel shifter supporting 32-bit and 64-bit configurations. The combinational RTL architecture performs singlecycle bidirectional shift operations with zero latency. A reusable UVM environment integrating constrained-random testing, directed corner-case sequences, cross-coverage, and SystemVerilog Assertions (SVA) was developed using Cadence Xcelium 25.03. Simulation results achieved 263 successful transactions with 100% functional coverage, zero mismatches, and zero UVM errors. FPGA synthesis on Xilinx Artix-7 utilized 138 LUTs for 32-bit and 372 LUTs for 64-bit designs, demonstrating scalable high-performance datapath verification.

Divya Gupta +20International Symposium on VLSI Design and Test
Scholarhardware-design

High-Performance Implementation of AES-128 Encryption and Decryption on FPGA Using Pipeline Architecture with UART Communication

With the growing number of sophisticated cryptographic attacks and the need for high-speed secure data communications, the design and implementation of hardware-based encryption appears inevitable. This paper aims to design a highspeed pipelined Advanced Encryption Standard (AES-128) architecture using the latest 28 nm Xilinx Artix-7 Field Programmable Gate Array (FPGA). Particularly, we envisioned a pipelined AES architecture with parallelised pipelined and expanded key setup stages to maximise the throughput. To validate the proposed ideas, the architecture was carefully described in Verilog Hardware Description Language (HDL). Furthermore, both encryption and decryption pipelined round components were replicated and encapsulated into a single module. This module was also interconnected with a newly proposed real-time UART IP core for immediate hardware debugging. Additionally, functional verification based on NIST standard test vectors was utilised to verify perfect data recovery with no timing violations. Finally, the Place and Route on the target device reported the logic utilisation and throughput of 59.1 Mbps (13.1 Mbps) for encryption (decryption).

H. K +20
Scholarhardware-design

A High-Power-Efficiency, Third-Order DT Delta-Sigma ADC Using Time-Division Controlling Technique in the First Integrator

This paper presents a high-power-efficiency switched-capacitor, 3rd-order, and 1-bit delta-sigma modulator (DSM). Compared to conventional 3rd-order and 1-bit DT-DSM, a time-division controlling (TDC) technique by simply increasing the time period of the integration phase, decreasing the time period of the sampling phase, and avoiding part-time wasted in the sampling phase is proposed in its first integrator. Due to the proposed TDC technique used in the first integrator, an operational-transconductance amplifier (OTA) with significantly less quiescent current can be used to satisfy the requirement of the first integrator’s (integrator1) operation. In addition, two body-effect-less (BEL) switches with a higher linearity are adopted as sampling switches in integrator1 to achieve a better peak signal-to-noise-distortion ratio (SNDR). Both the conventional and the proposed DSMs have been implemented with a standard 180 nm complementary metal-oxide-silicon (CMOS) process. At a supply voltage of 3.3 V, a signal bandwidth of 20 kHz, and a sampling frequency of 3.2 MHz, two designed DSM achieve a similar SNDR of 90 dB and a similar effective-number-of bits (ENOB) of 14.8 bits; however, the proposed DSM obtains a reduction of around 40% of power consumption. In addition, based on the proposed DSM a dedicated digital filter is also designed by using Verilog language, where the out-bandwidth noise is well filtered.

Jinning Li +80IEEE transactions on consumer electronics
Scholarhardware-design

A Multi-Core Winograd F(2,3) FPGA Accelerator with Low On-Chip Memory Footprint for VGG16 Convolution

Convolutional layers remain one of the main computational bottlenecks in deep neural networks, motivating specialized hardware accelerators for efficient inference. Winogradbased FPGA accelerators reduce the cost of $3 \times 3$ convolutions, but their efficiency depends heavily on how transformed activations and weights are stored, buffered, and delivered to the compute fabric. This paper presents a parameterized multi-core Winograd F(2,3) convolution engine described in Verilog and implemented on a Xilinx ZCU104 FPGA. The design streams the input feature map as narrow row-stripes and caches weights on chip, so that only a few image rows ever reside in block RAM (BRAM); this is the source of its distinctive low-memory profile. Tile-level parallelism across seven compute cores and a broadcast PE array sustain throughput, and a bit-accurate flow based on Pythongenerated VGG16 tensors validates the hardware. On the target FPGA the accelerator reaches 1452 GOPS for the VGG16 conv1 workload at 222 MHz using 16-bit fixed-point arithmetic, while occupying only 3.2% of BRAM, 24% of LUTs, 22.6% of FFs, and $\mathbf{1 0 0 \%}$ of DSPs, at $\mathbf{9, 8 2 ~ W}$. The proposed design therefore offers a strong low-memory operating point for FPGA Winograd acceleration, with the current evaluation focused on the VGG16 conv1 layer.

Thomas Henrique Schmitz +30Symposium on Integrated Circuits and Systems Design
Scholarhardware-design

Complex Data Leakage Analysis for Verilog Hardware Designs Based on Quantitative Information Flow

In complex Verilog RTL designs, covert data leakage mainly exhibits multi-source joint characteristics in the relational dimension and low-frequency temporal characteristics in the time dimension. Existing RTL-level analysis methods insufficiently characterize the relational semantics and temporal features of complex data, which can lead to false negatives. To address this issue, we propose a quantitative information flow analysis framework for complex data leakage. First, based on multi-source relational structural dependency analysis, we design a joint relational-sensitive semantic method to achieve relational node identification and relational semantic reconstruction. Then, a Temporal Aggregated Posterior Vulnerability mechanism is proposed, based on window-level exposure risk under multi-cycle observation, to enable multi-cycle aggregated computation for low-frequency temporal data leakage. Experimental results show that our method maintains good analysis effectiveness on public benchmarks and significantly improves the identification and quantitative characterization of complex data leakage, thereby providing more fine-grained quantitative analysis support for the pre-silicon security evaluation of Verilog hardware designs.

Yi-Bo Peng +60International Test Conference in Asia
Scholarhardware-design

Phase Unwrapping Module for Real-Time Biomedical Vibration Detection Using FPGA

The phase unwrapping is one of the most important processes in radar-based biomedical sensing that allows the extraction of the physiological signals, including the respiration and heartbeat. Nonetheless, traditional software based solutions have high latency and poor real-time response. The paper introduces a phase unwrapping module in real-time on the PYNQ-Z2 platform with the use of a hardware-software co-design methodology. The algorithm is implemented as a custom Verilog IP core that is interfaced using the AXI and can be efficiently hardware-accelerated in terms of time. Experimental data show that the proposed system can attain about 1.89x speed improvement over CPU-based implementation with throughput improved by about 1803 samples/sec to 3401 samples/sec. The system also demonstrates better energy efficiency (971 samples/Joule) and less memory use. The suggested solution allows real-time biomedical signal processing with low latency and low energy consumption, which is suitable in embedded healthcare applications.

Harshitha M. P. +30
Scholarsynthesis-pnr

Implementasi  Sistem Pakar Deteksi Dampak Penggunaan Gadget Pada Balita Dengan Metode Certainty Factor (Studi Kasus Klinik Utama Loka Medika)

The rapid development of digital technology has increased the intensity of gadget use in toddlers, so that it has the potential to have a negative impact on physical, cognitive, social and behavioral aspects. Limited parental knowledge and access to consultation with experts cause the early detection process to not be carried out optimally. This study aims to implement a web-based expert system using the Certainty Factor method to help parents detect the level of impact of gadget use on toddlers and provide treatment recommendations according to the diagnosis results. System development is carried out through the stages of needs analysis, design, implementation and testing. The diagnostic process applies the Certainty Factor method through a combination of expert and user belief values. The results of the study show that the system is successfully implemented according to the needs of the user and produces a diagnosis based on the symptoms entered. Based on the test results, the system obtained a score of 83.2% from 40 respondents, which indicates that the user acceptance rate is in the very good category. Thus, the Certainty Factor-based expert system is suitable for use as a medium for early detection of the impact of gadget use on toddlers to support handling decisions quickly, objectively and in accordance with expert knowledge.

Putri Fayza Pramestia Putri Fayza Pramestia +10InTech
Scholarsynthesis-pnr

Forecast-Driven Triggering for Switched Cyber–Physical Systems With Adaptive Interactions: Theory and Experimental Validation

This article investigates forecast-driven triggering (FDT) control for switched cyber–physical systems with limited communication resources. An adaptive FDT protocol is developed, in which the next triggering instant is predicted based on available sampled information rather than continuously detecting the triggering condition. To improve design flexibility, a dynamic coupling weight is further introduced into the controller synthesis. Compared with conventional event-triggered mechanisms, the proposed method reduces unnecessary transmissions and continuous monitoring requirements. By constructing a Lyapunov-based analysis framework, sufficient conditions are derived to guarantee exponential semiglobal ultimate boundedness without relying on the conventional Lyapunov–Krasovskii functional (LKF). In addition, a Zeno-exclusion condition is established to ensure the practical implementability of the triggering rule. Finally, numerical simulations and hardware experiments are provided to verify the effectiveness and applicability of the proposed approach.

Tan-We Chang +40IEEE Internet of Things Journal
Scholarsynthesis-pnr

Vulnerability Analysis of Multiagent Systems Under Node False-Data Injection Attacks With Two-Layer Stealthiness

This article proposes node false-data injection (NFDI) attacks with cyber-layer and physical-layer stealthiness launched by malicious agent nodes in multiagent systems (MASs), which inject unbounded attack sequences to destabilize the tracking error dynamics of all healthy agents. The difficulty faced by the existing node cyberattacks is that the launch of unbounded attack sequences will expose malicious agents from the cyber layer and physical layer. By contrast, the proposed NFDI attacks eliminate the influence on the agent detectors (cyber-layer stealthiness) and avoid the reverse influences on the physical dynamics embodied in the malicious agents (physical-layer stealthiness). First, in terms of the characteristic space of the agent dynamics, a necessary and sufficient condition is established for the existence of the NFDI attacks with two-layer stealthiness. On this basis, with the decomposition technology of topology incidence matrices, a design approach is developed to generate the unbounded attack sequences that still maintain two-layer stealthiness. The effectiveness of the proposed NFDI attacks is illustrated by two examples of vertical take-off and landing (VTOL) aircraft and autonomous underwater vehicles (AUVs).

Tian-Yu Zhang +10IEEE Internet of Things Journal
Scholarsynthesis-pnr

Secrecy Rate Optimization in Multiuser RIS-Assisted MIMO–NOMA IoT Networks

This article investigates physical-layer security (PLS) in multiuser reconfigurable intelligent surface (RIS)-assisted multiple-input multiple-output nonorthogonal multiple access (MIMO–NOMA) systems, which are key enablers for massive connectivity in Internet of Things (IoT) networks. We formulate a max–min secrecy rate optimization problem to ensure fairness among users in the presence of an active eavesdropper (EV) by jointly designing transmit beamforming, RIS phase shifts, and power allocation. The resulting problem is highly nonconvex due to the strong coupling among optimization variables. To address this challenge, we develop an efficient block coordinate descent (BCD)-based framework, where successive convex approximation (SCA) is employed for beamforming and RIS optimization, and a low-complexity dual-based method is used for power allocation. The proposed algorithm guarantees monotonic convergence to a Karush–Kuhn–Tucker (KKT) point with polynomial complexity. The simulation results demonstrate that the proposed design significantly improves the minimum secrecy rate compared with benchmark schemes, including maximum ratio transmission (MRT), RIS-free transmission, and semidefinite relaxation (SDR)-based methods. Furthermore, it achieves faster convergence than the SDR-based approach.

Ahmad Beigian +40IEEE Internet of Things Journal
Scholarsynthesis-pnr

Isolating Mobile Operating System Performance Through Hardware-Controlled Benchmarking

Existing mobile benchmarking tools conflate hardware capabilities with software efficiency, complicating fair architectural comparisons. This article introduces MobiOSBench, a standardized and flexible evaluation framework designed to isolate the mobile operating system (OS) performance through the hardware-controlled benchmarking. By executing workloads natively and holding physical hardware constant, MobiOSBench controls for silicon disparities to measure true software overhead. The framework evaluates systems across four contextual pillars: privacy, resources, performance, and efficiency, using configurable weighted user profiles to translate raw throughput, memory consumption overhead, and ambient telemetry into real-world performance indices. MobiOSBench is validated through extensive evaluations of four diverse mobile OSes deployed systematically across three distinct physical devices. Results confirm the framework reliably exposes intrinsic architectural tradeoffs, distinguishing systems optimized for power management from those prioritizing privacy or raw computational throughput. Extended case studies on iOS and a Raspberry Pi edge device further demonstrate its cross-architecture adaptability. A sensitivity analysis additionally confirms that these rankings remain largely stable under varying weight assignments, and a systematic comparison against existing benchmarking tools establishes the framework’s methodological novelty. Ultimately, MobiOSBench provides a mathematically rigorous methodology for transparent, context-aware mobile systems evaluation.

Naser AlDuaij0IEEE Internet of Things Journal
Scholarsynthesis-pnr

A Collaborative Perception Framework for Efficient Communication and Accurate Prediction

Collaborative perception enables supervisual sensing, alleviating the blind spots and limited detection range inherent to single-vehicle perception. In real-world deployments, however, constrained communication bandwidth restricts the amount of information that can be exchanged and introduces intervehicle latency. As a result, the delayed features received by the ego vehicle cannot be exploited to fully realize the benefits of collaboration. To address bandwidth limitations, we propose a collaborative perception system designed to make effective use of both the available communication capacity and temporally delayed perception features. Its efficient data communication mechanism can fully compress and transmit full feature-level data. To better exploit temporally misaligned features from neighboring vehicles, we introduce a second-order flow generator grounded in a physical motion model. By leveraging historical observations, it can reliably estimate the current-time feature representation, thereby enhancing collaborative perception performance. In addition, we develop a two-stage training scheme and an end-to-end loss function tailored to the overall framework. Experimental results show that, compared with the traditional approach of transmitting complete feature maps, our method reduces the amount of data transmitted and communication latency by about 4 times. Meanwhile, our proposed two-order flow generator exhibits superior robustness under varying communication delays (100–400 ms) on the V2XSet dataset, maintaining stable performance where competing methods degrade significantly. Finally, we validate the effectiveness of each component through ablation experiments.

Chuan Hu +60IEEE Internet of Things Journal
Scholarsynthesis-pnr

Loci-GS: Full-Automatic Disentangled 3D Asset Placement in 3D Gaussian Scene via Language Grounding

Text-guided asset placement in 3D Gaussian Splatting (3DGS) scenes presents a compelling frontier for immersive content creation. However, current progress is hindered by two primary limitations: (i) the lack of frameworks for purely text-guided layout within entangled representations, and (ii) the reliance of existing methods on cumbersome manual supervision or explicit geometric priors. In this paper, we introduce Loci-GS, a novel framework designed to automatically place assets into 3DGS scenes guided solely by textual instructions. Specifically, we employ a hierarchical Multimodal Large Language Model (MLLM) strategy for precise semantic grounding, coupled with a multi-view consensus mechanism to accurately localize interaction anchors. Subsequently, we derive optimal placement via geometry-aware alignment, utilizing physical regularization and appearance harmonization to ensure realistic integration. Extensive experiments demonstrate that our method generates physically plausible and visually coherent inserts, achieving state-of-the-art performance compared to editing baselines on a single RTX 3090 GPU.

Yang Zhou +20
Scholarsynthesis-pnr

DynaCartoon: 3D Dynamics-Guided Cartoon Video Generation from a Single Illustration and Forces

Animating a single anime illustration into a controllable video is of significant artistic value and remains challenging. Existing image-to-video models typically synthesize motion at the pixel level, lacking an understanding of physical dynamics. Meanwhile, explicit modeling approaches are often limited to the deterministic refinement of simulation results, failing to bridge the gap between physical reality and artistic style. To address this, we propose DynaCartoon, a framework for generating stylized cartoon videos guided by physics-based motion priors. Given a single illustration and user-specified forces, our method first explicitly models the geometric and attribute information of the scene. Coupled with a physics simulator, it derives force-driven motion sequences to serve as 3D dynamic priors. To overcome the visual rigidity of simulation results, we design a sketch-guided video cartoonization model. By introducing the Background Dynamic Enhancement (BDE) and Foreground Deformation Compensation (FDC) modules as two decoupled structural control branches, this model injects a dynamic atmosphere to the background and compensates for the deformations absent in physical solvers, synthesizing the artistic tension characteristic of cartoons. Experimental results demonstrate that compared to existing methods, DynaCartoon achieves significant improvements in motion plausibility and visual quality.

Zhuo Chen +40
Scholarsynthesis-pnr

When Environments Breathe With You: Passive Biofeedback for Accessible Restorative XR

Traditional horticultural therapy (HT) delivers proven psychological benefits yet remains inaccessible to hospitalized patients, individuals with mobility impairments, and urban residents due to physical barriers and spatial constraints. This paper presents BreathForest, a proof-of-concept breath-responsive VR system demonstrating the technical feasibility of translating HT's therapeutic mechanisms into passive virtual environments. We introduce light interaction, a design paradigm where unconscious breathing drives environmental responses without controllers or learned techniques. Through CO2-based sensing (<100ms latency, >90% accuracy), the system translates respiratory patterns into synchronized multimodal feedback across three progressive forest scenarios. Evaluation with 12 participants reveals strong system responsiveness (M=6.35/7), enhanced embodiment (M=6.12/7), and positive therapeutic impact (M=5.75/7). Our contributions include: (1) establishing light interaction as an accessible therapeutic XR paradigm, (2) providing a replicable translation framework mapping physical HT mechanisms to breath-responsive interactions, and (3) validating CO2 sensing technical specifications. By preserving HT's experiential elements (agency, responsibility, achievement) while eliminating participation barriers, this work establishes foundations for future clinical validation with excluded populations.

Meng-Yao Guo0
Scholarsynthesis-pnr

PPI-MDD: Phonological Prior-Infused Mispronunciation Detection and Diagnosis for L2 Mandarin Speech

Automatic mispronunciation detection and diagnosis (MDD) is a crucial component of computer-assisted language learning systems. However, adapting existing MDD paradigms to Mandarin faces critical challenges: the severe scarcity of Mandarin as a second language data and the lack of interpretability. To address these limitations, we leverage Mandarin's rigorous bottom-up phonological hierarchy, which provides a strong prior knowledge of phonological compositional structures and physical articulatory mechanisms, and propose the phonological prior-infused MDD (PPI-MDD) method. Specifically, we introduce a phonological representation learning module to map continuous speech onto physically meaningful articulatory representations and improve fine-grained interpretability at the underlying level. Furthermore, we design a phonology-based dual-stream network to explicitly model phonological structures, which effectively constrains the hypothesis space in low-resource scenarios. Finally, we propose a training-free text-fusion decoding algorithm based on the goodness of pronunciation to preserve explicit decision logic while effectively utilizing multimodal information, which also enables dynamic adjustment of MDD strictness. Extensive experiments show that PPI-MDD outperforms state-of-the-art methods with a 10.6% relative F1-score improvement without text prompts, and a 33.4% relative F1-score improvement with text prompts, while delivering multi-level and fine-grained diagnostic feedback. The implementation code is available at https://github.com/BNU-ERC-ITEA/PPI-MDD.

Zhi-Cheng Lian +20
Scholarmodel-compression

HE-FR: A High-Efficiency Fault-Resilient Framework for Stuck-at-Faults Mitigation in ReRAM-Based In-Memory Computing

ReRAM-based in-memory computing (IMC) architectures enable efficient neural-network inference on edge devices, due to their high density and non-volatility. However, ReRAM is prone to the stuck-at faults (SAFs) which would distort weight mappings and reduce model accuracy substantially. To tackle this problem, a high-efficiency fault-resilient (HE-FR) framework is introduced in the paper, which couples layered-precision quantization based on the generalized gaussian distribution cumulative distribution function (GGD-CDF), preserving weight distributions with the minimal information loss. Further, a fault-aware weight re-decomposition method is proposed. Specifically, a Q-agent-based method is proposed for sparse SAFs, which employs offline reinforcement learning to construct a globally optimized hash table with reduced compilation complexity, while incorporating both weight matching and cell-state stability into the reward function to suppress electro-stress. For dense SAFs, we exploit a greedy search over the remaining cells to provide rapid and accurate mappings. Experimental results show that GGD-CDF quantization improves accuracy by 2.18% over Float32. Under fabrication faults, HE-FR surpasses a fault-free method by 12.97% and accelerates compilation by $197\times $ with only 0.024% picojoule-level energy overhead. In electro-stress evaluations, HE-FR reduces performance degradation by 46.27%. It demonstrates that the proposed framework has advantages in robustness, efficiency, and reliability.

Junlong Lin +40IEEE Transactions on Circuits and Systems Part 1: Regular Papers
Scholarmodel-compression

Predicting Conversion from Mild Cognitive Impairment to Alzheimer's Disease: A Systematic Review of Deep Learning Models for Early-Stage Disease Classification.

INTRODUCTION Alzheimer's disease (AD) is a progressive neurodegenerative disorder for which early diagnosis-particularly the accurate prediction of conversion from mild cognitive impairment (MCI) to AD-is essential to enable timely and effective therapeutic interventions. Deep learning (DL) models have demonstrated substantial promise in this domain; however, critical challenges persist, including multiclass staging of disease progression, longitudinal data modeling, and effective multimodal data integration. This systematic review provides a critical appraisal of DL-based approaches for predicting MCI-to-AD conversion, with particular emphasis on these three key challenges. METHODS This systematic review was designed and conducted in accordance with the PRISMA guidelines. A comprehensive literature search was performed across PubMed, Scopus, IEEE Xplore, and Web of Science for articles published between January 1, 2019, and February 20, 2026. Following rigorous screening of titles, abstracts, and full texts, 60 studies were included that employed deep learning models for AD stage classification and/or MCI-to-AD conversion prediction. Data were extracted and synthesized regarding study design, datasets, input modalities, DL architectures, and classification tasks. RESULTS Cross-sectional approaches remained predominant (47 studies), while longitudinal designs were less common (13 studies) and showed heterogeneous but promising performance for MCI-to-AD conversion prediction. Heavy reliance on the ADNI dataset (42 studies) represents a major limitation to generalizability. Multimodal models were used in 24 studies and often reported strong performance, particularly for challenging MCI-related tasks; however, direct cross-study comparisons should be interpreted cautiously because of substantial heterogeneity in datasets, prediction tasks, validation strategies, and methodological quality. Convolutional neural networks (CNNs) dominated neuroimaging-based modeling (21 studies), while recurrent neural networks (RNNs) (4 studies) and transformers (4 studies) have emerged for capturing longitudinal dependencies and global relationships. The binary pMCI vs. sMCI classification proved the most challenging task (accuracy range: 71.71-96.3%), with performance declining as the number of classes increased in multiclass settings. Key limitations across studies include lack of diverse datasets, overfitting, and poor model interpretability. CONCLUSION Deep learning models hold considerable potential for predicting MCI-to-AD conversion, yet substantial barriers remain to their translation into routine clinical practice. Greater emphasis on longitudinal analysis, intelligent multimodal fusion, and interpretable architectures is essential for clinical impact. Future research should prioritize the development of diverse, multicenter datasets, advancement of explainable AI (XAI) techniques, and the design of personalized time-to-event models. This review offers a comprehensive roadmap to guide subsequent investigations toward more accurate, reliable, and clinically actionable diagnostic tools in Alzheimer's disease.

Mozhgan Ghasabi +20Ageing Research Reviews
Scholarmodel-compression

Deep learning-based metal artifact reduction in CT for head and neck radiotherapy

Background and purpose Metal artifacts from dental fillings and implants compromise CT image quality, target delineation, and dose calculation accuracy in head and neck (H&N) radiotherapy. Previous reviews have largely focused on traditional metal artifact reduction methods and image quality metrics rather than radiotherapy-specific endpoints. This systematic review evaluates deep learning-based metal artifact reduction (DL-MAR) approaches for H&N radiotherapy, emphasizing technical advances and clinical dosimetric validation. Materials and methods Following PRISMA guidelines, PubMed, IEEE Xplore, and Scopus were systematically searched (updated April 2026) for studies assessing DL-MAR methods in H&N CT imaging and radiotherapy. Of 227 identified records, 40 studies (11 traditional MAR and 29 DL-MAR studies) met the inclusion criteria. Results Reported DL architectures included convolutional neural networks, generative adversarial networks, dual-domain networks, diffusion models, and task-oriented loss functions. Among studies with clinical dosimetric validation, GAN-based and task-oriented approaches consistently reported dose deviations below 3%. Gamma pass rates exceeding 95% (3%/3 mm) were reported in a subset of these studies. Dual-domain and diffusion-based methods were evaluated primarily using image quality metrics without direct radiotherapy-specific dosimetric validation. Conclusions Current evidence suggests that DL-MAR may achieve clinically acceptable dosimetric accuracy while improving artifact suppression and target visualization in H&N radiotherapy planning. Task-oriented loss functions currently demonstrate promising potential for clinical translation based on available retrospective evidence. Broader clinical implementation remains limited by reliance on synthetic training data, restricted access to sinogram data, and the absence of prospective multicenter validation studies.

Ehsan Mihandoust +40Physics and Imaging in Radiation Oncology
Scholarmodel-compression

AI-driven drug repurposing: Methods, platforms, and translational applications.

The conventional drug discovery pipeline is resource-intensive, time-consuming, and characterized by high attrition rates, often requiring more than a decade and billions of dollars to bring a single therapeutic agent to market. Drug repurposing identifying new therapeutic indications for approved or investigational drugs offers a pragmatic alternative by leveraging established safety profiles and known pharmacokinetics. Despite growing interest, comprehensive syntheses that span the full spectrum of artificial intelligence (AI) methodologies from classical machine learning (ML) through deep learning (DL), graph neural networks (GNNs), and generative AI while also critically evaluating platforms, validation status, and translational readiness remain limited. This review addresses that gap. A narrative review methodology was employed. Literature was searched across PubMed, Scopus, Web of Science, and Google Scholar using a structured keyword strategy (terms: drug repurposing, AI drug discovery, machine learning, deep learning, graph neural networks, generative AI, reinforcement learning, clinical translation combined with Boolean operators). Studies published between 2018 and 2025 were prioritized, with foundational works included regardless of date. Inclusion criteria encompassed primary research articles, systematic reviews, and platform descriptions reporting AI-based drug repurposing methods or outcomes. Non-English publications, conference abstracts without full text, and purely theoretical works without empirical validation were excluded. We identified and synthesized evidence across six thematic domains: (1) classical ML strategies, including target-based and phenotypic screening; (2) DL architectures, including convolutional and recurrent networks; (3) GNNs for biological network modeling; (4) generative AI and reinforcement learning for de novo molecular design; (5) disease-specific applications in oncology, neurology, and COVID-19; and (6) bioinformatics databases underpinning model development. We explicitly compared ML, DL, GNN, and generative AI approaches in terms of data requirements, interpretability, scalability, and translational readiness and summarized these dimensions in a comparative table. Key platforms reviewed RepurposeDrugs, NeuroCADR, and COVID-19 repurposing frameworks were selected using predefined criteria that emphasized methodological transparency, accessibility, and reported validation. Across approaches, AI models frequently demonstrate strong internal predictive performance in silico, with graph-based and transformer models representing the current state of the art on benchmark datasets. However, external validation, reproducibility, and successful clinical implementation remain comparatively scarce. We therefore place particular emphasis on external cohort validation, prospective and experimental follow-up, and the distinction between computational promise and realized clinical benefit. Persistent challenges include dataset bias, limited model interpretability, regulatory uncertainty, and infrastructure gaps in low-resource settings. AI-driven drug repurposing holds considerable translational promise, but many claims remain preliminary. Interdisciplinary collaboration, shared benchmarks, transparent reporting of negative results, and prospective clinical validation frameworks are essential for realizing the full therapeutic potential of these approaches.

Z. Malik +20Computational biology and chemistry
Scholarmodel-compression

Artificial intelligence for stromal cell quality assessment: A systematic review of human mesenchymal stromal cell properties and viability

Reliable assessment of mesenchymal stromal cell (MSC) viability and stem cell–associated properties is essential for ensuring the consistency, safety, and therapeutic potential of MSC cultures before clinical use. However, incomplete understanding of the mechanisms underlying MSC therapeutic effects remains a major barrier to clinical translation and to establishing reliable, standardised quality attributes. Artificial intelligence (AI) may enable automated assessment of these properties, but the performance and clinical relevance of AI-based approaches have not been systematically evaluated. This systematic review evaluated the accuracy and clinical utility of AI for assessing viability and stem cell–associated properties in human MSC cultures. Following PRISMA guidelines, we searched five electronic databases (Cochrane, Scopus, PubMed, EMBASE, and MEDLINE) for eligible studies reporting quantitative AI-based assessments of human MSC cultures. Two reviewers independently screened studies, extracted data, and assessed risk of bias using PROBAST. The review protocol was prospectively registered with PROSPERO (CRD420256393235). Of 3,872 records identified, 18 studies met the inclusion criteria. Supervised AI models, particularly convolutional neural networks, were most used. Across included studies, reported performance metrics varied substantially because of differences in AI models, imaging modalities, datasets, and outcome definitions. Considerable heterogeneity in AI model architectures, imaging modalities, outcome definitions, and reporting metrics limited direct comparison across studies. Overall, available evidence suggests that AI can assess key MSC quality attributes and may support more objective, scalable, and standardised quality-control workflows. However, broader validation, standardised reporting, external testing, and evaluation across a wider range of MSC characteristics are required before routine clinical implementation can be recommended.

William Li +80Cell and Tissue Research
Scholarmodel-compression

Research on a Lightweight LSTM Time Series Forecasting Framework

In the prediction of industrial and transportation time series data, traditional recurrent neural networks face issues such as gradient explosion in the early stage of training, slow convergence, and difficulties in edge deployment due to massive model parameters. To address these issues, this paper proposes a lightweight Long Short-Term Memory (LSTM) prediction framework that combines adaptive learning rate scheduling with dynamic gradient clipping. First, a learning rate scheduling strategy integrating linear Warmup and Cosine Annealing is designed to achieve smooth initialization in the early training stage and fine-grained convergence in the later stage. Second, a dynamic gradient clipping mechanism based on Exponential Moving Average (EMA) is proposed to replace the traditional fixed threshold, adaptively suppressing numerical divergence during optimization. Finally, a lightweight architecture of "single-layer LSTM + single-head linear attention" is constructed, which significantly compresses the parameter scale while enhancing the ability to capture long-range dependencies. Experimental results on public time series datasets show that, compared with baseline models such as standard LSTM, GRU, and SVR, the proposed framework reduces the Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) by 16.6% and 18.9%, respectively. The training time per epoch is reduced by approximately 39.9%, and it does not require complex hyperparameter search, demonstrating high value for engineering applications and edge deployment.

Ling-Zhu Jiang +10
Scholarmodel-compression

CPU-based fast cardiac cine MRI segmentation using lightweight CNNs and knowledge distillation.

BACKGROUND Cardiac cine MRI segmentation is essential for quantifying ventricular function, but current deep learning methods typically require GPU hardware for inference, limiting deployment in clinical settings with constrained computational resources. This study develops and evaluates lightweight CNN architectures trained via knowledge distillation for fast and accurate cardiac cine segmentation on CPU devices, including deployment in web browsers. METHODS A residual-encoder U-Net teacher (9.22M parameters) was distilled into three lightweight student networks - MobileUNetV2 (2.09M), MobileUNetV3 (4.52M, with squeeze-and-excitation attention), and EfficientLiteUNet (5.08M, fused mobile convolutions) - trained with deep supervision at four output scales. Distillation matched softened teacher and student probabilities via Kullback-Leibler divergence, with a grid search over temperature T ∈ {1.0, 3.0, 5.0} and KD loss weight W ∈ {0.5, 0.99}, giving 6 KD configurations per architecture. Models were trained on 560 image series from the combined ACDC and M&Ms datasets, evaluated on 280 held-out series, and on a 29-patient external cohort (UVA). Paired t-tests on per-case mean Dice with Holm-Bonferroni correction (m = 6) compared each KD configuration with its baseline. Inference was benchmarked on CPU using PyTorch, native ONNX Runtime (C++), and ONNX Runtime Web - the latter running entirely inside the user's web browser with no images sent to any server. RESULTS The teacher reached a mean Dice of 0.894 across the three structures (LV cavity, LV myocardium, RV cavity). The best KD students reached 0.883 (MobileUNetV2), 0.883 (MobileUNetV3), and 0.888 (EfficientLiteUNet) at T=3.0, W=0.99, versus baselines of 0.871, 0.871, and 0.876. KD improved every architecture (Holm-adjusted p ≤ 5 × 10⁻⁷ at the best configuration), the top configurations forming a plateau rather than a unique optimum. Students tracked the teacher on the external cohort (0.903 versus 0.891 - 0.895) and on segmentation-derived ejection fraction in both cohorts, where no student's error differed from the teacher's (internal LV_EF bias -0.2 to +0.8%). On CPU, EfficientLiteUNet ran at 11 ms per frame natively (ONNX Runtime, 4 threads) while MobileUNetV2 was smallest (8.3 MB) and second-fastest in the browser (183 ms per frame), behind MobileUNetV3 (159 ms). CONCLUSIONS Lightweight CNNs trained with knowledge distillation come within 0.006 - 0.011 Dice of the teacher and match its downstream clinical measurements, while running on CPUs and inside ordinary web browsers. The browser deployment performs every step locally on the user's device, so no patient images leave the local environment - a practical path to clinical use without GPU infrastructure or outbound transfer of imaging data. A reference instance is at https://cmrsegmentation.org and the code is at https://github.com/xf4j/cmrsegmentation. PLAIN ENGLISH SUMMARY Cardiac MRI is the reference standard for measuring how well the heart pumps, but it requires a clinician or technologist to draw outlines around the heart's chambers and walls on dozens of images per scan - a slow and subjective task. Computers can do this automatically using artificial intelligence (AI), but most modern AI segmentation tools require expensive graphics hardware (GPUs) and often involve sending the patient's images to an external server, which raises both cost and data-privacy concerns. In this study we trained smaller, faster AI models that match the accuracy of a much larger reference model (within about one percentage point on standard accuracy benchmarks) while running on an ordinary computer's main processor in a fraction of a second per image - about a hundredth of a second natively, under a fifth of a second inside a web browser. We compared three different small-model designs and a grid of training settings, and found a range of settings that worked equally well across all three - useful guidance for other groups doing similar work. We then packaged the trained AI directly into a web page so that anyone with a modern web browser can drop in a cardiac MRI scan and see the segmentation result in seconds. Crucially, the web page does not send any images to a server; the AI runs entirely on the user's own computer, so patient data never leaves the local device. A live demonstration is available at https://cmrsegmentation.org and the source code is open-source. This makes accurate, automated cardiac MRI segmentation accessible to clinical sites that lack specialised hardware or that cannot share imaging data with external services.

Xue Feng +30Journal of Cardiovascular Magnetic Resonance
Scholarmodel-compression

Info-Driven Zero-Cost Proxy: Rethinking Vision Transformer Architecture Evaluation via Information Quantification

Neural Architecture Search (NAS) automates the design of Vision Transformer (ViT) architectures. However, the high computational cost of training-based methods has made training-free, zero-cost proxies a key research direction. While existing proxies can estimate model potential, they fail to capture critical information transmission characteristics due to the black-box nature of neural networks. Consequently, they suffer from high computational latency and weak correlation with ground-truth performance. Although information transmission determines performance, it has not been effectively quantified or utilized. This limitation remains the core bottleneck for current zero-cost proxies. In this paper, we propose Info-NAS, a zero-cost proxy based on architectural information. This method achieves the first structured quantification of information transmission in ViT. Specifically, it evaluates architectures using three core components: global information volume, local information gradient, and global consistency. Info-NAS derives proxy scores solely from architectural parameters, requiring no forward or backward propagation. Consequently, the calculation and search processes are highly efficient. Moreover, we construct the ViT-Info-Bench dataset to facilitate correlation analysis and algorithm evaluation. Experimental results demonstrate that Info-NAS significantly reduces search overhead while achieving superior ranking accuracy. With a single evaluation requiring only 1.8 ms, Info-NAS outperforms existing methods in both efficiency and performance.

Yue Yang +70
Scholarframeworks

ProCerDCL: Query-Efficient Black-Box Probabilistic Verification of Deep Learning Compiler Optimization

We present \tool, a query-efficient framework for black-box probabilistic verification. Given a fixed, sampleable input distribution and a binary property oracle, \tool assesses the property's violation probability $p$ against a threshold $\theta$. Under independent sampling and stateless oracle behavior, acceptance asserts $p\le\theta$, rejection asserts $p>\theta+\eta$ for a user-specified separation margin, and an inconclusive verdict indicates insufficient evidence for either conclusion. The probability of each type of erroneous verdict is bounded by its specified error budget. To support this three-valued contract efficiently, \tool uses a single sequential test, eliminating the interval family required by the baseline. The test accounts for repeated sampling checks through either corrected exact binomial confidence intervals or an anytime-valid confidence sequence. We apply \tool to probabilistic local model semantic consistency, where the property oracle detects decision-level inconsistencies between reference and compiled executions under a specified perturbation distribution. Our evaluation covers 24 tasks across TorchInductor and TensorRT and examines sampling efficiency, conclusiveness, and diagnostic behavior. For sampling efficiency, we use paired one-model clean-label radius searches: the baseline requires 3.6-4.2 times as many unique samples under matched numerical parameters while providing a weaker decision contract. For model semantic consistency, checks at $\theta=10^{-3}$ yield 817 acceptances, 19 inconclusive verdicts, and 4 rejections across 840 input--backend instances from 21 tasks. Case studies of two documented compiler faults further demonstrate what these local verdicts can diagnose.

Zexin Li +40
Scholarframeworks

A deep learning model for automated identification of age-related macular degeneration atrophy

Age-related macular degeneration (AMD), a leading cause of visual impairment and blindness among the elderly, is projected to affect 288 million individuals globally by 2040. Advanced AMD, including complete retinal pigment epithelium and outer retinal atrophy (cRORA), pose significant challenges for diagnosis and monitoring due to the labor-intensive, costly, and variable nature of manual annotation of volumetric optical coherence tomography (OCT) scans. Automating cRORA diagnosis offers the potential to improve annotation consistency and reduce clinical burden, which could facilitate, for example, the evaluation of recently FDA-approved treatments that delay disease progression. In this study, we compiled two large independent cohorts totaling nearly 5,000 3D OCT scans, labeled them for cRORA presence, and developed a deep learning model for cRORA automated detection. The model achieved state-of-the-art performance, with a ROC AUC of 0.97 on internal validation, and demonstrated robust translatability (zero-shot learning) with a ROC AUC of 0.88 on external evaluation. Notably, it exhibited high accuracy for both non-neovascular (non-nv) and neovascular (nv) AMD subgroups (ROC AUC 0.98 and 0.93, respectively), including complex cases with exudation. This model and dataset combination could facilitate clinical research and trial analyses by providing scalable, standardized assessments across non-nv and nv AMD patient subgroups. Identifying cRORA and related AMD atrophic changes on 3D OCT scans requires expert review and substantial time, which limits large scale research and clinical workflows. Existing automated approaches for AMD atrophy detection show variable performance and limited generalizability across datasets and graders. Identifying cRORA and related AMD atrophic changes on 3D OCT scans requires expert review and substantial time, which limits large scale research and clinical workflows. Existing automated approaches for AMD atrophy detection show variable performance and limited generalizability across datasets and graders. This study presents the first evaluation of SLIViT, a state-of-the-art AI model for 3D scans, on a large real-world dataset, tailored for automated cRORA identification across nearly five thousand OCT volumes from two independent cohorts. The model achieves high accuracy internally and robust zero shot generalization externally, outperforming benchmark methods while remaining consistent with multiple specialist grading groups. The method demonstrates reliable performance across AMD subgroups, defined by neovascular status, including complex exudative and anti VEGF treated cases, supporting practical use in clinical and research settings. The final optimized model checkpoint is publicly released to enable further research and retrospective clinical studies, including analyses of atrophy occurrence across datasets and investigations into longitudinal progression (which would require dedicated segmentation/regression endpoints). This study presents the first evaluation of SLIViT, a state-of-the-art AI model for 3D scans, on a large real-world dataset, tailored for automated cRORA identification across nearly five thousand OCT volumes from two independent cohorts. The model achieves high accuracy internally and robust zero shot generalization externally, outperforming benchmark methods while remaining consistent with multiple specialist grading groups. The method demonstrates reliable performance across AMD subgroups, defined by neovascular status, including complex exudative and anti VEGF treated cases, supporting practical use in clinical and research settings. The final optimized model checkpoint is publicly released to enable further research and retrospective clinical studies, including analyses of atrophy occurrence across datasets and investigations into longitudinal progression (which would require dedicated segmentation/regression endpoints).

O. Avram +220Graefe's Archive for Clinical and Experimental Ophthalmology
Scholaredge-ai

Urban Theories for Smart Cities: AI-Driven Predictive Modeling for Sustainable Infrastructure Design

Urban planning and civil infrastructure are deeply altered by artificial intelligence (AI) design through the use of data - driven, prognostic , and flexible frameworks. As global populations rapidly urbanize, traditional, static planning paradigms are proving insufficient to address the complex, interwoven challenges of modern city management, such as carbon emissions, resource strain, and urban sprawl. The objective of the paper is to examine the meeting of traditional urban theory and trending AI technologies in addressing the development of sustainable smart city. By adopting a systematic literature review between the period of years 2010 to 2025, the study projects a route from elementary Internet of Things (IoT) sensor networks to modern , entirely combining AI and digital twin systems . The paper explores various AI applications around the urban systems to propose two theoretical frameworks: a continuous AI - driven data pipeline and a digital twin feedback loop. Combining progressive machine learning algorithms with large urban datasets suggestively improves operational efficiency, shifting municipal governance from sensitive maintenance to evidence - based decision - making is one such findings . Other findings such as, digital twins and geospatial analytics allows active scenario simulation, analytical infrastructure maintenance, and optimized resource allocation, dynamically transforming metropolises toward spirited , carbon - neutral environments. Also , the research identifies critical, systemic barriers to adoption which contain determined cybersecurity exposures , extreme privacy concerns, the ethical consequences of algorithmic bias on reasonable expansion , and a requirement for homogeneous governing frameworks for public AI disposition . In conclusion, the paper highlights that although AI proposes metamorphic potential for the built environment the implementation process needs severe data control , interchange among legacy systems, and severe obedience to essential sustainability principles. At the end , the paper’s findings donate to the pool of knowledge on algorithmic urbanism, providing practical , socio - technical understandings for officials , urban planners, and researchers.

Kasim A. Korkmaz +20Inżynieria Mineralna
Scholaredge-ai

LiteML-Edge: A Contract-Centered Framework and Methodology for Replay-Driven 1:1 Python--Firmware Validation in TinyML

Tiny Machine Learning (TinyML) enables data-driven inference on resource-constrained microcontrollers; however, offline evaluation and firmware execution are often treated as loosely coupled stages, leaving reproducibility gaps between reference experiments and on-device behavior. This paper presents LiteML-Edge, a contract-centered TinyML framework and methodology for replay-driven 1:1 PythonFirmware Validation. LiteML-Edge integrates training, optimization, quantization, artifact export, firmware deployment, and per-stage validation while exporting replay artifacts, preprocessing rules, tensor I/O metadata, output mapping, post-processing, and the Rolling-24 metric policy as a shared validation contract. The framework supports fully quantized, fully floating-point, and hybrid tensor interfaces within a single workflow. The evaluation covers one environmental sensing scenario, a fixed W = 24 policy, and the TensorFlow-to-TFLM deployment path. On a WEMOS LOLIN32 (ESP32) and a NUCLEO-F411RE (ARM Cortex-M4), case studies based on a multilayer perceptron (MLP), a long short-term memory network (LSTM), and a compact one-dimensional convolutional model (Conv1D Tiny) showed how the framework localizes agreement or mismatch across preprocessing inputs, model payloads, raw outputs, final predictions, and Rolling-24 metrics. Under replay on both targets, MLP and Conv1D Tiny achieved contract agreement across the critical path, whereas LSTM preserved input-stage agreement but showed its first reproducible divergence at the raw-output stage. Field execution is retained as operational evidence collected under the same contract, complementing replay without replacing it for equivalence claims. The resulting workflow makes validation failures more attributable, reproducible, and auditable before deployment decisions are made, while distinguishing contract conformance across platforms from hardware-dependent deployment costs.

Charles Pereira dos Santos +30IEEE Latin America Transactions
Scholaredge-ai

Application of the Internet of Things in Wind Energy—Validation of Wind Turbine Power Models Using Open SCADA Data

The article asks the following questions: what is the Internet of Things, how do wind farms operate, and why is it worth integrating the IoT with renewable energy? It highlights the monitoring of wind turbine operating parameters, process automation and optimization, and failure prediction and predictive maintenance. Examples of wind farm systems utilizing the Internet of Things, as well as technologies and platforms used in monitoring, are presented. By identifying the challenges and limitations of data management, the future of the Internet of Things in wind energy is outlined. A secondary analysis of measurements from the Kelmarsh turbine illustrates machine learning in the analytical layer of an IoT system. The study assesses whether a power model selected using earlier data retains an advantage over a simple empirical power curve in a later operating period. The results highlight temporal validation, large-error assessment, and interval coverage monitoring. The calculations estimate power within the same measurement interval and do not forecast future production. In the calculations presented, the anticipated input signals and the recorded power output are derived from the same 10 min interval. The estimate is generated only after the measurements for this interval become available. No forecast horizon is defined; therefore, this analysis does not constitute a prediction of future production, nor does it demonstrate an ability to anticipate faults. To make claims regarding predictive maintenance, one would require verified records of failures or service interventions, information on operational constraints and states, diagnostic signals (e.g., vibration, component temperatures, gearbox parameters) along with their historical trends, and an evaluation of the lead time and classification error rates based on an independent dataset.

P. Radek +130Energies
Scholaredge-ai

AI - Driven Optimization of Indoor Environmental Quality in Buildings: From Data Analysis to Adaptive Control

Indoor environmental quality (IEQ) is one of the most direct ways in which buildings influence the health, comfort and performance of their users, yet in many buildings it is still controlled mainly by delayed, threshold - based rules. Such operation is increasingly insufficient, because the individual IEQ domains do not behave separately. Indoor air quality, thermal comfort, visual comfort and the acoustic environment interact through ventilation, material emissions, daylight, occupancy and energy use, and a control action that improves one component may worsen another. This paper reviews how artificial intelligence (AI) and machine learning (ML) can support a shift from reactive monitoring to predictive and adaptive management of IEQ. The reviewed methods are discussed according to their practical role rather than as an isolated catalogue of algorithms. Classical supervised - learning models, especially random forests and support vector machines, remain useful where datasets are smaller, training speed is important and the result must be understandable for building operators. Deep recurrent architectures, including long short - term memory and gated recurrent unit networks, are better suited to sensor time series and have shown strong potential for short - and medium - term prediction of parameters such as CO₂, particulate matter and thermal conditions. Convolutional - recurrent hybrids extend this approach by combining the detection of local patterns with the modelling of their development in time. Deep reinforcement learning then moves from prediction to control, because it can learn HVAC strategies that balance comfort, air quality and energy demand. Physics - informed and hybrid models add physical constraints to data - driven learning and are especially promising where measured data are limited or operating conditions change. The paper also places these methods within an integrated IEQ management system that links IoT sensors, data cleaning, calibration, model training, inference, explainable AI, multi - criteria optimisation and feedback to the knowledge base. The reviewed evidence shows that AI can improve building operation, but only when reported performance is interpreted carefully. Very high prediction accuracies or energy savings usually depend on the specific dataset, building type, sensor quality and validation protocol. The main obstacles to wider use are heterogeneous and drifting sensor data, the lack of long - term standardised datasets, limited interpretability of deep models and the computational cost of continuous deployment. AI - based IEQ management is therefore most valuable when data - driven prediction is not used as a black box, but is combined with building - physics knowledge, transparent decision support and a clear understanding of the real behaviour of the indoor environment.

Michal Kraus +20Inżynieria Mineralna
Scholaredge-ai

An Integrated IoT-Based Smart Farming Framework for Tomato Cultivation Using Fuzzy Logic Irrigation Control and Deep Learning Classification

Inefficient resource utilisation and changing environmental conditions pose significant challenges to tomato cultivation. This paper presents the design, implementation and experimental validation of a prototype smart farming system that combines Internet of Things (IoT) sensing, fuzzy-logic irrigation control and image-based classification of tomato status. An ESP32 microcontroller acquires soil-moisture, temperature, humidity and rainfall readings and publishes them through MQTT and HiveMQ Cloud to a PC, where a Mamdani fuzzy inference system implemented in MATLAB/Simulink R2025b computes the irrigation demand. The prototype was operated for four months with one potted plant. Separately, five classifiers were compared offline for tomato maturity classification. The five classifiers include ResNet18-based Convolutional Neural Networks (CNN), YOLOv8s cls, Random Forest, Support Vector Machine and K Nearest Neighbours. The ResNet18 achieved the highest test accuracy (95.00%; F1-score 94.97%), followed by YOLOv8s-cls (93.89%), whereas the classical models reached 85.6–88.1% on raw pixels and up to 90.8% on hand-crafted colour and texture features.

A. Sahbel +30Automation
Scholaredge-ai

A Severity-Aware Machine Learning Framework for Intrusion Detection and Risk Assessment in IoT Networks

The widespread use of Internet of Things (IoT) devices across critical infrastructure sectors has created extensive and heterogeneous attack surfaces that conventional rule-based security frameworks are structurally unable to address. This paper proposes and evaluates a machine learning-based cybersecurity risk assessment framework for IoT networks using Convolutional Neural Networks (CNNs), Random Forest, Support Vector Machines (SVM), and Autoencoders. Models were trained and validated on a harmonized pool of two publicly available IoT network datasets, TON_IoT and IoTID20, using stratified 5-fold cross-validation within the training partition and an independent held-out test set reserved for the final risk assessment. Empirical results demonstrate that the CNN model achieved the highest vulnerability detection performance across all evaluated metrics: 96.7% accuracy, 95.8% precision, 96.1% recall, and a 95.9% F1-score, outperforming Random Forest (95.1%), SVM (92.3%), and Autoencoder (89.5%). The proposed four-stage risk assessment pipeline, which comprises Data input module, Anomaly detection module, Risk scoring module, and Alert generation module, identified a total of 177 network anomalies within the held-out test set. Notably, 80.2% of these detections (n = 142) fell into the medium- or high-severity categories under a heuristic prioritization scheme, indicating that most CNN-flagged events in the evaluated data warrant analyst attention. This framework provides a proactive defense mechanism, which offers critical implications for deploying adaptive security protocols in heterogeneous IoT ecosystems.

Alaa Ali S. Almohanna +10Electronics
Scholaredge-ai

A hierarchical ensemble learning framework for cyberthreat detection in IoT platforms

The proliferation of Internet of Things (IoT) and Internet of Medical Things (IoMT) devices has dramatically escalated the risk of cyberattacks, necessitating robust and highly responsive Intrusion Detection Systems (IDS). However, existing solutions frequently struggle with severe class imbalance and high-dimensional data, leading to a suboptimal trade-off between detection accuracy and computational efficiency. To address these challenges, this paper introduces a novel hierarchical machine learning-based IDS optimized for real-time multiclass attack detection. We construct a comprehensive unified dataset by integrating the recent CICIoT2023 and CICIoMT2024 datasets. Our framework first employs a Genetic Algorithm (GA) for feature selection, extracting a lightweight subset of critical features to minimize computational overhead. The core classification engine utilizes a two-stage hierarchical Random Forest architecture coupled with selective SMOTE oversampling. The first stage identifies major attack categories while grouping minority instances into a meta-class, which are then precisely classified by the second stage. Experimental evaluations demonstrate that the proposed hierarchical framework establishes an optimal accuracy-efficiency trade-off and improves the classification performance on minority attack classes compared to traditional flat classifiers. While simple data undersampling methods achieve fast runtimes at the cost of severe performance degradation, our hierarchical approach preserves exceptional detection capabilities and operates faster than conventional models trained on full datasets. The framework achieves robust F1-scores of 90% on the unified dataset and 89% on CICIoT2023. Finally, the integration of Explainable AI (XAI) techniques, including SHAP and LIME, ensures model transparency and interpretability, making the proposed IDS a highly practical, reliable, and scalable solution for deployment in resource-constrained IoT environments.

Masoud Mostafaei +10Peer-to-Peer Networking and Applications
Scholaredge-ai

HSTENet: A deep learning-based scene text recognition system

In many computer vision applications such as intelligent transportation systems, autonomous navigation, digital content retrieval, and multilingual translation, detection and recognition of scenes play a vital role. These systems are fairly effective at document-oriented applications but less effective in real-world applications where an increased number of languages, orientation variation, noise, and distortion are present. To overcome these shortcomings, we are introducing a Hybrid Structured Text Extraction Network (HSTENet) for robust and accurate scene-text recognition that combines deep learning and machine learning. HSTENet features advanced modules such as edge detection with image preprocessing, support vector machine classification for detecting the textual part, and a line opening algorithm for fast segmentation of the image into regions. The recognition part relies on deep architecture, convolutional neural networks, attention-based connectionist temporal classification decoding, and bidirectional long short-term memory networks to recognize the high-precision target text. HSTENet was evaluated on three benchmark datasets—the International Conference on Document Analysis and Recognition, Common Objects in Context for Text, and a self-designed high-noise multilingual dataset—and obtained an F1-score of 91.9% and recognition accuracy of 93.8%, significantly outperforming other state-of-the-art models, including DeepText, convolutional recurrent neural networks, and connectionist text proposal networks. The framework is also polarization-invariant, real-time compatible (inference time 0.69 s), and deployable on edge devices. The results showed that HSTENet is highly efficient and automated for textual scene recognition, which could be applied to systems embedded in mobile platforms, self-contained platforms, and intelligent human–computer interaction.

Vishnukant Soni +50AIP Advances
Scholaredge-ai

Maintenance Strategies for Medical Equipment: A Systematic Review of Their Effects on Reliability, Availability, Safety, and Costs

Medical equipment is essential for patient diagnosis and treatment; however, its maintenance strategies remain heterogeneous. This systematic review aimed to characterize the maintenance strategies applied to medical equipment and analyze their effects on reliability, availability, safety, and costs. Following PRISMA 2020 guidelines, original studies were searched in Scopus, ScienceDirect, PubMed, and Web of Science, published between January 2015 and July 2026. Studies were included if they applied maintenance strategies (preventive, corrective, predictive, condition-based, risk-based, or reliability-centered) and evaluated their effects on reliability, availability, safety, or costs. Seventy-four studies were selected. Eleven equipment categories were identified, with diagnostic imaging, life support, and infusion therapy being the most studied. Hybrid strategies—particularly preventive combined with corrective—were the most frequent (14 studies), followed by predictive maintenance (13 studies). Advanced analytics, IoT, and machine learning were the most applied technologies. A gap persists between model performance metrics and real maintenance indicators: several studies reported high predictive accuracy, but limited evidence of improvement in MTBF, downtime, availability, or cost. This review offers an integrative framework for healthcare technology managers and identifies priorities for future research.

N. Chambi +20Technologies
Scholaredge-ai

Smart Logistic System: A Multidisciplinary Integrated System for Driving Safety and Transportation Management

The logistics and transportation industry faces several critical challenges, including the absence of real-time vehicle data, inadequate safety measures, and inefficient resource management. This paper presents the Smart Logistic System (SLS), a unified intelligent transportation architecture that integrates embedded systems, Augmented Reality (AR), Machine Learning (ML), Internet of Things (IoT), and digital-twin technology within a single operational framework, rather than as independent subsystems. The system continuously monitors key vehicle parameters — location, speed, fuel consumption, and engine performance — to support proactive maintenance and minimize downtime. An edge-deployable driver-behavior-monitoring module employs computer vision and a lightweight MobileNetV3-Small classifier to detect drowsiness, distraction, and other hazardous behaviors in real time. Special-needs support includes seizure detection, GPS-based wandering alerts, and AR-guided vehicle-operation assistance. A route-optimization module formulates logistics planning as a constrained, multi-objective vehicle routing problem that jointly considers transportation cost, travel time, and estimated carbon emissions, while a digital-twin layer maintains a continuously synchronized virtual representation of fleet state to support fleet-manager decision-making. A mobile application delivers real-time data, alerts, and navigation to drivers; a desktop application allows managers to decode raw Controller Area Network (CAN) data; and a cloud platform provides secure storage, synchronization, and visualization. On a held-out benchmark, the driver-behavior-monitoring module achieves accuracy and inference-latency performance competitive with, and in several respects exceeding, recent state-of-the-art driver-monitoring methods; the route-optimization, predictive-maintenance, and digital-twin modules are validated through architectural and feasibility analysis rather than large-scale benchmarking. Taken together, these results indicate that SLS’s principal contribution is the integration of safety monitoring, logistics optimization, predictive maintenance, digital-twin synchronization, and accessibility features within a single operational architecture — a step toward more secure, efficient, and inclusive logistics and transportation systems.

H. Mogahed +20IEEE transactions on intelligent transportation systems (Print)
Scholarai-hardware

Surrogate Modeling of Stochastic Hysteresis Response in Irregular RC Shear Walls Using CNN-KAN and Manifold Random Fields

Accurately predicting the hysteretic response of irregular reinforced concrete (RC) shear walls is essential for seismic performance assessment. However, spatial variability in material properties and geometric complexity introduce significant stochasticity, making conventional stochastic finite element (SFEM) methods computationally inefficient due to difficulties in random field assignment and the high cost of nonlinear analyses. This study proposes an efficient deep learning framework that uses a convolutional neural network-Kolmogorov–Arnold network (CNN-KAN) model to predict stochastic hysteresis responses of irregular shear walls with manifold random fields. The random field is constructed using the isometric mapping (isomap) method to accurately represent the spatial variability of material properties in SFEM. Then, hysteresis curves obtained from SFEM analysis are used to train the CNN-KAN model, the CNN-long short-term memory (LSTM) model, and the CNN-multilayer perceptron (MLP) model. While all three models achieve high prediction accuracy, only CNN-KAN successfully preserves the stochastic diversity of structural responses. The traditional CNN-MLP and CNN-LSTM models suffer from an averaging effect, failing to capture the inherent randomness. The trained CNN-KAN model is then used as a surrogate model for Monte Carlo simulation in large-scale sample analysis. The results demonstrated that it can efficiently generate hysteresis curves and energy dissipation distributions to study the influence of spatial correlation on structure stochastic responses. The proposed model significantly enhances computational efficiency, reducing the computing time to only 1/300 of the conventional SFEM analysis. Further, the proposed framework enables efficient uncertainty quantification and is extensible to other complex structural systems involving spatial randomness.

Shi-Xue Liang +30Journal of computing in civil engineering
Scholarai-hardware

A Theoretical Framework and Efficient Algorithm for Computing Condensed Multi-Level Bisimulation Graph Summaries

Summary graphs have shown to be a valuable tool in graph analysis and processing, with applications in query optimization, retrieval, and visualization, among others. Many summarization approaches are based on the concept of stratified bisimulation, which determines vertex equivalence based on outgoing paths of length k. Summaries based on bisimulation typically also store information about all paths of length less than k, and thus explicitly encode the 1, 2, ..., k bisimulations. To this end, we introduce condensed summaries: a formal model for representing stratified bisimulations in a compact manner. By leveraging repeating structures between consecutive bisimulation levels, condensed summaries represent all levels of bisimulation up to some k, while containing O(|V|) vertices, independent of k. We also prove that condensed summaries are equivalent to an established, uncondensed representation. Furthermore, we provide an algorithm for computing condensed summaries, along with an optimization that can yield a 20-fold speed-up. To demonstrate the efficacy of our model and algorithm, we provide experimental results on various edge-labelled graphs with up to 28~billion edges and a bisimulation depth of up to 4,757. All experiments successfully reached full bisimulation in less than 10 hours, while those with up to 1~billion edges only took minutes or even seconds, with no more than 64 GiB of memory. The resulting summaries were all substantially smaller than the original graphs, sometimes with a 95% reduction in vertices and edges. Code is available at: https://github.com/R-van-Bakel/Multi-Summaries.

Ruud van Bakel +40
Scholarai-hardware

BPB-GNN: Resource-Efficient Distributed GNN Training via Boundary-Aware Prefetching and Cost-Balanced Batch Construction

Distributed graph neural network (GNN) training is essential for large-scale graph learning, but existing mini-batch frameworks still suffer from substantial inter-partition communication overhead and indirect, coarse-grained GPU memory control. In distributed mini-batch training, remote sampled neighbors cause frequent feature fetching, while fixed seed-node batch sizes cannot accurately reflect the computation graph expansion induced by multi-layer sampling. This paper presents BPB-GNN, a resource-efficient distributed GNN training system that jointly optimizes inter-partition feature access, hierarchical storage management, and mini-batch construction. First, BPB-GNN introduces a training-centric boundary-aware prefetching strategy to localize high-value remote neighbor features associated with boundary nodes in training-node receptive fields. Second, it uses a DRAM/NVMe SSD hierarchy with priority-aware placement and dynamic migration to manage prefetched features under limited DRAM capacity. Finally, it constructs seed-node batches according to estimated sampling-induced load costs rather than equal seed-node counts, thereby balancing batch-level workloads and reducing peak GPU memory usage and its inter-batch fluctuation. These resource-efficient mechanisms may also improve the efficiency of large-scale GNN computation in computing power networks and JointCloud environments. Experiments on real-world graph datasets show that BPB-GNN achieves up to 5× speedup over the baselines while reducing peak GPU memory usage by up to 36%.

Zhi-Chao Chen +40
Scholarai-hardware

GSPARC: A Mode-Aware GPU-Based Method for Large-Scale Sparse Tensor Contraction

Sparse tensor contraction (SpTC) is a core operation for high-order sparse data analytics in recommendation, knowledge graph analytics, large-scale tensor mining, and scientific computing. Existing methods are limited by SpTC-misaligned representations, overlapping outputs across partitions, and memory-inefficient accumulation over large output spaces. We propose GSPARC, a GPU-based SpTC method that combines a new mode-aware bit-wise representation with mode-decoupled output partitioning and memory-efficient accumulation. These techniques improve index matching, eliminate inter-partition merging, and avoid full output-space allocation. Extensive experiments demonstrate that GSPARC outperforms state-of-the-art SpTC methods in contraction speed and scalability. Our code is available at this link: https://github.com/seohyeon0322/GSPARC.

Seohyeon Kim +30
Scholarai-hardware

ML Inference Acceleration via Scalable In-network Computing

Modern inference engines have significantly advanced the scalability of large models through innovations in serving techniques, proving instrumental in driving the widespread adoption of AI. However their true potential is constrained by traditional server architecture where reliance on host CPUs and operating systems introduces unnecessary overhead and degrades performance. We identify a significant opportunity to eliminate these bottlenecks by offloading data management directly to the network layer by proposing a unified architecture based on our previous research and ongoing work using programmable Data processing units (DPUs)/SmartNICs for in-network computing. Our previous work, NICTokenizer is a framework that offloads the initial tokenization phase to programmable network dataplane eliminating the host OS Kernel overhead lowering the overall tokenization latency by 61.3% and increasing the token throughput by 407.62% compared to conventional server based implementations and was previously published as a poster at SIGCOMM' 25. Alongside we introduce CacheFlow a disaggregated prefix cache pool which utilizes the DPU memory and GPUDirectRDMA for efficient sharing, storage and retrieval of KV cache across a cluster. CacheFlow reduces the average Time to first token (TTFT) by 82% imporves overall throughput by 39% and decreases memory storage requirements by 44% compared to standard inference engines like vLLM. As a natural progression our future research explores offloading semantic routing to a DPU enabling the network plane to dynamically select the appropriate model experts for each incoming sequence.

Shaunak Galvankar0
Scholarai-hardware

Learn to Compress or Compress to Learn? Shannon and Turing meet again in the LLM era

The year 1948 gave birth to two transformative pillars of modern computing: Claude Shannon's information theory, which gave rise to data compression, and Alan Turing's writings on intelligent machinery, which ignited the field of artificial intelligence. Today, nearly eight decades later, we are witnessing a profound hybridization of these once-distinct fields. As we push to efficiently store, process, and retrieve massive datasets across diverse data types, users, devices, and systems — from vector and graph databases to cutting-edge agentic memories — AI and compression can no longer operate in silos. In this talk, we investigate how Learning is fundamentally reshaping the fields of Indexing and Compression. We will focus on two challenging research directions: designing Learned Data Structures that fuse classical indexing with machine learning to achieve unprecedented efficiency, and developing next-generation Data Compressors driven by powerful autoregressive and diffusion neural models to reach new frontiers of efficacy. We will also examine the reverse direction — how Compression can reshape Learning through model compaction and comprehension. By unpacking the distinct computational advantages and current practical pitfalls of these lines of investigation, we will outline a roadmap for these research frontiers and hopefully inspire new cross-disciplinary collaborations.

P. Ferragina0
Scholarai-hardware

Curvature Graph Mergeability: Predicting Model Compatibility from Per-Layer Hessian Geometry

Merging independently trained neural networks --- via weight averaging, model soups, or task-vector arithmetic --- is increasingly central to large-scale model reuse, yet the success of any merge hinges on whether two parameter vectors occupy the same loss basin, a property currently certified only by expensive multi-point interpolation barriers. We introduce Curvature Graph Mergeability (CGM), a scoring framework that predicts the interpolation barrier from only cheap, per-layer Hessian curvature features at each endpoint. CGM extracts two scalar curvature statistics per layer using Hutchinson's trace estimator and the power method, organizes them into a chain-structured factor graph whose latent compatibility states encode qualitative regimes of layer alignment, and computes a mergeability score from the negative variational free energy of the chain. For deep linear networks under balanced factorization, we prove that the interpolation barrier is upper-bounded by a multilinear function of per-layer curvature mismatch, and we identify a diagonal-dominance condition under which the score is monotone in mismatch; the bound is stated as an implicit joint bound on the per-layer mismatch vector, with the regime of validity made explicit. Across four meta-training architectures and four held-out OOD architectures (ResNet-110, ViT-B/ImageNet, WRN-28-10, 3-layer LSTM), CGM attains Spearman ?=0.86 on leave-models-out splits and ?=0.81 pooled OOD, outperforming zero-order distance, CKA, Fisher overlap, REPAIR, midpoint-barrier, and architecture-agnostic curvature-MLP baselines at Holm--Bonferroni-corrected p<0.01. It is approached only by Git Re-Basin + barrier (?=0.90), which costs ~14,500× more per pair; CGM closes the gap to permutation-based merging at 35× amortized speedup over direct 11-point barrier when screening 50 candidates, with negligible memory overhead. Per-layer posterior marginals additionally pinpoint specific layers that act as merge bottlenecks.

Ze-Hua Cheng +20
Scholarai-hardware

Synapse‐Integrated Physical Unclonable Functions for Trusted Inference in Neuromorphic Computing

Secure neuromorphic computing requires compact hardware primitives that can authenticate devices and protect synaptic states without relying on stored keys or large physical arrays. Here, we propose a noise‐spectroscopy physical unclonable function (NS PUF) based on a compact floating‐body (FB) transistor with a charge‐trap layer (CTL). In this platform, the low‐frequency noise power spectral density serves as a frequency‐domain fingerprint of stochastic FB dynamics, while the CTL provides programmable device states for reconfigurable synaptic security. Device geometry diversifies the spectral response landscape, enabling multiple device‐specific responses to be generated from the same compact physical array. The spectral responses are converted into bitmap challenge‐response pairs and optimized to suppress redundant readouts and improve response diversity. The NS PUF‐derived keys can be used to lock or reconfigure synaptic weight‐mapping rules based on bipartite‐sort coding, linking intrinsic device entropy with trusted inference in secure neuromorphic operation. By assigning new security roles to two technologies already compatible with modern microelectronics, FB dynamics and CTL programmability, this work establishes a scalable hardware foundation for secure neuromorphic computing.

Minil Kang +60Advanced Functional Materials
Scholarai-hardware

Achieving Deterministic and Safety-Certified Operation in Software-Defined Vehicles through Hardware–Software Co-Design

<div class="section abstract"><div class="htmlview paragraph">The transition toward Software-Defined Vehicles (SDVs) [<span class="xref">1</span>] is driving a fundamental shift in automotive electrical and electronic architectures [<span class="xref">1</span>], where centralized high-performance computing platforms must support mixed-critical workloads under strict requirements for determinism, performance isolation [<span class="xref">2</span>], and functional safety. Consolidating safety-critical and non-critical applications onto shared System-on-Chip (SoC) platforms [<span class="xref">3</span>] introduces significant challenges in ensuring predictable execution behavior, bounded latency, and compliance with safety standards [<span class="xref">4</span>, <span class="xref">15</span>]. Addressing these challenges is essential for enabling scalable and reliable next-generation automotive systems [<span class="xref">1</span>] [<span class="xref">5</span>]. This paper presents a hardware–software co-design approach to achieve deterministic and safety-certified operation in SDV [<span class="xref">1</span>] architectures. Unlike traditional implementations that rely on software hypervisors or hardware isolation independently, the novel core contribution of this work is an integrated, dual-layer hardware-software orchestration model. It dynamically coordinates static Stage-2 translation with hardware Extended Resource Domain Controllers (XRDC) and Memory Partitioning and Monitoring (MPAM) to enforce end-to-end spatial, temporal, and interconnect-level bus isolation.</div><div class="htmlview paragraph">The proposed solution combines hardware-assisted resource isolation mechanisms, including Hardware-enforced resource isolation and Memory Partitioning and Monitoring (MPAM) [<span class="xref">6</span>], with a lightweight partitioning hypervisor and Virtual Machine Monitor (VMM). XRDC [<span class="xref">7</span>] provides fine-grained access control for memory and peripheral resources, while MPAM [<span class="xref">6</span>] enables memory bandwidth allocation and monitoring to minimize resource contention across competing workloads. On the software side, the partitioning hypervisor [<span class="xref">8</span>] enforces strong temporal and spatial isolation between Virtual Electronic Control Units (vECUs), allowing mixed critical applications to coexist safely on a shared platform. VMM supports efficient virtualization [<span class="xref">9</span>, <span class="xref">10</span>] and lifecycle management of virtual machines while maintaining deterministic execution characteristics.</div><div class="htmlview paragraph">The proposed architecture aligns with the requirements of ISO 26262 [<span class="xref">4</span>], enabling the certification of safety critical applications alongside non-critical services. By integrating hardware-enforced isolation with software-level partitioning, the approach ensures predictable system behavior and reduces interference between workloads. The results demonstrate that a coordinated hardware–software strategy is essential to bridge the gap between high-performance computing and functional safety, thereby enabling scalable, deterministic, and safety-compliant SDV [<span class="xref">1</span>] platforms for future automotive applications. System-level validation conducted on an NXP S32G3 [<span class="xref">7</span>] [<span class="xref">10</span>] vehicle network processor platform demonstrates the efficacy of the proposed co-design. Under artificial memory-bus contention generated by background non-critical (QM) Linux traffic, the hardware-enforced MPAM bandwidth regulation reduced maximum worst-case execution time (WCET) jitter for an ASIL real-time control loop from 142.6 μs down to 8.4 μs—achieving a 94.1% reduction in latency variation. Paravirtualized hypercall operations achieved a 40.2% latency reduction compared to standard Trap-and-Emulate approaches.</div></div>

Misbah Ullah Khan0SAE technical paper series
Scholarai-hardware

Model-Based Quantized Edge Vision Inference: Component Detection System for the STM32N6 Circuit Diagram

Against the backdrop of the rapid development of Edge AI, deploying object detection models on resource-constrained microcontrollers (MCUs) presents simultaneous challenges in terms of computing power, memory, and accuracy. This paper proposes an improved algorithm for detecting components in circuit diagrams and deploys it on an STM32N6 edge device. At the algorithmic level, using YOLOv8n as the baseline model, the study addresses the characteristics of circuit diagram components—which are extremely small and often arranged horizontally or vertically—by introducing an NWD-SIoU joint loss function to replace the default CIoU loss, thereby providing stable gradient signals and directional alignment for tiny objects.Additionally, Dropout regularization is incorporated into the detection head structure of YOLOv8n to address overfitting in circuit diagram training scenarios. Compared to the original algorithm, the improved algorithm achieves a 0.036 increase in F1-score, with precision and recall improving by 0.017 and 0.031, respectively. At the deployment level, post-training quantization from FP32 to INT8 was completed, reducing the weight size from 12 MB to 3.23 MB. To address the limited SRAM capacity of the STM32N6, a heterogeneous storage hierarchy allocation scheme was designed: static weights are stored in external Flash, while runtime activation tensors are locked in the NPU’s dedicated RAM. This approach frees up SRAM capacity while accelerating model execution. Experimental results demonstrate that the proposed approach achieves a balance of size, accuracy, and speed on the STM32N6 platform. It validates the feasibility of running complex vision models on edge devices and provides a reusable technical path for the implementation of edge AI in the field of industrial vision inspection.

Jie-Yao Li0