2026-11-02

7 repos · 22 papers · 50 scholar

Showing 79 of 79 items

GitHuboptimization

http-streaming

HLS, DASH, and future HTTP streaming protocols library for video.js

JavaScript2,665440
GitHubedge-ai

libreyolo

LibreYOLO is a MIT licensed open source computer vision library

Python97666
GitHubai-hardware

VOID-Nonbinary-Physical-Computing

Open-source research into nonbinary physical computing, integrating analog material circuits, HX89 native execution, CPU feedback, and hardware verification.

Haxe10
GitHubai-hardware

zoe

Project Zoe - connectome-based reservoir computing companion with evolving LoRA, virtual avatar, MicroVM sandbox, and companion mode. Working MVP of SRS v4.0.

Python10
arXivhardware-design

RTL-Interact: Training EDA-Interactive Models for Multi-Task RTL Design

Large language models offer a promising route to register-transfer-level (RTL) design, but reliable implementations require checking and revision based on electronic design automation (EDA) tools' feedback. Existing RTL-specialized models often lack learned interaction with such EDA tools or rely on separate generation and repair components. We introduce RTL-Interact, a multi-task post-training framework that trains a single RTL-specialized model to generate RTL, construct task-appropriate EDA checks, and revise implementations using tool feedback. The framework combines automatically constructed multi-task demonstrations with staged supervision from chain-of-thought to tool-integrated reasoning. We further introduce Pattern-Aware Policy Optimization (PAPO) for multi-task reinforcement learning. PAPO augments Group Relative Policy Optimization with correctness-gated rewards that favor concise successful checking while rewarding effective revision, and uses task grouping to preserve coverage across objectives. On CVDP, RTL-Interact-32B achieves 48.7% weighted Pass@1 and 68.9% Pass@5, surpassing its teacher model, DeepSeek-V3.2, while RTL-Interact-8B achieves 44.9% weighted Pass@1. Additionally, we show that PAPO can reduce average tool calls by 63.8% while improving Pass@1. Both models also outperform all RTL-specialized baselines on VerilogEval v2 and RTLLM v2, demonstrating the effectiveness of our method.

Wenxuan Shi +15PDF
arXivhardware-design

Prediction-Induced Sparsity for Tracked FMCW Radar: A Track-Conditioned Event-Gap FPGA Architecture

Tracked FMCW radar already carries a prediction of the next target state, yet the downstream digital hardware is commonly provisioned for a fixed spectral search. We present prediction-induced sparsity (PIS), a tracked-mode architecture that uses the prior before event formation and hardware scheduling. The predicted beat is derotated to an innovation-centred coordinate, while tracker covariance is propagated to a bounded residual support K. Sparse level-crossing packets are then executed by an exact event-gap resonator engine with a shared complex processing element and K-dependent state/coefficient memory. On a public FMCW dataset, causal prior-only con- ditioning reduces event records by 33.1% on average, and the representative fixed-point stream is bit-exact in RTL. And on an Artix-7 implementation, reducing physical support from 129 to 23 states lowers LUT and BRAM use by 65.0% and 88.1%, respectively, and reduces measured VCCINT power by 19.9% at matched throughput. The results show how tracking uncertainty can become a hardware-capacity implication for future adaptive radar accelerators.

Huy TrinhPDF
arXivsynthesis-pnr

Enhancing Sentiment Classification on Small Datasets through Data Augmentation and Transfer Learning

Small-scale sentiment classification often suffers from data scarcity, which limits the generalization ability of the models. This study evaluates and compares the effectiveness of three data augmentation strategies: Easy Data Augmentation (EDA), back-translation, and contextual token substitution (nlpaug-style), with both traditional machine learning classifiers (Logistic Regression, Random Forest) and transformer-based models (BERT). We perform a comprehensive empirical comparison with low-resource sentiment datasets by summarizing the results of recent studies and performing targeted head-to-head experiments. Our findings indicate that all augmentation methods improve performance. Contextual augmentation yields the most consistent gains for BERT models, while EDA and back-translation provide greater benefits for traditional classifiers. These insights help guide the selection of data augmentation techniques tailored to model type and dataset size, filling a critical gap in research on data augmentation for sentiment classification on small datasets.

Mahmoud S. Mayaleh +1PDF
arXivsynthesis-pnr

PolarFlow: Spectral Geometric Flow Matching for Directed Graph Generation via Polar Decomposition

Directed graphs model asymmetric relationships in biology, transportation, social networks, and electronic design automation. Recent discrete diffusion and flow-matching methods generate directed graphs effectively, but their edge-wise categorical formulations require dense predictions over all ordered node pairs and repeated categorical updates during generation. This motivates a fundamental question: can comparable generative quality be achieved through a more computationally efficient formulation that better exploits the structure of directed adjacency matrices? Existing benchmarks also cover limited network structures, restricting comprehensive evaluation. We introduce PolarFlow, a continuous geometric flow-matching framework based on polar decomposition and truncated spectral coordinates. Its representation comprises four coupled components: a Stiefel frame and rotation angles encoding the orthogonal factor, and a second Stiefel frame and nonnegative eigenvalues characterizing the positive-semidefinite factor. PolarFlow learns a joint velocity field along geometry-aware conditional paths on this product space, reconstructs a continuous adjacency-score matrix, and discretizes it into a binary directed graph. By replacing discrete edge-wise generation with joint continuous transport, PolarFlow improves computational efficiency while retaining global adjacency structure. To broaden evaluation, we introduce complementary Circuit and Twitter benchmarks spanning engineered functional and naturally evolved social networks. Across synthetic and real-world datasets, PolarFlow achieves generative quality comparable to competitive discrete methods while substantially reducing training and sampling time.

Yanan Zhao +6PDF
arXivaccelerators

Migrating a Legacy Scientific Application with an AI Coding Agent: NICAM from Fortran to Kokkos

Migrating from Fortran to C++/Kokkos broadens access to hardware accelerators through backends such as CUDA and HIP while offering greater flexibility in software design. However, legacy Fortran applications often have large codebases, making manual porting time-consuming and costly. Recent advances in large language models (LLMs) and AI coding agents have made large-scale code migration increasingly feasible. In this study, we ported NICAM (Nonhydrostatic ICosahedral Atmospheric Model), a legacy Fortran climate model accelerated with OpenACC, to C++/Kokkos with the assistance of Claude Code. The agent translated the code directly from Fortran to C++/Kokkos under human review. Approximately 140,000 lines of Fortran were analyzed and ported into 35,000 lines of C++/Kokkos, covering the dynamical core and seven physics packages. A 100-member ensemble comparison with the original Fortran model showed that the port is statistically consistent with the original. Performance was evaluated on Miyabi-G, an NVIDIA GH200 cluster, and Sirius, an AMD MI300A cluster. Compared with the Fortran/OpenACC baseline on GH200, the port achieved application speedups of $1.71\times$ on GH200 and $1.54\times$ on MI300A. The CUDA build achieved a GPU kernel-time speedup of $1.38\times$ on GH200. After normalization by peak memory bandwidth, the HIP build on MI300A achieved $0.96\times$ the kernel performance of the Fortran/OpenACC baseline.

Xuanzhengbo Ren +2PDF
arXivmodel-compression

When Does Test-Time Search Help World Action Models? An Empirical Audit of Candidate Selection and Simulator Integrity

World-action models increasingly apply test-time scaling, spending extra inference computation on candidate selection and adaptive allocation to improve control. Such gains are credible only when candidates carry exploitable signal, nominal compute savings become real speedup, and evaluation stays reliable. We audit all three on Cosmos Policy and RoboCasa with a premise-first, pre-registered methodology. On clean data, a learned value selector is 1.66 pp worse than random; a hindsight oracle gains only 1.40 pp, with an episode-level interval spanning zero; and 82% of states show identical candidate outcomes, leaving little selection headroom. Token pruning cuts per-forward FLOPs by up to 47% but yields at most a measured 1.15$\times$ end-to-end speedup on the deployed path (1.25$\times$ on a decode-excluded path), short of a 1.5$\times$ target, and our adaptive-allocation experiment is inconclusive under a power gate. We further identify and fix a silent rendering corruption whose rate rises sharply along compute-heavy planning paths while leaving the official direct-policy result unaffected, and release a detector, fix, and scripts. These findings show why test-time scaling should be audited as a chain of premises (decision signal, realized efficiency, evaluationintegrity) rather than by success rate alone. Conclusions are scoped to the tested configuration.

Jun Liu +10PDF
arXivmodel-compression

Prune Gradually, Adapt Sparsely: Efficient Pruning of Large Language Models at Extreme Sparsity

Weight pruning can reduce the storage and inference cost of large language models (LLMs), but preserving quality at high sparsity levels remains challenging. Existing methods face a quality-cost trade-off in this regime. One-shot pruning is cheap but degrades sharply as sparsity increases, even when followed by post-pruning recovery. A growing line of work instead optimizes the mask directly from the dense model under the end-to-end objective; among these methods, [Lee et al.,2026], the recent state of the art at extreme sparsity, reaches high quality but at the cost of model-sized optimizer state and auxiliary variables. To address this gap, we propose ETAPS, which prunes gradually and adapts sparsely: it tightens a global sparsity constraint in progressive stages while training a sparse delta with a small, fixed budget. Our key observation is that gradual weight removal keeps each pruning step small enough for sparse adaptation to compensate, allowing the adapted model to provide end-to-end gradients that guide the next pruning decision without model-sized optimizer state. Across six models from 125M to 13B parameters, ETAPS achieves lower perplexity than ELSA at every evaluated sparsity from 50% to 90%, with perplexity reductions of up to 42%. ETAPS is also far cheaper: at 90% sparsity, it prunes LLaMA2-7B on a single H200, where ELSA uses four, with $5.0\times$ smaller total persistent state and 49% fewer GPU-hours. Together, our findings point to a promising direction for achieving high quality in demanding regimes such as extreme sparsity at substantially lower cost.

Jean Lucien Randrianantenaina +6PDF
arXivaccelerators

Dynamical Memory Transformer: Selective Recurrent Memories with Constant Cost Attention

Transformers retain an explicit history of the sequence, allowing attention to directly retrieve past representations. However, the size of the KV-Cache and the cost of a new token increase with the context length. State Space Models compress the past into a fixed-size recurrent state, at the cost of reduced addressability. We present the Dynamical Memory Transformer (DMT), a hybrid architecture that separates sequence encoding from information reading. The encoder uses several memory units, each implemented by a dynamical system with a complex diagonal recurrent matrix. An input-dependent leak rate controls their updates. The decoder then uses attention to read and combine their outputs, without going through the token history. The formulation of the recurrence allows an associative scan with O(T ) total work and O(log T ) parallel depth, with fixed dimensions. During autoregressive inference, the recurrent states and the convolution cache remain fixed in size. We compare DMT with Mamba and Transformer on cognitive and language modeling tasks, and show its strengths and limitations: DMT achieves the best results on several cognitive tasks and language modeling performance close to that of Transformer, but remains behind Mamba in language modeling. On selective copy, DMT also shows stronger length extrapolation than Mamba and Transformer. These results demonstrate that parallel dynamical memory units are a promising mechanism for retaining and using information over sequences.

Yannis Bendi-Ouis +1PDF
arXivmodel-compression

Where KV-Cache Error Lives in Looped Language Models: A Diagnostic Study of Ouro and Huginn

A looped language model applies one block of layers 4 times per token and keeps a key/value (KV) plane for each loop. We study where storage error changes the next-token distribution and how shared-weight error accumulates, using Ouro-1.4B (4 loops) on GSM8K and MATH and Huginn-0125 (32 loops) on GSM8K-derived chains, with dense, simulated quantization. Per unit of error, values carry 94% of the diagonal Fisher mass, but measured KL is dominated by keys with larger quantization errors. Local second-order prediction follows the measured KL of typical sequences, including 136 held-out chains, but misses a few confident rows, and a local second-order factor for shared weights at 32 loops ranges from 1.65 (INT8) to 4.09 (INT4), so it depends on the error direction. A cell-level allocator lowers one-step KL at 25-48% of BF16 bytes, while paired task-accuracy comparisons depend on budget and the 25% gap against K4V2 shrinks to near zero under a post hoc answer-extraction check. A later Ouro-2.6B test retains the key/value split but falsifies measured-cost allocation's advantage over Fisher at 31%; across all 21 ledger registrations, 60 of 176 criteria failed.

Dezhen Lu +7PDF
arXivmodel-compression

Trainable Vector Quantization for Large Language Models

Vector Quantization (VQ) is a rapidly evolving technique for compressing Large Language Model (LLM) weights, reducing memory usage, and speeding up inference. However, extreme compression deteriorates model performance, which is challenging to recover due to the discrete nature of VQ parameters. Existing methods employ training-free approaches that are limited to layer-wise reconstruction loss minimization. To address this limitation, we propose TVQ, a novel method that leverages straight-through estimation to enable gradient descent-based optimization of vector-quantized LLMs. By reformulating VQ LLM optimization with trainable latent weights, TVQ enables arbitrary objective minimization. We constitute our approach by applying TVQ in a block-wise knowledge distillation paradigm, which balances quality and training efficiency, and propose several practical solutions to accelerate training. In experiments on modern LLMs, we demonstrate that TVQ is compatible with modern VQ schemes. As a result, TVQ achieves state-of-the-art quality in low-bit post-training quantization settings with a modest training budget and straightforward PyTorch implementation. For instance, TVQ achieves near-lossless 2-bit quantization for LLaMA2-70B with only 1% accuracy drop on popular Question-Answering benchmarks.

Anastasiia Prutianova +5PDF
arXivmodel-compression

ElTab: Elastically Looped Tabular Transformers

Tabular transformer models such as TabPFN achieve strong predictive performance but incur substantial inference costs, making them difficult to deploy in resource-constrained settings. While looped transformers reduce parameter count through weight sharing, they can lose significant predictive power when stopped early. To address this challenge, we introduce the Elastically Looped Tabular Transformer (ElTab), a single model trained to operate across different computational budgets through supervision at randomly sampled early exits. Across 30 OpenML classification datasets, a single ElTab model matches six depth-specific looped models while cutting training cost by more than half, and matches a 12-layer TabPFN-style baseline with 25% fewer inference FLOPs. These results show that elastic-depth training yields a single model that adapts to the deployment budget at hand, offering a favorable trade-off among predictive performance, storage, and computational cost.

Mieszko Komisarczyk +4PDF
arXivmodel-compression

DIFFNAS: Predictor-Free Neural Architecture Search via Diffusion Model and Zero-Cost Evaluation

Neural Architecture Search (NAS) has advanced the automated design of deep networks, but remains computationally expensive due to reliance on model training or performance predictors. We introduce DIFFNAS, a predictor-free NAS framework that leverages a discrete graph diffusion model guided by zero-cost performance proxies. Unlike prior diffusion-based NAS methods that require learned performance predictors, DIFFNAS conditions the generative process on training-free metrics such as MeCo and SWAP, enabling efficient architecture search with minimal computational overhead. Experiments on NAS-Bench-201 show that DIFFNAS finds the optimal architectures for CIFAR-10 and CIFAR-100 (94.37% and 73.51%) without a pretrained predictor, while reducing search cost by up to 17.5x compared to DiffusionNAG. On the MobileNetV3 search space, its accuracy is lower than that of DiffusionNAG. These results highlight the promise of diffusion models combined with zero-cost proxies for scalable and effective architecture search.

Yize Sun +5PDF
arXivaccelerators

SACQ: Structure-Aware Codebook Quantization for Ultra-Low-Bit Vision-Language-Action Models

Vision-Language-Action (VLA) models enable capable robot behavior, but their large memory footprint limits deployment on resource-constrained robotic platforms. Weight quantization is a natural remedy, yet post-training quantization (PTQ) recipes inherited from large language models degrade when applied to VLAs at ultra-low precision. Existing low-bit PTQ methods often fail to explicitly account for the structure of downstream operations and allocate precision without considering its effect on the policy actions. We introduce SACQ, an ultra- low-bit weight-only vector-codebook quantization framework for VLAs. SACQ combines (1) structure-aware codebook fitting with role-specific quadratic metrics that reflect input statistics and downstream sensitivity; (2) action-aware mixed-bit allocation based on the increase in final-action error as individual layers are compressed; and (3) two-stage post-quantization refinement that reassigns codeword indices with second-order error compensation and calibrates codebooks and row-group scales against the decoded action. Using GR00T N1.7, SACQ achieves 93.85% average success across four LIBERO suites at 1.5 bits per language/action weight while reducing model memory by 4.83x. With INT8 activation quantization and low-precision attention, it retains 93.55% success while running 1.51x faster than BF16. On SimplER, SACQ further achieves 63.76% and 52.27% success under Visual Matching and Variant Aggregation at 2 bits per language/action weight, compared to 58.90% and 53.45% for BF16, respectively.

Haesung Lee +1PDF
arXivmodel-compression

Distilling Knowledge from Search: Decoder Model Factory for Quantum LDPC Codes

We introduce GraphMP, a learning-based message-passing decoder for quantum low-density parity-check codes that transfers without fine-tuning across code families, code sizes, physical error rates, and noise models. GraphMP uses local message passing to keep computation scalable as the decoding graph grows. Guided by its predictions, beam search progressively refines candidate error patterns, trading additional decoding time for lower logical error rates. We develop the decoder model factory as a paradigm for training successive models on experience generated by earlier models during decoding. Model-guided search supplies verified trajectories for knowledge distillation, folding computation spent on decoding into reusable training data. Repeated collection and distillation improve the decoder across generations and adapt its strategy as the noise changes. These results motivate a coupled quantum--classical architecture with two interacting cycles: rapid decoding protects quantum information, while slower learning renews the models from operating experience. The long-term vision is a scalable decoder model factory that produces high-performance, low-latency decoders and continually adapts them to evolving quantum hardware.

Anil Bilgin +3PDF
arXivaccelerators

What We Keep of What We Forget: Replacement Operators and Routing Semantics in Training-Free MoE Expert Pruning

Training-free expert pruning makes two decisions: which experts to remove, and what the layer computes when its router still selects a removed expert. Published methods change both at once. Holding the mask fixed, we compare replacement operators from three families on OLMoE, DeepSeek-V2-Lite and Moonlight, with Granite as a fourth router, using paired, multiplicity-corrected tests. The largest effect is a choice usually left implicit: the survivors' gate scale. On routers trained without top-k renormalization, renormalizing the survivors' gates, a common implementation choice, costs 0.8-8.5 points and reverses whether rerouting beats MoNE's mean-output replacement; on two routers trained with renormalization, the trained scale wins at 50% pruning. We introduce the tilted novice, a closed-form low-rank map of the pruned expert's input whose rank-0 case is MoNE's constant. At 50% pruning it adds 1.2-1.8% parameters and gains 3.45 and 1.37 points over the constant on OLMoE and Moonlight, stable across calibration draws and largely unchanged by 4- and 3-bit quantization. No operator dominates: trained-scale rerouting is best on DeepSeek-V2-Lite, and a partial copy of each pruned expert is statistically indistinguishable from the tilted novice on two models. Evaluation can change conclusions: at 25% pruning MoNE's constant keeps 92-97% of dense multiple-choice accuracy but only 5-65% of GSM8K and 0-17% of code scores, and the calibration domain moves the mask 4-9 times more than resampling. We propose a five-item reporting checklist.

Ezza Ali +3PDF
arXivaccelerators

Experts Exhibit Structured Redundancy: Structure-Preserving Expert Pruning for Mixture-of-Experts Large Language Models

Sparse MoE mechanism significantly improves the trade-off between model capacity and inference efficiency. However, practical deployment still requires storing all experts, resulting in substantial memory overhead. We show that expert outputs in MoE layers exhibit structured redundancy. Across different models and task categories, the stable rank and entropy effective rank of expert outputs are consistently far below the nominal number of experts, while expert output distributions also exhibit fewer effective modes. Motivated by these observations, we reformulate expert pruning as a structure-preserving expert subset selection problem and efficiently optimize it via a greedy strategy, rather than only independently scoring and ranking individual experts or exhaustively searching over expert combinations. We propose two expert pruning methods under the unified structure-preserving framework: Projection Energy Expert Pruning (PEEP), which preserves the functional subspace by maximizing weighted projection energy, and Distributional Coverage Expert Pruning (DCEP), which preserves the distributional geometry by minimizing weighted coverage error. Both methods incorporate router-derived weights as priors, enabling them to jointly preserve the global structural relationships among experts while accounting for the local importance of individual experts. Experiments on DeepSeek-V2-Lite and Qwen3-30B-A3B demonstrate that, at the expert retention ratios of 75% and 50%, our methods outperform the state-of-the-art baselines on the majority of benchmarks and consistently achieve the best average performance across both general tasks and mathematical reasoning tasks.

Tian Wang +5PDF
arXivaccelerators

Koopman Greedy Spiking Control for Nonlinear Systems under Firing-Rate Limits

Neuromorphic controllers interact with physical systems through fixed-amplitude spikes whose timing conveys control information. Existing Lyapunov-based greedy spiking controllers guarantee stability but are restricted to linear systems and neglect firing-rate limits. We extend greedy spiking control to nonlinear systems using Koopman operator theory. For observable dictionaries invariant under differentiation along the input direction, spike-induced state translations admit an exact linear representation in lifted coordinates, enabling exact modeling of spiking jump dynamics. This allows firing decisions to be formulated as comparisons of quadratic Lyapunov functions in the lifted space. We prove practical stability of a clocked, rate-limited controller with an explicit ultimate bound and develop a data-driven design based on Koopman eigenfunctions learned from unforced data by generator-based regression. The weights of the Lyapunov function are synthesized by a linear program that enforces a sampled relaxation of the rate-limited control-Lyapunov condition. Applied to an inverted pendulum, the learned eigenfunctions recover the system energy, and the proposed controller achieves swing-up and stabilization from nearly all initial conditions despite severe actuation limits.

Nader MeskinPDF
arXivaccelerators

MemX: Memory-Aware Automatic Parallelism for Cross-Vendor Heterogeneous GPU Training

Cross-vendor heterogeneous GPU clusters are increasingly used for large-scale model training. However, their GPUs differ substantially in compute power, memory capacity, and interconnect characteristics. Existing automatic parallelism training systems usually minimize iteration time while treating memory as a hard constraint, almost failing in resolving the compute--memory asymmetry of cross-vendor clusters regardless correlation between memory and throughput. In this paper, we present MemX, a new memory-aware automatic parallelism system that jointly optimizes iteration time, peak memory usage, and cross-GPU memory balance. MemX consists of a weighted-product objective function, heterogeneous recomputation strategies, a minimal-set profiler, interpolation-based data augmentation, a dual-output XGBoost predictor, and a hash-signature prediction cache, in order to accelerate search of large-scale auto-parallel best configurations at lower cost. We implement MemX on Megatron-LM and evaluate it against four automatic parallelism baselines on two heterogeneous clusters. MemX achieves the lowest prediction Mean Absolute Percentage Error (MAPE) of 12.17% and is the only method that identifies the real Top-1 configuration under both time-based and memory-balance rankings. In terms of search cost and search space, MemX searches 31,240 configurations in 26.5 seconds, while an end-to-end profiling baseline is extrapolated to require about 17 days at the same scale. The simulation study that extends the measured 20-GPU data to 40-GPU and 60-GPU clusters further illustrates the effectiveness of MemX at larger scales.

Lin Meng +2PDF
arXivaccelerators

SparseGDN: Chunk Sparse in Linear Attention

Gated Delta Networks (GDNs) offer linear-time sequence modeling, yet long-context prefill remains costly due to dense intra-chunk computation, recurrent-state updates, and intermediate state storage. We introduce SparseGDN, a hardware-aware framework that combines structured intra-chunk approximation with input-dependent recurrent-event selection while producing outputs for every token. A block-adjacent solver reduces local Delta-solve computation, while a dynamic selector allocates recurrent-update budgets across head pairs. The GPU implementation materializes updates only for selected events and reconstructs token outputs using compact state snapshots and group-local interactions, reducing computation and memory traffic. We evaluate SparseGDN on Qwen3.5-4B and Qwen3.5-9B across RULER tasks and context lengths from 4K to 128K. Retaining approximately 29% of recurrent events on average, SparseGDN achieves 1.50-1.94x GDN-core prefill speedups over native FLA kernels, including online selection overhead.

Jiedong Zhuang +3PDF
arXivaccelerators

Input-dependent Neuromodulation Enhances Mixed-Signal Neuromorphic Hardware

Gating mechanisms allow a fixed neural network to adapt its computation to the current input by dynamically modulating its effective parameters. In the brain, neuromodulation uses this principle to move neural circuits between dynamical regimes, and deep learning exploits input-dependent gating, e.g., in gated recurrent networks and self-attention. Input-dependent modulation has also recently been shown to improve spiking neural networks, with particularly strong benefits in small, parameter-limited ones. This makes neuromodulation especially promising for neuromorphic processors at the edge, where area, memory and energy budgets constrain network size, while circuit physics limits parameter precision. Here we demonstrate input-dependent, within-trial neuromodulation of a spiking recurrent neural network on a mixed-signal neuromorphic processor. An off-chip modulator reads the on-chip network's input and activity and adjusts its population-level gains while the chip's analog circuits compute with the physics of the silicon. Neuromodulation raises binary spoken-digit classification accuracy by $+17.3 \pm 10.4$ percentage points, tested across five hardware network instantiations.

Aashman Mehta +4PDF
arXivhardware-design

HippocamFLEX: FPGA Prototyping of Hippocampus-Inspired Architecture for Context-Aware Sequence Prediction

Unlike conventional architectures that treat memory as static pattern storage, the hippocampus organizes experience as ordered sequences and retrieves them through temporally compressed replay spanning prior, current, and anticipated states. This capacity of context-appropriate sequence prediction remains absent from conventional deep learning accelerators and existing neuromorphic fabrics. We present HippocamFLEX, a fully synthesizable digital implementation of these hippocampal computational principles, realized with SystemVerilog RTL on the AMD Zynq UltraScale+,XCZU9EG. The design maps the hippocampal trisynaptic pathway onto three pipeline stages: a Dentate Gyrus / Entorhinal Cortex (DG/EC) stage for hypervector proximity matching, a CA3 stage for context-gated recurrent sequence prediction with multi-phase temporal tagging, and a CA1 stage that integrates past, present, and predicted future events into a theta-sequence inspired output. While much brain-inspired hardware relies on analog mixed-signal design for energy efficiency, this work targets FPGA prototyping as a deliberate architectural choice: FPGAs offer rapid design-space exploration without tape-out cost, bit-exact determinism that simplifies verification. The design supports 64 events across 16 contexts and is verified on physical silicon across context-isolation, temporal-ordering, and mid-sequence sliding-window test cases and is inherently scalable to larger event and context sets.

Bhuvanesh S. Kayarabettu +6PDF
arXivedge-ai

Predictive Maintenance at the Edge: An Embedded Anomaly Detection Pipeline Trained with Quantum Optimization

This article presents a predictive-maintenance pipeline that detects anomalies in bearings from vibration signals, adapted for deployment at the edge on low-cost microcontrollers, and that uses quantum optimization in two steps of its training. The pipeline consists of a step of vibration feature extraction and a step of classification with a one-class support vector machine (OCSVM). The features are selected by minimum redundancy and maximum relevance (mRMR) with the variational quantum eigensolver (VQE), under a limit on computing time. The OCSVM is adapted for inference at the edge without floating-point operations and is trained by a variational quantum algorithm with Pauli-correlation encoding (PCE). In small-scale tests on a simulator, with the Paderborn bearing dataset, the VQE found the optimal selection in 36% of the runs with Adam, and PCE chose better support vectors than chance, until a classical local search made the two equal. A classical baseline, the usual OCSVM trained by SMO with its weights rounded, matched them without any search.

Filipe Chagas FerrazPDF
Scholarsynthesis-pnr

TeethMark: Enabling Tongue-Tooth Interaction with Capacitive Sensing and Electrotactile Feedback using an Intraoral Wearable

We present TeethMark, an intraoral interface that enables tongue–tooth interaction by leveraging the tongue as the active medium for gesture input and electro-tactile feedback, while utilizing teeth as positional cues for localization. By applying a flexible PCB on a 3D-printed brace, TeethMark can recognize tapping, sliding, and area-touch gestures, and deliver two forms of electrotactile feedback: co-located feedback when the tongue touches the inner side of the teeth, and always-available feedback on the lip-facing side. With our design tool, users can customize and generate their own flexible PCB layouts, selecting interaction locations at the level of individual teeth. We conducted a formative study examining the physical comfort and social acceptability of tongue gestures across different tooth locations. The validation study evaluates the perception of electrotactile feedback and the accuracy of tongue–tooth gesture recognition. Our results indicate that TeethMark provides a promising modality for discreet, hands-free interaction. We further demonstrate its versatility through a set of applications, including language education, VR sensory sharing, gamified myofunctional therapy, and fabrication tool support.

Yi-Jing Jiang +51
Scholarhardware-design

Design and Simulation of a Five Stage Pipelined RISC-V Processor using Vivado

This paper describes the design, implementation and validation of a pipelined processor with five stages that uses the RV32I set of instructions. The proposed system increases the efficiency of the execution by allowing overlaps between the phases of the instruction-Instruction Fetch, Decode, Execute, Memory Access and Write Back, therefore increasing throughput than a single-cycle design of the same system without raising the clock frequency. It has been modeled using System Verilog and synthesized on an FPGA platform using Xilinx Vivado 2023, with both functional correctness and hardware feasibility. All types of instructions are supported: arithmetic, logical, control flow, and memory operations, immediately values and interaction between registers are correctly handled. Mechanisms like data forwarding, hazard detection, stalling, and pipeline flushing are added to ensure the correct execution in a concurrent environment to resolve data and control dependencies. Individually verified components of the complete data path, such as the ALU, control unit and memory interface, are then integrated together. The outcomes of the simulations prove the adequate instruction flow, efficient hazard assessment and sufficiently functioning in a variety of test conditions. The offered design can be characterized by a high efficiency of resources use, as well as low energy consumption on the chip level which reaches about 43 mW. In general, the offered processor is a scalable and energy optimized implementation that offers a powerful starting point in the development of future greater implementations like smart cache and RV32M/F extensions.

V. A. Chowdary +50
Scholarhardware-design

Design of a Low-Latency VLSI Router with Dual-Buffer Packet Handling Mechanism

High traffic loads can degrade router performance by causing packet accumulation, increased latency, and buffer overflow, particularly in conventional single-buffer routers. This paper presents a low-latency VLSI router architecture based on a dual-buffer packet-handling mechanism. By separating packet reception and transmission, the proposed design improves packet management and reduces congestion under heavy traffic conditions. The architecture is implemented in Verilog HDL and evaluated using Xilinx Vivado simulation. Results show a 20–35% improvement in throughput and a 15–25% reduction in latency compared to conventional router designs. Under congested conditions, packet loss decreases by 40–60%, while buffer utilization improves by up to 50%. Additionally, clock gating and selective activation reduce power consumption by 30–45%. These results demonstrate that the dual-buffer router is an effective solution for enhancing packet transmission efficiency in high-speed communication networks and Network-on-Chip applications.

P.Shruthi Sharanya +10
Scholarsynthesis-pnr

AreaCode: Encoding Conductive Tags Via Surface Area Modulation For Inductive Sensing

Contact-based object recognition enables interactive systems to infer how users engage with physical objects. However, existing inductive sensing approaches are limited to conductive materials or rely on geometric tag patterns that require sensor arrays, restricting their use on small or irregular objects. We present AreaCode, a method that expands the design space of conductive tags by encoding identity through conductive coverage within the sensing region. This approach enables tags with identical shapes and materials to produce distinct and reliable signatures, even when sensed using a single inductive coil. Through a series of experiments, we characterize material choices, pattern designs, and encoding capacity, demonstrating up to 17 distinguishable tag levels with high recognition accuracy. We further demonstrate applications on everyday objects and surfaces, including small items that cannot accommodate conventional sensor grids. Our results highlight how modulating conductive coverage enables expressive, passive tagging and extends inductive sensing to a broader range of interactive scenarios.

Zi-Hang Chen +10
Scholarsynthesis-pnr

ProtoCall: Preemptive Physical Feedback for Sequential Skill Augmentation

Mastering sequential skills is essential for everyday and professional tasks such as cooking and assembly. However, executing actions in the correct order and with appropriate timing remains challenging, particularly during transitions between consecutive steps. Existing approaches, including haptic guidance, error-based feedback, or post-error correction, either increase cognitive load, reduce users’ agency, or intervene too late to prevent errors. We developed ProtoCall, a custom wearable physical feedback system, combined with a Vision–Language Model (VLM) for interaction inference, and designed three feedback strategies targeting different stages of the error process: 1) preemptive attentional nudging (attention-level guidance by head actuation), 2) preemptive reaching nudging (action-level prevention of reaching), and 3) post-error nudging. We conducted a user study with 16 participants on a sequential step-transition task, revealing trade-offs between physical feedback strategies across augmentation, retention, and user experience.

Kyungyeon Lee +40
Scholarsynthesis-pnr

TwistForm: Shaping Reconfigurable 3D Surfaces via Pre-programmed Torsional Deformation of Chains

In this paper, we propose TwistForm, a novel shape-forming method that leverages torsion-actuated chains to reconstruct pre-programmed 3D surfaces. By parametrically designing the geometric interlocking of individual ring units, our approach encodes 3D surfaces into continuous, 1D kinematic chains. Upon the application of torsion, these chains undergo a structural phase transition, physically deploying into the pre-programmed 3D shapes. Because this 1D encoding inherently enables significant volumetric compression, TwistForm effectively addresses spatial constraints associated with the transportation and storage of physical objects. To realize this, we developed an end-to-end fabrication framework for designing and physicalizing these chains. Users can utilize this system to import arbitrary 3D models, design the corresponding interlinked chain structures, and export optimized print data for fabrication. We detail the underlying algorithms for unit design and spatial layout, and demonstrate practical use cases, including compact shipping, reusable molds, and shape-changing interfaces.

Hiroki Sato +20
Scholarsynthesis-pnr

MetaBeads: Enabling Shape- and Stiffness-Changing Interfaces through Beaded Metamaterials

As interactive systems increasingly extend into the physical world, designers need accessible ways to prototype interfaces that can change both shape and stiffness. In this paper, we investigate how simple, widely available materials—beads and string—can be used to construct shape- and stiffness-changing interfaces without specialized machinery. We introduce a set of reusable structural mechanisms that enable sturdy, wearable, and load-bearing structures from bead-based assemblies, each supporting a distinct mode of deformation and stiffness change, such as transforming from a flexible sheet into a jammed cuboid, a stiff tube, or a bistable curvature cell. To facilitate the creation of these structures, we provide an interactive design tool that guides craftspeople designing and fabricating their own MetaBead interfaces. We demonstrate the possibilities of MetaBeads through four application prototypes across wearable devices, educational tools, and adaptive furniture. Together, this work offers an accessible approach to prototyping shape- and stiffness-changing interfaces with everyday materials.

Yu-Jia Liu +50
Scholarsynthesis-pnr

A Gaussian Reinforcement Aggregation Network for Multimodal Classification

To address the heterogeneity in the distribution of discriminative features in remote sensing land-cover classification, we propose a Gaussian reinforcement aggregation network (GRAN). First, we design a sample spatial-spectral score decoupling (S3D) strategy based on category-specific characteristics. The S3D dynamically partitions the original dataset into multiple expert training subsets according to the spectral and spatial complexity of different categories, thereby enabling each expert to focus on modeling the discriminative boundaries of specific category groups. Second, a physics-inspired Gaussian derivative expert network (GDE-Net) is introduced and independently trained on each subset. The GDE-Net constrains the gradients in the spectral and spatial dimensions through low-frequency spectral derivative enhancement (LSDE) and high-frequency spatial derivative awareness (HSDA) modules, implicitly injecting physical prior information from the imaging process into the network, thus effectively extracting deep, fine-grained features highly correlated with the category. Finally, to mitigate the fragmentation caused by independently trained expert models and to achieve globally optimal classification performance, we propose an adaptive coupling factor aggregation network (ACFA-Net). The ACFA-Net adaptively integrates the complementary knowledge of multiple expert models. As a result, the final model exhibits strong global discriminative capability while maintaining high sensitivity to tail categories. Extensive comparative experiments conducted on multiple remote sensing datasets demonstrate that the proposed GRAN framework achieves superior classification performance. Our code is available at the following URL:https://github.com/ChenC1027/GRAN

Wen-Ping Ma +50IEEE transactions on circuits and systems for video technology (Print)
Scholarsynthesis-pnr

Predefined-Time Event-Triggered Practical Output Consensus Tracking of Multiagent Systems With False Data Injection Attack Under Digraph

This study addresses the predefined-time event-triggered practical output leader-follower consensus problem in high-order strict-feedback nonlinear multiagent systems (MASs) with unknown dynamics, subjected to a false data injection attack (FDIA), over a digraph. This problem introduces fundamental challenges, including estimating the leader’s states, achieving predefined-time consensus tracking under unreliable information due to FDIA, mitigating the influence of FDIA on the triggering conditions and consensus protocols, and excluding Zeno behavior that such an attack may induce. To address these challenges while designing MASs resilient to FDIA, this study proposes a new FDIA-resilient cooperative control framework that adopts a three-layer modular architecture comprising a physical layer, a communication layer and a control layer. At the control layer, a reputation-based attack detection module is proposed to detect FDIA. Then, the reputation value is transmitted to the event-triggered distributed observer module, and a new reputation-value-based event-triggered prescribed-time distributed observer is proposed to mitigate the effects of FDIA and estimate the leader’s states within the prescribed time. Afterward, based on the reconstructed leader’s states, an innovative relative threshold event-triggered predefined-time neural network resilient controller with minimum updating parameters is designed to achieve predefined-time practical output consensus tracking with improved control performance and to eliminate Zeno behavior. Theoretical analysis proves that the proposed control framework achieves time-guaranteed consensus tracking and excludes Zeno behavior in a resource-constrained and FDIA environment. Finally, the simulation results demonstrate the validity and advantages of the proposed control framework.

Jun-Kang Ni +40IEEE Transactions on Systems, Man & Cybernetics. Systems
Scholarsynthesis-pnr

Jacobian-Guided Weighting-Factor Design for MPC-Controlled DAB Converters

DAB converters are widely used in DC power conversion systems, but their strong coupling among power transfer, phase shift, and operating conditions complicates the controller design. Model predictive control (MPC) is attractive for DAB converters because it can explicitly handle multivariable dynamics and system constraints within a unified control framework. However, MPC parameters are still commonly tuned by empirical experience or heuristic rules, which may lead to limited interpretability, poor robustness, and even unstable operation under varying conditions; systematic stability-oriented design guidelines therefore remain limited. To address this issue, this brief proposes a Jacobian-guided weighting-factor design method for DAB converters. A discrete-time controller-oriented closed-loop mapping is established to relate MPC weighting factors to the local eigenvalue movement, from which a stability-oriented weighting-factor boundary is identified for the considered operating condition. The method converts MPC weighting-factor tuning from heuristic trial-and-error into an interpretable preselection procedure, and the same mapping can be extended to include additional physical states when full-system dynamics need to be evaluated. Experimental results validate the effectiveness of the proposed approach.

Yuan Li +30IEEE Transactions on Circuits and Systems - II - Express Briefs
Scholarsynthesis-pnr

A Fine-Grained Data Reconstruction Framework for Sparsely Sampled Industrial Time-Series

With the rapid development of the Industrial Internet of Things (IIoT), the large-scale acquisition of industrial time-series data has become increasingly feasible, providing valuable temporal insights for timely decision-making. Nonetheless, efficient data transmission is often constrained by communication bandwidth limitations. Consequently, the monitoring data are frequently sampled at sparse intervals in practical industrial scenarios, potentially obscuring fine-grained temporal patterns and reducing the effectiveness of downstream analytical tasks. To tackle this challenge, a novel fine-grained data reconstruction framework is proposed for sparsely sampled industrial time-series. Specifically, a semantic-aware loss function is designed to align network output with the ground truth while preserving structural patterns and semantic richness. Additionally, domain knowledge that incorporates measurement and correlation constraints is embedded into data reconstruction via an augmented Lagrangian-based learning paradigm. In this scheme, the reconstructed time-series maintain numerical alignment with coarse-grained observations and preserve intrinsic physical dependencies, facilitating reliable downstream analytical tasks. Extensive experiments on the commercial modular aero-propulsion system simulation (C-MAPSS) dataset, as well as a real-world steel conveyor roller dataset, demonstrate that the proposed framework outperforms state-of-the-art approaches across different downstream analytical tasks.

Yu-Di Zhang +50IEEE Transactions on Artificial Intelligence
Scholarmodel-compression

MURAL: Multimodal Uncertainty-aware Recommendation via Adaptive edge Learning

Multimodal Graph Neural Networks have become standard for recommendation by augmenting sparse interaction data with content features. Yet current architectures face two bottlenecks: structural rigidity, from a reliance on static precomputed similarity graphs that cannot adapt to evolving preferences; and semantic fragility, where noisy modality signals are indiscriminately fused, distorting the collaborative signal. We propose MURAL (Multimodal Uncertainty-aware Recommendation via Adaptive edge Learning), a unified framework that shifts multimodal recommendation from fixed structural augmentation to dynamic topology discovery. To address structural rigidity, an Adaptive Edge Learner combines a differentiable retrieval-augmented strategy with an approximate nearest neighbor search to discover latent item-item correlations that are both semantically adaptive and computationally scalable (O(NlogN)). To address semantic fragility, an Uncertainty-Aware Fusion module models the aleatoric uncertainty of heterogeneous modalities, dynamically down-weighting unreliable features while prioritizing high-confidence signals as a defense against cross-modal noise. We further employ a contrastive teacher-student alignment that anchors modality-specific representations to stable behavioral signals, ensuring optimization stability without gradient leakage. Experiments on large-scale benchmarks including TikTok and Amazon show that MURAL significantly surpasses both structural and generative state-of-the-art baselines, achieving superior accuracy while offering interpretability through domain-specific modality dominance and robustness under extreme data corruption.

Ahmad Mousavi +40
Scholarmodel-compression

A Unified Physics-Aware Quantum Machine Learning Framework across Power GaN HEMTs and Logic Nanowire FETs: Predicting Unseen Process Splits and Held-Out Geometry Combinations with Lower Error and Tighter Split-to-Split Variability

We present a unified reinforcement-learning (RL) framework that discovers compact parametrized quantum circuits (PQCs) for data-scarce device modeling. A graph neural network (GNN) policy optimized by proximal policy optimization (PPO) searches circuit architectures using leave-one-group-out cross-validation (LOGOCV) error on held-out process or geometry groups as the reward. The framework achieves the lowest mean absolute error (MAE) on all 11 targets versus six classical baselines, with 59% lower error (Ioff) and 81% tighter fold variability (VTH) for HEMTs and 84% lower error (VTH, SS, Ioff) and 82% tighter fold variability (Ioff) for NWFETs. These results demonstrate the potential of RL-selected, classically simulated PQCs as compact surrogates with low OOD error and improved physical consistency, despite imposing no explicit physical constraints, penalty terms, or device-specific equations, on the two evaluated device datasets.

Rushat Rai +120
Scholarmodel-compression

Temporal Spectral Analysis of Late-Time Error in a Physics-Informed Neural Network Solution of the One-Dimensional Advection–Diffusion Equation

Persistent late-time variation can remain in physics-informed neural network (PINN) solutions after the governing transient has effectively decayed, while conventional error norms do not reveal whether this variation has a systematic temporal–frequency structure. This study develops an offline temporal–spectral diagnostic and postprocessing workflow for a one-dimensional advection–diffusion benchmark. A high-accuracy analytical reference and three fixed-resolution finite-difference baselines are used to assess a PINN whose architecture is selected by a fully supervised neural architecture search and whose parameters are trained with progressive temporal windowing. Candidate late-time intervals are selected without using the reference solution by applying the Bayesian information criterion (BIC) to a breakpoint model for the inter-reconstruction sensitivity; the selected field is subsequently reconstructed by retaining a prescribed fraction of its temporal spectral energy and is evaluated independently through reference-error and physics-consistency measures. For [tcut,tmax]=[1.8,5], the zero-frequency component contains 0.9999996 of the raw-field energy, so the q=0.95 reconstruction retains only the temporal mean. This projection reduces the final-time spatial error norm from 2.70×10−3 to 1.06×10−3, a factor of approximately 2.5, while changing the discrete governing-equation residual by less than 0.3% over the filtered window. Mean-removed tests for q=0.90,0.95,0.99 show that the discarded fluctuation is dominated by low-frequency approximation error rather than high-frequency noise. The result supports the proposed selection–validation workflow for this controlled benchmark but does not establish a universally transferable filter.

David Díaz-León +30Mathematics
Scholarmodel-compression

AI-driven insights into electronic Word-of-Mouth and repurchase intention: A Neural Network approach in multisectoral contexts

Electronic word-of-mouth (E-WOM) is widely associated with consumer trust and repurchase intention, but evidence on how artificial intelligence (AI) and neural networks analyse these relationships remains fragmented across industries. This PRISMA-guided systematic literature review synthesizes research across eight sectors: healthcare, banking, tourism, education, insurance, real estate, entertainment, and food and beverages. The review examines sentiment analysis, repurchase intention, customer loyalty, and the fit between neural-network architecture and data structure. Searches identified 3,280 records; 230 studies were included in the qualitative synthesis, and 68 studies reporting quantitative model information were examined descriptively. This is a review of published studies rather than primary model training, and no pooled meta-analysis was conducted because datasets, outcomes, and validation procedures were heterogeneous. The synthesis indicates that contextual language models fit review text, graph neural networks fit relational or coordinated-review signals, multimodal networks require visual evidence, and reinforcement learning fits sequential recommendation settings. Cross-sector conclusions therefore emphasize recurring mechanisms and architecture-to-data fit rather than unsupported comparisons of accuracy values from non-equivalent studies.

A. Azizi +40
Scholarmodel-compression

Optimised Machine Learning for MASW-Based Zonation of Weak Alluvial Soils Using Anchor-Based SPT/SCPT Calibration

Weak alluvial soils are difficult to delineate where penetration-test data are sparse and stiffness varies rapidly with depth. This study presents a novel goal-attainment-optimised machine learning framework that combines multichannel analysis of surface waves (MASW) with anchor-based standard penetration test (SPT) and seismic cone penetration test (SCPT) calibration to improve the depth continuity of weak-soil zonation. A dimensionless potential weak-soil index, Pweak, was derived from SPT-N, SCPT cone resistance and MASW-derived shear-wave velocity for 265 site-depth observations from 13 Taipei Basin sites. SPT/SCPT measurements at predefined anchor depths within the 0–10, 10–20 and 20–30 m zones were condensed into a site-zone calibration coefficient and combined with representative depth and MASW-derived Vs as artificial neural network (ANN) inputs. An exhaustive architecture search evaluated 650 candidate architectures, after which the goal-attainment procedure selected a two-hidden-layer network with 16 and 22 neurons, yielding Rval=0.896, MSEval=0.019 and an overall correlation coefficient of 0.875. The ANN-derived zonation reproduced the principal weak and relatively strong intervals of the reference calculation and preserved local depth-dependent transitions more effectively than simple interpolation between sparse anchor levels. The framework therefore provides a practical means of extending anchor-based penetration-test calibration along MASW profiles for preliminary ground-risk interpretation.

A. B. Ekmen0Applied Sciences
Scholarmodel-compression

Firefly Algorithm for Medical Image Segmentation: A Systematic Review of Methods, Applications, and Future Directions

The firefly algorithm (FA) is a nature-inspired metaheuristic increasingly applied to medical image segmentation because of its flexible search mechanism and optimization capability. This paper systematically reviews FA-based approaches for medical image segmentation and related image-analysis tasks published between 2010 and 2026. A structured search of Scopus, PubMed, IEEE Xplore, Web of Science, and Google Scholar was conducted using predefined eligibility criteria. The included studies were categorized into multilevel thresholding, entropy-based segmentation, clustering-assisted methods, adaptive and chaotic FA variants, and hybrids with deep learning and other metaheuristics. Major applications included brain, lung, breast, and microscopic image analysis. The reviewed studies generally reported competitive performance and demonstrated FA's flexibility in optimizing thresholds, clustering centers, feature subsets, and model hyperparameters. However, direct quantitative comparisons remain difficult because of heterogeneity in datasets, imaging modalities, evaluation metrics, and experimental protocols. Major limitations include computational overhead, parameter sensitivity, inconsistent benchmarking, limited clinical validation, and insufficient investigation of three-dimensional and multimodal imaging. Future research should emphasize self-adaptive FA, GPU-accelerated and parallel implementations, integration with neural architecture search and explainable artificial intelligence, extension to three-dimensional and multimodal imaging, and validation on standardized multicenter datasets. Overall, FA represents a promising optimization framework for medical image segmentation, although methodological standardization, computational improvements, and rigorous clinical validation are required for broader practical deployment.

Mehdi Jalali-Varnamkhasti +10
Scholarmodel-compression

Artificial intelligence in breast cancer imaging: a systematic review of detection, segmentation, explainability, and clinical translation

Breast cancer remains the most commonly diagnosed malignancy in women worldwide, with approximately 2.3 million new cases and 670,000 deaths in 2022. Artificial intelligence (AI) is increasingly being applied across breast imaging for detection, characterization, segmentation, risk stratification, explainability, and clinical translation. This systematic review synthesized 284 eligible peer-reviewed publications published between 2015 and 2025, identified through a PRISMA 2020-compliant search of four indexed databases. A supplementary relevance-ranked Google Scholar screen was reported separately. A targeted May 2026 narrative update of recent prospective and implementation studies was also conducted without adding these records to the PRISMA denominator. The reviewed evidence encompassed classical machine learning and radiomics, convolutional neural networks, YOLO-family detectors, vision transformers and hybrid CNN-Transformer architectures, U-Net variants, Mask R-CNN, SAM, and MedSAM across mammography, tomosynthesis, MRI, ultrasound, contrast-enhanced mammography, and emerging photoacoustic imaging. The evidence highlights advances in explainable AI, multimodal and radiogenomic fusion, federated learning, and open-source deployment, while persistent challenges include data heterogeneity, class imbalance, domain shift, bias, incomplete calibration, and limited prospective validation. Retrospective benchmark findings were derived from the 284-study systematic corpus, whereas prospective clinical utility and real-world implementation were informed by the separate 2026 narrative update. AI shows substantial potential to enhance breast cancer imaging, but broader clinical translation requires robust external and prospective validation, improved calibration, assessment of generalizability and bias, and integration into clinical workflows. Supplementary Data Sheet S1 should be interpreted as a provenance-tagged evidence map rather than a formal study-level risk-of-bias assessment, with 38 entries verified from full text, 48 based on abstracts, and 198 based on DOI/source pages or other web-accessible records.

I. Khan +100
Scholarmodel-compression

Diagnostic accuracy of an artificial intelligence-driven cytopathological tool in detecting cervical pre-cancerous lesions from Pap smear images

Background Cervical cancer remains a leading cause of cancer death among women in sub-Saharan Africa, with Tanzania bearing a disproportionate burden. The critical shortage of trained pathologists, coupled with the unprecedented disease burden in low-resource settings, underscores the urgent need for point-of-care screening and diagnosis to enable timely decision-making. We assessed the diagnostic accuracy of AI-driven cytopathological tools to improve diagnostic efficiency and accessibility. Methods This retrospective secondary data analysis evaluated five convolutional neural network (CNN) architectures: EfficientNetB7, MobileNet, ResNet50, ResNet152, and InceptionNetV3, for semi-automated classification of cervical cell abnormalities. A total of 11,955 Pap smear cytological images were used from the publicly available Center for Recognition and Inspection of Cells (CRIC) Searchable Image Database, spanning six cellular classes: Normal, ASC-US, LSIL, ASC-H, HSIL, and carcinoma. Performance was assessed on a hold-out test set (n = 961) using macro-averaged metrics. Results EfficientNetB7 achieved the highest overall performance, with a macro F1 score of 0.9324 (95% CI: 0.920–0.945), an accuracy of 0.9775 (95% CI: 0.968–0.987), and a sensitivity of 0.9324 (95% CI: 0.920–0.945). ResNet50 ranked second (F1 score: 0.9282 [95% CI: 0.916–0.941]) and ResNet152 third (F1 score: 0.9240 [95% CI: 0.911–0.937]), showing minimal performance gaps. InceptionNetV3 followed closely (F1 score: 0.9220 [95% CI: 0.909–0.935]). MobileNet achieved the lowest F1 score (0.8918 [95% CI: 0.875–0.908]), but its lightweight architecture is suitable for edge deployment. The Carcinoma (CA) class achieved near-perfect recall across all models (> = 0.978). Notable interclass confusion was observed between ASC-US and LSIL, attributable to cytomorphological overlap. Conclusion EfficientNetB7 offers promising diagnostic accuracy for automated cervical cancer screening using Pap smear images and shows potential for integration into point-of-care workflows in low-resource settings. Future work should focus on training models on locally representative datasets and exploring whole smear analysis to reduce interclass misclassification.

S. Sang'udi +100Frontiers in Digital Health
Scholarmodel-compression

Closing the Semantic-Edge Gap: Tiny Language Models for 6G Wireless Intelligence

Sixth-generation (6G) wireless networks are envisioned as AI-native systems in which semantic communication - transmitting task-relevant meaning rather than raw bits - moves beyond Shannon's classical bit-pipe model. Large language models (LLMs) dominate semantic encoding but are unsuitable for 6G user equipment and IoT devices, given prohibitive memory, energy, and latency costs. Tiny language models (TinyLMs) - compressed via TinyML techniques into kilobyte-to-megabyte memory and milliwatt power budgets - are the missing bridge between LLM-level semantic encoding and 6G edge hardware, yet no prior work systematically maps TinyML techniques onto semantic communication architectures for this purpose. This survey closes that gap through a two-axis taxonomy connecting six compression families (quantization, pruning, knowledge distillation, low-rank adaptation, neural architecture search, hybrid pipelines) to five semantic communication architectures (end-to-end joint source-channel coding, split learning, federated learning, knowledge-graph-assisted, and multi-task/cross-modal communication), synthesized with a quantitative meta-analysis of the model-size-versus-semantic-fidelity Pareto frontier. Representative results include a CNN-Transformer encoder achieving 22 dB PSNR at 33.33% semantic-representation size reduction; a symbolic protocol machine reducing a neural MAC protocol from 4.55 MB to 1 KB (99.98% smaller) with zero performance loss; federated bidirectional knowledge distillation converging under joint model-and-data heterogeneity where FedAvg-style averaging underperforms; and knowledge-graph-assisted probability graphs cutting transmission energy by 65%. The survey identifies nine open research challenges for TinyLM-enabled 6G semantic communication, including two not previously articulated in the literature.

Srikanth Kamath +30
Scholarmodel-compression

Animal Image Classification Using Deep Learning Algorithms

The process of animal image classification plays a vital role in establishing essential requirements that enable effective monitoring of ecosystems and the preservation of biodiversity and agricultural practices and ecological research work. The task of animal face recognition faces difficulties because animals from the same species present different facial features which results in difficulties to identify their species through the use of facial recognition technology. The research presents a deep learning-based system which uses convolutional neural networks and transfer learning techniques together with document enhancement methods to create an animal concept classification system. The main objective of this work search out create a precise and robust classification system which will evaluate its performance against existing deep learning-based systems. The research provides three main contributions which include the creation of a systematic exercise that tests various deep learning architectures in a standardized research environment and the application of transfer knowledge to enhance feature representation and the establishment of a comprehensive achievement assessment system that uses standard versification. The proposed model reaches a 95.4% accuracy rate which surpasses the performance of conventional CNN models that were developed from scratch (88.2%) and standard transfer learning models (91.6%) and inconsequential CNN architectures (89.4%) and ensemble-based methods (93.1%). The proposed method reaches high accuracy (95.1%) together with recall (94.6%) and F1-score (94.8%) which demonstrates its capability to properly categorize data while generalizing across various situations. The research demonstrates that the deep learning foundation functions as a dependable and scalable solution for tasks involving animal representation categorization

Wurood A. Jbara0
Scholaredge-ai

MIRA: An Edge-Cloud Multimodal LLM Framework for DESH Assessment in IoT-Enabled MRI Workflows

Connected MRI scanners operate as networked medical devices within DICOM-based healthcare Internet of Things (IoT) workflows, but conventional connectivity primarily moves images to a Picture Archiving and Communication System (PACS) rather than adding task-specific reasoning near acquisition. Building on DESH-App 2.0, a preceding machine-learning contribution accepted for oral presentation at SIN Firenze 2026, this paper specifies MIRA (Multimodal Imaging Reasoning and Assessment), a vendor-neutral edge-cloud module for assessing disproportionately enlarged subarachnoid spaces (DESH). A hospital-edge gateway validates and de-identifies 3D T1-weighted DICOM studies, derives ventricular, high-convexity, and Sylvian-fissure masks with SegFormer3D, and converts each volume into a compact evidence package comprising 12 representative slices with overlays, 17 global, zonal, and ratio features, and quality-control tokens. LLaVA-Med-7B and Lingshu-7B are adapted through 4-bit QLoRA to return constrained JSON containing class, confidence, named anatomical evidence, and an abstention or quality warning. The edge-cloud partition keeps raw volumes inside the institutional network; only the compact evidence package may reach an authenticated inference service, while structured outputs return to the PACS workstation. The evaluation protocol jointly covers diagnostic discrimination, calibration, evidence agreement, unsupported claims, abstention, latency, transmitted bytes, memory demand, link-loss recovery, and multi-device interoperability. MIRA is designed to add auditable multimodal reasoning to connected MRI workflows without modifying safety-critical scanner controls.

Salvatore Quattropani +50
Scholaredge-ai

TinyShareNet: Toward Extremely Tiny Network for Memory Efficient On-Chip Inference at the Edge

The recent progress in TinyML technologies facilitates the deployment of machine learning on low-cost edge devices strictly constrained by computation, memory and energy. However, the extremely limited on-chip memory makes it hard to balance efficiency and performance during the inference phase. To address this issue, we propose a memory-efficient framework termed TinyShareNet that enables convolution groups with minimal parameters to achieve powerful performance. Motivated by sparse dictionary learning, we present an innovative linear convolution sparse representation (LCSR) paradigm to construct linear convolution groups for improving the expressive ability. On this basis, we propose a novel parameter sharing strategy across stacked linear convolution groups for significantly shrinking the parameter size. For further promoting the precision of parameter sharing tiny network, we put forward an adaptive orthogonality regularization (AOR) method on shared basis convolutions to learn more diverse features. Extensive experiments demonstrate that our proposed TinyShareNet achieves outstanding precision compared to state-of-the-art lightweight networks with a record-low storage cost. When deployed on popular microcontroller units (MCU), TinyShareNet yields significantly higher precision with over $1.56\times $ higher frame rate and $1.88\times $ lower energy consumption than classical solutions, greatly promoting the development of on-chip deep learning inference at the edge.

Tong Wu +40IEEE transactions on circuits and systems for video technology (Print)
Scholaredge-ai

Containerized Anomaly Detection Framework for IoT Networks Using Machine Learning

The increasing use of Internet of Things (IoT) devices generates continuous sensor data that requires efficient monitoring and analysis. Detecting abnormal sensor readings is important for identifying unusual behavior and supporting reliable IoT operations. This paper presents a Containerized Anomaly Detection Framework for IoT Networks Using Machine Learning. The proposed system groups continuous sensor readings into fixed-size logical containers and applies machine learning techniques to identify anomalies. Isolation Forest is used as the primary unsupervised anomaly detection algorithm, while Local Outlier Factor (LOF) and One-Class Support Vector Machine (One-Class SVM) are available for comparison. The framework supports real-time humidity monitoring using an Arduino Uno with a DHT11/DHT22 sensor and dataset-based analysis through CSV uploads. Data preprocessing, feature extraction, container-level aggregation, SQLite storage, and a Flask-based web dashboard are integrated into a single application. The system displays sensor readings and anomaly results and supports report generation. The proposed framework provides a lightweight approach to IoT sensor monitoring and anomaly analysis.

A. K. Dr +40
Scholaredge-ai

Hardware–Software Co-Design for Reliable and Secure TinyML on Microcontrollers: A Narrative Review

Aims: To review how hardware–software co-design can make tiny machine learning (TinyML) on microcontrollers not only efficient but also reliable and secure, to classify existing approaches by design objective and stack layer, and to propose an integrated framework and research agenda for dependable on-device intelligence. Study Design: Narrative (non-systematic) review. Methodology: Peer-reviewed journal articles and conference papers, supplemented by a small number of influential preprints, technical reports, standards and one book, were identified from IEEE Xplore, the ACM Digital Library, ScienceDirect, SpringerLink, arXiv and Google Scholar. Studies were included if they addressed machine learning on microcontrollers or reported a reliability or security technique or finding that transfers directly to such devices. The 61 sources retained, published between 1999 and 2026, were classified by design objective (efficiency, reliability, security), layer of the hardware–software stack, platform and type of evidence, and synthesised narratively. Results: Co-design of network architectures, quantisation, kernels, memory scheduling and hardware has made image-, audio- and sensor-based inference feasible within kilobytes of memory and milliwatts of power. However, reliability and security are usually treated as separate afterthoughts. Fault-injection studies show that the impact of soft errors depends strongly on data representation, bit position and layer, and that low-cost measures such as activation range restriction and selective hardening can mitigate many silent data corruptions. However, most of this evidence comes from software fault injection, often on platforms other than microcontrollers. Security studies demonstrate that physical access enables electromagnetic side-channel recovery of model architectures, instruction-skip and bit-flip attacks that alter predictions, and model extraction through query interfaces. Recent laser fault-injection studies on Arm Cortex-M devices have demonstrated practical bit-flip attacks and the extraction of model parameters from flash memory, although under laboratory conditions. Hardware isolation, secure boot, authenticated updates, masking and control-flow protection are available but rarely evaluated together with accuracy, latency and energy. Conclusion: Dependable TinyML requires reliability and security to be treated as first-class design objectives alongside accuracy and efficiency throughout the design, verification, deployment and operation lifecycle. Standardised fault and threat models, benchmarks that report robustness and security costs, and multi-objective co-design tools are the main priorities for future research. Because the evidence is heterogeneous and the review is narrative, these conclusions are indicative and should be confirmed by experimental validation of the proposed framework.

Shamsuddeen Mohammed Sada +20Journal of Engineering Research and Reports
Scholaredge-ai

AI-IoT Enabled Smart Agriculture Using Sequential Hybrid Feature Selection for Intelligent Irrigation Management

Efficient irrigation management is a prerequisite of sustainable agriculture, especially when considering water constraints and climate variability. Recent breakthroughs in Artificial Intelligence (AI) and Internet of Things (IoT) technology have made it possible to develop smart agricultural systems that monitor environmental conditions and automate irrigation decision-making. However, high-dimensional agricultural sensor data usually contain redundant and unnecessary features, which significantly impair machine learning model performance. This paper presents a smart agriculture system based on AI-IoT with an intelligent irrigation control using a Sequential Hybrid Feature Selection (SHFS) method. The SHFS model sequentially combines Principal Component Analysis (PCA), Particle Swarm Optimization (PSO) and L1-regularized Logistic Regression (Lasso) to minimize the dimensionality and preserve the discriminative information. The suggested framework was evaluated on 100,000 IoT-based agricultural observations consisting of ambient as well as soil-related factors. We evaluated the prediction capability of some classifiers, such as k-Nearest Neighbors (kNN), Na¨ıve Bayes, Logistic Regression and XGBoost. The experimental results show that the SHFS framework can achieve a significant improvement in classification performance compared to traditional feature selection methods. The best accuracy was 97.5%, attained by XGBoost in combination with SHFS, which also had higher precision, recall, and F1-score values than the other models. The collected findings demonstrate that the suggested framework efficiently reduces data redundancy, enhances the reliability of irrigation forecasts, and promotes sustainable applications of precision agriculture.

Hussein A. Abdul Wahid +30Sakarya University Journal of Computer and Information Sciences
Scholaredge-ai

Explainable deep learning for predicting temperature breaks and identifying their causal drivers in fresh produce cold chains

This study proposes an integrated framework combining deep learning, causal discovery, explainability, and domain thresholds for proactive monitoring in temperature-sensitive fresh fruits and vegetables cold chains, a domain that remains largely unexplored. The framework predicts temperature breaks (a critical physical hazard that compromises food safety and accelerates postharvest losses), prospectively identifies their causal drivers, and prioritizes countermeasures. Using two IoT-gathered datasets from an Apple cold room in South Africa, eight machine learning models and three baselines were compared for temperature prediction, including a proposed LSTM-GRU hybrid (LSGR) model. Causal relationships between measured parameters and predicted temperature were identified using Latent Peter & Clark Momentary Conditional Independence (LPCMCI; compared with PCMCI). At the same time, SHapley Additive exPlanations quantified feature importance, and domain-based thresholds complemented them for countermeasure prioritization after temperature-break prediction (TBP). On the extended dataset, deep neural networks' (DNN) R 2 and normalized mean squared error improved by up to 30% and to 61%, respectively, over the smaller dataset, demonstrating strong generalizability to substantially different cold room operating conditions. The LSGR demonstrated the best overall performance across both temperature and TBP tasks among the recurrent models assessed, raising the fewest false alarms per day. All DNNs flagged breaks in advance; no baseline did. LPCMCI recovered most direct causal links after break prediction, especially for lagged temperature, identifying it as the dominant factor associated with predicted breaks. Notably, post-TBP, causality, importance, and thresholds strongly agreed on the same top two critical parameters or countermeasures (temperature and humidity), providing consistent and actionable insights for decision-making. These findings demonstrate the viability of combining causal discovery, explainability and thresholds for cold chain management, with implications for reducing food wastage and enhancing quality, safety, and sustainability from farm to fork in temperature-sensitive FFVs supply chains. LPCMCI recovered a few links overall, indicating the influence of unmeasured drivers and warranting further investigation, including additional features, their interactions, and seasonal data, as well as a fuller characterization of causality-regime dependence.

Jeremiah Taguta +20Frontiers in Artificial Intelligence
Scholaredge-ai

A lightweight and explainable framework for prompt injection detection in AI applications

Large language models (LLMs) are increasingly embedded in chatbots, document assistants, retrieval-augmented generation (RAG) pipelines, and tool-using applications. These systems may interpret untrusted user or retrieved text as instructions, creating a prompt-injection risk. A practical detector must identify attempts to redirect model behaviour while allowing harmless security explanations, quotations, and ordinary instructions. This paper proposes a lightweight and explainable text-screening framework based on word- and character-level TF-IDF features, classical machine-learning classifiers, and a transparent rule-based baseline. The study compares Logistic Regression, linear Support Vector Machines, and a combined word-character model under source-separated evaluation. Particular attention is given to hard-benign inputs, controlled obfuscation, duplicate leakage, calibration, and inference cost. Candidate data sources include public prompt-injection examples, NotInject-related resources, and context-rich benchmark materials, subject to access, label, overlap, and licence checks. The framework returns a risk estimate, a concise evidence summary, and a review recommendation for uncertain inputs. It is intended as a screening layer, not a replacement for instruction separation, access control, least privilege, or human approval. Since implementation and testing are proposed work, no accuracy or performance improvement is claimed in advance. The intended contribution is a reproducible low-compute evaluation of when compact detectors are useful, where they over-block benign text, and how reliably their behaviour transfers to unseen sources.

Vats Vaibhav +10International Journal of Innovative Research in Engineering
Scholaredge-ai

Integrating Fuzzy Rule Based Reasoning with Machine Learning for Intelligent and Scalable Irrigation Control

This study presents a hybrid intelligent irrigation decision support system that integrates fuzzy logic and Random Forest (RF) regression to support irrigation decision-making in agriculture. A fuzzy inference engine was designed using expert-defined rules based on soil moisture, air temperature, and rainfall. This system generated irrigation recommendations on a scale of 0 to 10 using Mamdani-type fuzzy logic. An RF model was trained on the same inputs with the fuzzy output as the target to evaluate how accurately a data-driven model can reproduce the fuzzy decision mapping. Using five-fold cross-validation, the RF achieved a root mean squared error (RMSE) of 0.2459 , mean absolute error (MAE) of 0.0801 , and coefficient of determination (R2) of 0.9682, indicating strong alignment with fuzzy decision logic. This hybrid approach offers a transparent, data-driven, and scalable solution, effectively bridging the gap between interpretability and prediction accuracy. The system has the potential to support precision irrigation and informed water management decisions and was evaluated using one year of daily aggregated Internet of Things (IoT) sensor observations from Narrabri, Australia.

Bharati Wukkadada +60Journal of Artificial Intelligence and Technology
Scholaredge-ai

AI-Based Smart Microgrid Energy Management System

The AI-Based Smart Microgrid Energy Management System aims to improve energy utilization and power management using IoT and machine learning. The system integrates solar power, battery storage, and transformer supply, with an ESP32 monitoring electrical and environmental parameters. MQTT enables real-time data communication, while a web dashboard supports remote monitoring and load control. Relay-based switching manages power sources according to predefined operating conditions. An LCD displays system status, and a Python-based backend stores data and generates alerts. Machine learning is proposed for estimating battery and transformer health to support predictive maintenance. The system provides a cost-effective approach to smart energy monitoring, automated power management, and reliable microgrid operation

Prof. Manisha landge, Rahane Sayali Babasaheb +20International Journal of Advanced Research in Science, Communication and Technology
Scholaredge-ai

A labeled dataset for resource scheduling in heterogeneous IoT-fog-cloud computing systems

Fog computing brings computation closer to IoT devices to reduce latency, but assigning tasks to the right execution node demands scheduling decisions that are difficult to automate without labeled training data. Machine learning-based schedulers require such labels, yet publicly available datasets either carry no assignment labels, record only post-execution metrics, or assign binary tier labels that discard node-level information. This paper describes FogMLS, three labeled tabular datasets generated by a genetic algorithm scheduler with greedy initialization across a three-tier IoT-fog-cloud simulation, modeling three heterogeneous fog nodes and a cloud server. Three load configurations are deposited under fixed random seed 42, each spanning 5,000 scheduling slots. The primary high-load dataset contains 404,919 task-resource assignments across 200 task positions per slot, with pre-execution features encoding task type, instruction count, data size, and the live remaining capacity of each fog node at scheduling time, alongside GA-derived node assignment labels. A Random Forest classifier achieves 76.35% overall accuracy across four scheduling classes with consistent results across seeds 42, 123, and 456 (76.14-76.35%), confirming that the labels carry sufficient signal for training a scheduling model.

Harry John +10Scientific Data
Scholarai-hardware

Second-Order Training for Neural Networks Using RRAM-Based Analog Matrix Computing

Second-order training methods leverage curvature information to achieve faster and more stable convergence than first-order approaches. However, their practical adoption has been limited by the prohibitive computational cost of inverting the second-order information (SOI) matrix. In this work, we present an experimental demonstration of second-order training enabled by RRAM-based analog matrix computing (AMC). We validate the proposed optimizer by training a two-layer convolutional neural network (CNN) for handwritten letter classification, reducing the required training epochs by 26% and 61% compared to SGD and Adam, respectively. On a larger-scale task, the proposed accelerator achieves a <inline-formula> <tex-math notation="LaTeX">$23.3\times $ </tex-math></inline-formula> improvement in throughput and a <inline-formula> <tex-math notation="LaTeX">$16.0\times $ </tex-math></inline-formula> gain in energy efficiency over state-of-the-art digital processors. These results demonstrate the potential and effectiveness of RRAM-based AMC for second-order neural network training, paving the way toward more efficient AI training hardware.

Sai-Tao Zhang +70IEEE Electron Device Letters
Scholarai-hardware

Integrated Low-Power Self-Coherent Kramers--Kronig Millimeter-Wave Receiver With Nonlinear Analog and Mixed-Signal Computing

A critical challenge in scaling up high-frequency wireless communication in the millimeter-Wave (mmWave) spectrum has been poor energy efficiency of the constituent transceivers, particularly on the highly energy constrained user equipment (UE). Receiving spectrally-efficient wireless signals (e.g., 64-quadrature amplitude modulation (QAM) etc.) require coherent transceivers with programmable frequency synthesizers, carrier recovery techniques, and carrier synchronization, all of which pose significant energy and latency burden on the wireless systems. Here, we present a different approach to allow low-power mm-Wave/sub-THz receivers, where such complex frequency synthesis, control and conversion, can be eliminated in favor of a simple amplitude detection, and energy-efficient nonlinear analog and mixed-signal computing fabric. The approach, inspired by the Kramers–Kronig (KK) principle, allows us to estimate phase from the amplitude detector by establishing a Hilbert relationship between the logarithm of amplitude and phase. The chip implemented in a 65-nm CMOS process achieves coherent detection of spectrally-efficient modulation (e.g., 16-QAM) with Gb/s data links at tens of milliWatt of DC power, without requiring a receiver LO. The approach can inspire new innovations in low-power RF, mmWave and high-frequency wireless receivers.

Sherif Ghozzy +40IEEE Journal of Solid-State Circuits
Scholarai-hardware

GRAS: General-Purpose and Reconfigurable Analog Solver for Matrix Computing

Matrix computation is ubiquitous in modern scientific and intelligent applications. Conventional digital computing architecture faces fundamental limitations in the post-Moore era, prompting intensive exploration of alternative computing paradigms. In-memory analog matrix computing (AMC) with resistive random access memory (RRAM) represents a highly promising solution, as it enables performing matrix computations in one step while also reducing the cost of data movement between memory and processor. However, the existing AMC circuits each have a specific connection topology to implement a single computing function, and lack the universality of a matrix processor. In addition, equipping every AMC circuit with a set of analog amplifiers causes an overhead of hardware resources. In this work, we design a reconfigurable AMC macro for general-purpose matrix computations, which is achieved by configuring proper connections between the memory array and amplifier circuits. Based on this macro, we develop a hybrid system that incorporates an on-chip write-verify scheme and digital functional modules, thus presenting a general-purpose AMC solver for various applications. Typical matrix computations, such as matrix-vector multiplication (MVM), matrix inversion (INV), generalized inverse (GINV), and eigenvector (EGV), are examined with this solver. Relevant applications, including convolutional neural networks (CNNs) and BlockAMC, are also demonstrated. Compared to the state-of-the-art GPU for solving linear systems with <inline-formula> <tex-math notation="LaTeX">$512\times 512$ </tex-math></inline-formula>, the throughput and energy efficiency are improved by <inline-formula> <tex-math notation="LaTeX">$80\times $ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$14\times $ </tex-math></inline-formula>, respectively.

Lun-Shuai Pan +30IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Scholarai-hardware

COSH: An Efficient Cross-Snapshot Accelerator for High-Performance Dynamic Heterogeneous GNN

Dynamic heterogeneous GNNs (DyHGNNs) are increasingly employed to model the complex relationships between various types of vertices and edges in real-world applications. When performing DyHGNN inference, existing software/hardware solutions adopt the snapshot-by-snapshot execution approach and suffer from frequent and irregular access to identical neighbors for the same vertex across multiple snapshots, leading to redundant computations and irregular memory accesses. To address these challenges, we propose an efficient cross-snapshot hardware accelerator, i.e., COSH, which facilitates high-performance execution of DyHGNN inference. Specifically, we integrate a redundancy-aware asynchronous filter-load-aggregate-gather (FLAG) execution model into the accelerator design to minimize redundant computation and off-chip memory accesses. FLAG traverses the neighbors of the same vertex across different snapshots to identify all involved vertices and filter out changes in neighbors between consecutive snapshots. Then, it computes intermediate aggregation features for all encountered neighbors and incrementally calculates the results for the vertices in each snapshot, eliminating redundant computations and regularizing memory accesses. To minimize redundant accesses for vertex features, we also design a hierarchical data loading mechanism to provide a parallelized cross-snapshot filter (PCF) to enhance the efficiency of processing graph structures. Besides, a coordinated feature aggregator (CFA) is also proposed to synchronize feature aggregation and gathering, ensuring the timely utilization and management of generated intermediate features to boost computational parallelism. Our experimental results demonstrate that, compared to state-of-the-art software and hardware solutions, COSH achieves performance improvements of <inline-formula> <tex-math notation="LaTeX">$27.1\times $ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$3.6\times $ </tex-math></inline-formula> on average, and also energy efficiency gains of <inline-formula> <tex-math notation="LaTeX">$29.5\times $ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$3.7\times $ </tex-math></inline-formula> on average, respectively.

Qi-Yuan Niu +50IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Scholarai-hardware

MTDiff: A Multi-Task Diffusion Accelerator With Bandwidth-Aware Memory Partition and Efficient Digital Compute-in-Memory

Initially applied to image synthesis, diffusion models have rapidly expanded into various content generation tasks, such as 3-D scene reconstruction and video generation, due to the superior generative quality. Based on our observations, the imbalanced operational intensities across diffusion model layers lead to suboptimal hardware utilization due to inadequate bandwidth optimization. Additionally, emerging 3-D and video generation tasks require computationally intensive multi-view consistency operations to ensure smooth transitions between perspectives and frames. In this work, we present <italic>MTDiff</italic>, a multi-task diffusion accelerator featuring three key innovations: 1) a dynamic bandwidth-aware memory partitioning scheme optimizing hardware utilization and reducing external memory accesses; 2) an efficient bitline (BL)-segmented compute-in-memory (CIM) macro together with a reuse-aware weight update scheduler; and 3) a hierarchical multi-frame consistency management (MFCM) skips redundant operations at frame and pixel levels to improve computational efficiency. Implemented using a 22-nm process, the <italic>MTDiff</italic> accelerator achieves a 49.74–60.81 TFLOPS/W system efficiency across various generative tasks. It delivers a <inline-formula> <tex-math notation="LaTeX">$1.4\times $ </tex-math></inline-formula> performance improvement and a <inline-formula> <tex-math notation="LaTeX">$1.52\times $ </tex-math></inline-formula> system energy efficiency enhancement compared to prior state-of-the-art (SOTA) diffusion accelerators.

Jia-Qi Zhou +50IEEE Journal of Solid-State Circuits
Scholarai-hardware

A Two-Step Time-Domain Logic-Compatible Embedded-Flash Computing-in-Memory Macro for On-Device AI

Computing in memory (CIM) has emerged as a promising architecture that can significantly reduce power consumption and latency in edge artificial intelligence (AI) devices. For battery-constrained edge devices, non-volatile memories (NVMs) are ideal due to their ability to retain neural network parameters even when powered down, ensuring rapid response time upon device wake-up. However, the tradeoff between energy efficiency and computing precision in CIM implementation limits their ability to support various computational workloads in edge AI. In this article, we propose a logic-compatible embedded-flash macro with two-step time-domain computation. The proposed macro computes through dual capacitors operating in a ping-pong manner, enabling simultaneous activation of up to 128 cells with signed weights and achieving 10-bit output precision with the option of further scalable resolution. A pipelined time-to-digital conversion architecture further reduces conversion latency by performing computation in parallel, and an on-chip rectified linear unit (ReLU) suppresses unnecessary time-to-digital conversions for negative outputs, further improving energy efficiency. The proposed CIM prototype achieves a peak throughput of 639.38 GOPS and peak normalized energy efficiency of 1517.16 TOPS/W using 3-bit weights, 4-bit activations, and 10-bit outputs, with a measured matrix-vector multiplication (MVM) error deviation of less than 1%.

Ji-Ho Chun +50IEEE Journal of Solid-State Circuits
Scholarai-hardware

A 65-nm Compute-in-Read-Only-Memory Macro With Bidirectional Ternary Weight Storage and Sparsity-Aware Local Accumulation for 1.58-bit LLM Inference

Compute-in-Read-Only-Memory (CiROM) is a promising paradigm for weight reload-free edge inference. The recent emergence of BitNet, which represents weights in ternary values at an average precision of approximately 1.58 bits, opens a path toward single-chip LLM deployment by drastically reducing storage footprint. However, fully exploiting this opportunity requires addressing three key challenges: compact ternary weight storage, effective exploitation of high weight sparsity, and post-fabrication adaptability for downstream tasks. To this end, this paper presents a BitNet-oriented CiROM macro with three dedicated techniques. First, a Bidirectional ROM Array based on Configurable Source/Bit-Lines stores two ternary weights in each 1T ROM cell, significantly improving the weight storage density. Second, a Tri-Mode Local Accumulator is developed to support an accumulation-then-summation dataflow, where zero-weight operations are locally gated to translate BitNet sparsity into reduced switching activity and energy consumption. Third, a lightweight digital adapter co-designed with LoRA-based fine-tuning is introduced to provide post-fabrication downstream task adaptability with negligible hardware overhead. The proposed macro is fabricated in a 65nm CMOS process and achieves an ultra-high weight bit density of 4,479 Kb/mm2, corresponding to an $9.20\times $ improvement over prior digital CiROM designs. Silicon measurements demonstrate a peak energy efficiency of 33.7 TOPS/W and a peak compute density of 60.8 GOPS/mm2 over a 0.6-1.2 V supply range. Model-level evaluations further confirm that the proposed digital adapter consistently improves downstream performance across diverse tasks. These results demonstrate that the proposed macro extends CiROM toward efficient weight reload-free LLM acceleration, providing a promising path for compact and energy-efficient edge deployment.

Wen-Lun Zhang +40IEEE Transactions on Circuits and Systems Part 1: Regular Papers
Scholarai-hardware

Low-Cost Testing of 2.4 GHz Receiver on Standard Digital ATE: Practical Implementation

In this article, we present the implementation of a low-cost solution for the test of 2.4 GHz receivers on a standard automatic test equipment (ATE). The approach relies on the use of a customized binary signal generated by a digital tester channel as the test stimulus. This binary signal is specifically defined such that it presents similar spectral characteristics in the frequency band of the device under test (DUT) as a conventional analog test stimulus generated by an RF tester channel. Dedicated scripts are developed that permit: 1) the selection of an appropriate sampling frequency for the digital channel to avoid spectral corruption in the DUT frequency band and 2) the automatic conversion of the I/Q data computed by the ATE in case of a conventional RF stimulus into a binary sequence to be stored in the vector memory of the digital tester channel. Experimental results are provided that validate the efficiency of the proposed solution under different communication protocols [ZigBee (ZB) and Bluetooth low energy (BLE)] and signal strengths.

T. Vayssade +40IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems