evox
Infrastructure for evolutionary AI systems. GPU-accelerated, PyTorch-based optimization with an ecosystem of algorithm libraries, benchmarks, and tools.
6 repos · 55 papers · 53 scholar
Showing 114 of 114 items
Infrastructure for evolutionary AI systems. GPU-accelerated, PyTorch-based optimization with an ecosystem of algorithm libraries, benchmarks, and tools.
Practical guide to Spiking Neural Networks and Neuromorphic Computing
A Neural Processing Unit designed to run Neural Network Inferences and implemented using a Digital Compute In Memory architecture
Photonic computing challenge for hardhaq 2026
A small MoE language model built from scratch in numpy, used to study memory and compute for inference: a GPU memory simulator, grouped expert dispatch, mixed-precision quantization, and an expert cache.
Matrix multiplication dominates both PHY signal processing and Deep Neural Network (DNN) inference, yet these domains place opposing demands on the matrix engine. Signal processing requires low-latency execution of relatively small matrix operations, whereas DNN inference, particularly at large batch sizes, prioritizes high throughput and energy efficiency for large matrix multiplications. While offering efficient memory access profile, a Systolic Array (SA) adds fill and drain delays per matrix. Those can be amortized for large matrices but not for small matrices. Conversely, a tensor unit completes the kernel in few cycles but would require repeated reads of same operand tiles over large matrices. This makes SA too slow for signal processing and tensor too inefficient for AI. We show that this need not be a choice between systolic and tensor structuring of Multiply-Accumulate (MAC) units. We propose a single engine that operates primarily as a SA and, through low-cost reconfiguration, transforms into a tensor-style engine on a per-operation basis. MUXes create both the spatial reduction adder tree and temporal reduction accumulators from the existing MACs and registers within the systolic array. The proposed reconfigurable SA simulated using modified SCALESim simulator, verified on an Zynq 702 FPGA and synthesized using ASAP 7nm FinFET library. SCALE-Sim simulations shows up to 7.5x lower latency for small complex value PHY kernels. Even on DNN workloads, it achieves up to 1.65x speedups over vanilla SA execution with trivial increase of memory accesses, retaining the systolic array's memory access advantage. Synthesis results estimate reconfigurable fabric costs only 8% power and 13.6% area overheads.
Fully homomorphic encryption (FHE) enables privacy-preserving computation on encrypted data without decryption. Despite its strong security guarantees, FHE incurs substantial computational and memory overhead. In FHE-based deep learning inference, on average 63% of the CPU time is spent on the number theoretic transform (NTT) and its inverse (INTT), which reduce the complexity of polynomial multiplication from $O(N^2)$ to $O(N \log N)$. Prior FPGA works follow two approaches: accelerating entire FHE schemes, which yields significant speedups but ties them to specific parameters and algorithms, or accelerating only the NTT, which applies across schemes, algorithms, and libraries. Most FPGA NTT designs require interconnects that grow with parallelism and utilize significant resources, which complicates routing on multi-die FPGAs. Additionally, to the best of our knowledge, end-to-end NTT acceleration within an FHE context has so far been left unexplored. We design a fully unrolled, delay-commutator streaming NTT/INTT kernel, and floorplan it across the three dies of an AMD Alveo V80. At 400 MHz, the kernel transforms a polynomial of $2^{16}$ coefficients every 5.12 $μ$s, the highest throughput among FPGA designs at this size. Offloaded end to end, it is up to 22x faster than the CPU per transform. We integrate the kernel into the Lattigo FHE library and evaluate it end to end. With 128-bit secure parameters, it accelerates core FHE operations, such as key switching and bootstrapping, by 1.8-2.2x, and neural network inference by 1.8-2.1x.
In FPGA soft processors with shared backing block RAM, memory admission can stall a multiplication after its product has been prepared. We study how admission timing changes result retention in PaceRV, our board-tested five-stage RV32IMC core, using fixed binaries and matched dynamic instructions. Offering data requests one edge earlier reduces geometric-mean Embench cycles by 18.2% and retention opportunities from 375,063 to 4,373. Admission and retention cycle gains are conditional and non-additive. Under slower reads, a 99% capture target selects an 18-edge validity window with registered requests and 28 edges with early requests, although no multiplication matched within both measurement intervals waits longer. A predeclared sizing rule predicts opportunity recovery exactly in all finite tests, including kernels excluded from model development. Independent timing analysis closely estimates execution-cycle savings. The results support choosing admission first, then deciding whether to extend validity and sizing it against an explicit capture or missed-opportunity objective.
Read mapping algorithms represent a critical bottleneck in genomic analysis pipelines due to the resources required; a problem that is becoming more pronounced as sequencing technologies advance and datasets expand. We present an extensive workload characterization of leading read mapping algorithms across diverse sequencing technologies to identify hardware acceleration and optimization opportunities. We analyzed the following implementations: MashMap, wfmash, and Minimap2 for whole-genome alignment (WGA); NGMLR, Minimap2, and LRA for long reads; and Bowtie2, BWA-MEM, StrobeAlign, and minibwa for short reads. We used the Intel VTune Profiler for detailed memory analysis to identify memory- bound functions, cache usage patterns, and data transfer bottlenecks. We further analyzed a selected subset of alignment tools using the DAMOV-SIM simulator to obtain a more granular view of data movement characteristics for two different cache structures. Based on our results, Minimap2 emerges as the fastest alignment tool for long read scenarios, but falls behind in whole-genome alignment due to poor multithreading. Across all three categories, some programs (e.g., LRA, BWA-MEM) are highly memory-bound. More in-depth analysis via DAMOV-SIM revealed that some memory-bound tools, in particular, do not fully utilize the multi-level cache hierarchy in modern multi-core CPU architectures. Our results show that while some tools require algorithmic redesign for performance gains, some others can benefit from running on PIM architectures.
Neural network inference at the edge must fit tight energy budgets. Existing benchmarks compare few devices, do not report energy on a common basis, and omit the newest microcontrollers with an NPU or the 1 GHz Cortex-M85. Hardware-aware design also ranks networks by MAC count, yet no study has tested whether this proxy holds across accelerators. We benchmark sub-watt edge AI hardware by evaluating four CNNs (MobileNetV3, MCUNetV1, EfficientNet, SqueezeNet) and two tiny transformers (MobileViT-XXS, DeiT-tiny) on ten commercial devices: an in-sensor DSP (Sony IMX500), four MCUs with an NPU (STM32N6, Alif Ensemble E7/E8, GreenWaves GAP9), and five MCUs without an NPU (STM32U5, Apollo510B, ESP32-S3 Xtensa LX7, RA8M2, STM32V8). Four devices are novel (Alif E8, Apollo510B, RA8M2, STM32V8). Measured as energy above idle, the IMX500 (0.13-0.35 mJ) uses 1.3-3x less energy than the STM32N6 (0.31-1.05 mJ) on every model; the GAP9 uses the least on MobileNetV3 and the Alif E7 the least on MCUNetV1 and SqueezeNet. All four use 6-34x less energy than the STM32V8 (1.85-3.98 mJ), and the Alif E7 and IMX500 form the power-performance Pareto front. Latency on all four accelerators scales with memory traffic (bytes of weights and activations, r >= 0.98) and not with MACs, so they are memory-bound, while no MCU without an NPU is. The NPU advantage grows with intensity from 1.4x over a 1 GHz Cortex-M85 on MCUNetV1 to 8.2x on SqueezeNet. MAC count is therefore a misleading cost proxy for sub-watt accelerators.
Face anti-spoofing (FAS) models often learn attack- and dataset-specific cues that degrade performance under unseen acquisition conditions and presentation attacks. We present DINOFAS, a DINOv3-based framework that formulates FAS as bonafide-centric representation learning. The proposed framework combines parameter-efficient DINOv3 adaptation using low-rank adaptation (LoRA), an attention style-content disentanglement (ASCD) module for reducing domain-specific appearance variations, and a dual-branch architecture that separates prototype-based liveness learning from attack-semantic supervision. The model is jointly optimized using a bonafide-centric prototype objective, an attack-semantic classification loss applied to detached features, and orthogonality regularization. DINOFAS achieves an average HTER of 9.09\% under the standard OCIM leave-one-domain-out protocol, where training is performed on three source datasets and evaluation is conducted on the held-out unseen domain. We further evaluate synthetic-to-real generalization by training exclusively on SynthASpoof and testing on four unseen real-world datasets, where DINOFAS achieves the lowest average HTER of 21.41\% and the highest average AUC of 86.14\% among the compared methods. Evaluations on OULU-NPU and SiW-Mv2 additionally demonstrate robustness to variations in acquisition conditions, devices, and unseen presentation attacks.
Timely and accurate crop health diagnosis is essential for sustainable agricultural production, yet accurate diagnosis often requires expert knowledge that is not readily available. Recent vision-language models (VLMs) have demonstrated remarkable capabilities in visual understanding, reasoning, and natural-language interaction, creating new opportunities for intelligent crop health assessment. However, deploying VLMs on resource-constrained devices remains challenging due to their substantial computational and memory requirements. This paper presents PlantVLM-Edge, a quantized vision-language framework for conversational crop health diagnosis that can run entirely on resource-constrained edge devices. PlantVLM-Edge combines agricultural domain adaptation, post-training quantization, and on-device serving with an offline chat interface, allowing growers to ask open-ended questions and receive diagnoses, symptom explanations, and management guidance without cloud connectivity. The framework is backbone-agnostic: we validate it with two different VLM types with varying scale and architecture (Qwen3-VL-2B and LLaVA-1.5-7B), four quantization schemes, and two diagnostic domains (plant diseases and insect pests), and we deploy all variants on an NVIDIA Jetson AGX Orin. Eight-bit quantization affects answer quality minimally relative to the full-precision model, and 4-bit activation-aware weight quantization (AWQ) retains 98.6% of full-precision answer quality on both domains while reducing the weight memory of the 2B backbone by 49%. In its most compact form, PlantVLM-Edge requires 2.01 GiB of weight memory, returns its first token in about 0.12 s, and completes a 64-token diagnostic response in 0.78 s. These results demonstrate that timely and accurate crop health diagnosis is practical on edge devices, providing a foundation for next-generation crop monitoring and decision support systems.
Skin cancer diagnosis, particularly differentiating melanoma from benign nevi, is challenging due to the visual similarity between lesions. Although convolutional neural networks (CNNs) and vision transformers (ViTs) have shown promising performance, their effectiveness often deteriorates on data from heterogeneous clinical sources. Conventional domain adaptation methods address domain shift but require access to source data during adaptation, which is often infeasible due to privacy regulations. Multi-source-free unsupervised domain adaptation (MSFDA) addresses this limitation by leveraging multiple labeled source domains to generalize to an unlabeled target domain without requiring access to source data. However, existing MSFDA methods rely on full backbone fine-tuning, leading to catastrophic forgetting and overfitting on small clinical datasets, and address domain shift at the aggregation stage without establishing a shared domain-invariant feature space. Furthermore, their reliance on hard pseudo-labels or confidence-weighted aggregation introduces noisy supervision signals under domain shift. To address these limitations, we propose CAT-CKD, consisting of two components: (1) contrastive adapter training (CAT), which trains lightweight ConvPass adapters within a frozen ViT backbone using supervised contrastive learning (SCL) to establish a shared domain-invariant feature space before source-specific model training, and (2) consensus knowledge distillation (CKD), which aggregates logits from multiple source models into a consensus supervisory signal and adapts a student model on unlabeled target data using KL divergence. Experiments on five publicly available skin lesion datasets show that CAT-CKD achieves an average AUROC of 86.1%, outperforming existing MSFDA methods while requiring only 4.3M trainable parameters. The code is available at https://github.com/A-Abedi/CAT_CKD.
Optimal control of partially observable Markov decision processes (POMDPs), in general, requires a complete history. The belief state, a distribution over hidden states, summarizes this history and provides the classical sufficient statistic to control POMDPs. However, the continuous nature of the belief space generally prevents a finite-memory controller from maintaining it exactly. This raises the question we address in this work: what must the available finite memory preserve to achieve near-optimal control performance? We study this question through an information-theoretic converse that gives a fundamental lower bound on the loss of any finite-memory controller. The analysis introduces the concept of a decision witness that characterizes belief states whose confusion would lead to unavoidable loss. Motivated by this bound, we propose RECAP, a constructive quantization method for designing representative memory states. We prove an upper bound on the optimality gap of any RECAP codebook and use the components of this bound together with the converse analysis to define the codebook objective. Experiments on 11 POMDP benchmarks demonstrate that RECAP improves over finite-memory representation baselines under the same memory budget.
Attention remains a substantial computational cost in Transformer training, particularly at long context lengths. NVIDIA Blackwell GPUs provide high-throughput low-precision matrix multiplication, but exploiting it in attention requires addressing quantization overhead, causal information leakage through shared scales, and numerical inconsistencies between the forward and backward passes. We present FLAMIA, a family of attention kernels for Blackwell with MXFP8 and mixed NVFP4/FP8 training paths. Our design combines token-block scaling that reuses quantized operands across matrix orientations, targeted safeguards against scale-induced leakage, and consistent forward and backward computations. Fused quantization and architecture-specific kernel optimizations reduce overhead, while document-local scaling and workload-aware scheduling support variable-length documents directly in packed layouts without repacking. Across various causal packed-document workloads, the MXFP8 path accelerates the combined forward and backward attention call, including quantization, by factors of $1.10$--$1.40\times$ on B200 and $1.06$--$1.42\times$ on B300 relative to the fastest evaluated baseline for each workload. We validate training stability and quality preservation on pretraining Llama-style models. These results show how jointly designing quantization, training numerics, and document scheduling translates low-bit hardware throughput into practical attention acceleration.
Large Mixture-of-Experts (MoE) models are promising for local inference because sparse routing allows the routed-expert pool to reside on SSD while DRAM caches recently used experts. However, under practical DRAM cache capacities, substantial SSD I/O remains on the critical path despite I/O-computation overlap. Token-level dynamic expert precision offers a natural way to reduce this data demand by allocating bits according to expert importance and quantization sensitivity. Nevertheless, treating each bit-width of the same expert as an independent object prevents cross-precision cache reuse, causing overlapping data to be fetched repeatedly. The key insight is that precision should instead denote the refinement depth of one shared expert state. Based on this insight, we present PagedMoE, which stores each expert once using a prefix-compatible tree codebook and bitplane layout. Lower-precision representations are valid prefixes of higher-precision ones, allowing cached data to be reused or extended by fetching only missing bitplanes. PagedMoE selects the required prefix length before I/O using current routing weights and offline expert-specific error profiles, and directly executes cached 1-8-bit prefixes without materializing full-precision experts. At an average routed-expert precision of 2.5 bits, PagedMoE achieves quality close to GPTQ-INT4 on Qwen3.5-397B-A17B and to MXFP4 on most evaluated GLM5.2 tasks. With a 24 GiB expert cache, it reaches 8.89 tokens/s on Qwen3.5, 1.33x the throughput of the same dynamic requests using independent precision objects and 1.56x that of uniform 4-bit execution.
Generative recommendation has been widely adopted in large-scale industrial recommendation systems, where multimodal item embeddings are quantized into offline semantic identifiers (SIDs) and items are retrieved by autoregressively generating their SIDs. However, conventional SID construction often underrepresents item utility reflected in post-exposure behaviors such as shares, long views, and video completion, making semantically similar items with different utility insufficiently distinguishable while struggling to accommodate rapidly changing utility signals. Moreover, generation likelihood may be misaligned with downstream utility, causing high-utility candidates to be prematurely pruned during beam search before post-generation reranking or training-time reward alignment can effectively compensate. To address these limitations, we propose UAGR, which aligns both SID construction and decoding with downstream multi-objective utility. Our Utility-Enriched Semantic Codes preserve the original semantic information while incorporating item-level utility without increasing SID length or vocabulary size, supporting incremental utility updates with no additional tokenization or sequence-decoding cost. Meanwhile, our Utility-Guided Decoding estimates candidate utility by combining predictions of multiple post-exposure outcomes and incorporates these estimates into beam scoring, reducing the risk of prematurely pruning high-utility paths while enabling flexible utility control without interfering with generative backbone training. Extensive offline experiments and online A/B tests under real-world traffic demonstrate the effectiveness and deployability of UAGR, which has been fully deployed on Kuaishou and increases video watch time by 0.206\% online.
Block-scaled lookup-table (LUT) formats, which represent weights as indices into a small value set (the grid), have become standard for low-bit LLM quantization, ranging from NF4 and BOF4 to the 3-bit LUT instruction in NVIDIA Rubin. While these formats leave the grid definition flexible, existing implementations rely on fixed grids derived either from numerical data types (INT, FP) or assumed weight distributions (NF, BOF). We demonstrate that no single fixed grid is universally optimal; the best fixed grid varies with bit-width: BOFS at 2 bits, NF at 4 and 5 bits, and INT at 6 bits. We introduce SHq, a closed-form family of grids parameterized by two critical degrees of freedom: a companding strength α that interpolates level spacing from uniform to floating-point-like, and an offset c that introduces asymmetry. This offset allows the sign bit of the FP8 block scale, otherwise unused, to select between the grid and its mirror image, up to doubling the number of distinct values available per block. SHq modifies only the lookup table contents and serves as a drop-in replacement for the grid in any such format. Across five models ranging from 3B to 35B parameters, including a mixture-of-experts architecture, evaluated at 2-6 bits using RTN and GPTQ, SHq achieves the lowest KL divergence in every setting at 3 and 4 bits and in most settings at 2, 5 and 6 bits. It outperforms the best fixed grid by up to 34% and remains superior under mixed-precision configurations. At 4.5 bits per weight, the SHq format reduces the KL divergence of NVFP4 by 41-67% while improving downstream accuracy. Grid families incorporating both spacing and offset parameters perform comparably, whereas those lacking an offset exhibit clearly higher error, confirming that these two degrees of freedom drive the observed gains. We therefore propose SHq as the standard grid for block-scaled lookup-table formats.
This paper introduces a reinforcement learning-based approximate dynamic programming approach to optimize the deployment of Walker-type satellite constellations under uncertainties, including launch failures and satellite malfunctions. We formulate the problem as a Markov Decision Process that incorporates low-injection orbit launches, continuous low-thrust orbit raising, and variable satellite loads. We adopt the offline approximate value iteration solution method, using linear function approximation with basis functions and action pruning strategies to substantially reduce the computational and memory requirements for finding optimal decision policies. Preliminary results demonstrate convergence to policies with low expected costs and deployment times, both of which are lower than those of a rolling-horizon baseline, showcasing a novel application of approximate dynamic programming in space logistics.
On-device multimodal agents are limited by two deployment mismatches: pretrained attention structures may not divide evenly across the compute cores of an edge accelerator, causing irregular partitions; and standard prefix caching keeps historical visual tokens in the persistent KV cache, whose capacity grows with the number of rounds. We present PARC, which combines parallelism-aligned structured pruning with lifetime-aware KV-cache reorganization. At the model level, it uses first-order Taylor importance and grouped-query-attention-aware selection to make attention dimensions divisible by the target core count while preserving the query-to-key/value ratio and head dimension. At runtime, it separates the system prompt, persistent text history, and transient visual observations by lifetime, updating text history during tool execution so that historical images leave the persistent cache and more interaction rounds fit a fixed budget. On 2,062 GUI actions, the incremental interface shows no accuracy regression, and task-aware recovery training restores 98.7% of the action-type accuracy gap and the coordinate hit rate to the original level. Under seven-core deployment, hardware-aligned pruning reduces vision-encoder latency by 11.4%, improves decoding throughput by 23.8%, and lowers per-heap high-water memory by 12.7%; a parameter-matched control confirms that the speedup stems mainly from core alignment rather than parameter reduction.
Reasoning-augmented recommender systems (RecSys) enhance user preference modeling by generating reasoning trajectories over historical interactions before making recommendations. Knowledge distillation provides a practical paradigm for transferring such reasoning capabilities from powerful large language models to lightweight RecSys. However, existing methods typically rely on a single fixed teacher, remain constrained by the teacher knowledge ceiling, and indiscriminately distill teacher-generated trajectories without assessing their actual utility for downstream recommendation. To address these limitations, we propose MUSE (Multi-Teacher Utility-guided Selection with Self-Exploration), an adaptive reasoning distillation framework that jointly improves reasoning generation and reasoning utilization. MUSE first constructs a diverse reasoning pool using multiple teacher LLMs and progressively augments it via student self-exploration, enabling the student to leverage the strengths of different teachers and gradually explore reasoning strategies beyond teacher imitation. We then develop Rank-Gain Utility Assessment (RGUA), which evaluates each candidate according to the normalized full-catalog rank improvement of the ground-truth item and identifies the most and least recommendation-beneficial trajectories. Based on these utility-aware trajectory pairs, we introduce dual-level contrastive learning objectives that guide the student generator to prefer beneficial reasoning while training the recommendation encoder to more effectively exploit recommendation-relevant preference evidence. Extensive experiments on three real-world datasets demonstrate the effectiveness and robustness of the proposed framework.
Geometry-aware neural operators face substantial computational and memory demands when learning across complex physical domains. We introduce Architecture-Aware Mixed Precision (AA-MP), a replica-exchange training strategy that allocates numerical precision according to replica roles and architecture-specific operations. By retaining FP32 for refinement while selectively employing BF16 for exploration, AA-MP preserves predictive accuracy while improving training efficiency. Across fluid dynamics and three-dimensional structural mechanics benchmarks, AA-MP achieves comparable accuracy to full-precision training over an ensemble of runs, while reducing hot-replica peak GPU memory by 46.9% and training time by up to 25.8%. It also outperforms naive Automatic Mixed Precision (AMP) in memory efficiency and computational cost, with increasing memory advantages at larger model and geometry scales. Ablation studies show that retaining FP32 in the cold-replica is important for predictive accuracy, while additional FP32 operations in the exploratory replica provide limited benefits relative to their computational cost. Finally, conformal prediction for uncertainty calibration shows that AA-MP achieves near-nominal coverage with narrower prediction intervals than the alternatives. These results advance the scalability of geometry-aware scientific machine learning by reducing computational costs without compromising predictive accuracy or uncertainty quantification.
LLM agents for GPU kernel optimization sit at two ends of a structure-freedom tradeoff. Scripted agents fix what the model does at each step, via a learned policy, an optimization menu, or fixed roles, and are sample-efficient where the recipe applies and stall where it does not. Open-ended agents let a self-directed model mutate a population of kernels and pay for the freedom in trials: thousands of evaluations, repeated edits, and wrong candidates. Both respond to hard targets by equipping the agent with more: parsed profiler output, disassembly analysis, retrieval over documentation and libraries, and stores of lessons. We argue that both lack direction, a hypothesis about where the remaining gap is, and that a short question asked at the right moment supplies it without narrowing what the agent may do next. We present KernelSeek, a question-driven agent that makes an existing CUDA kernel faster with no operator-specific knowledge and no faster implementation to imitate: a general-purpose coding agent, a minimal harness that enforces a verifiable goal, and questions at decision points, first from a human, then from a bank the agent applies to itself. On the same task, evaluator, hardware, and week, a tool-rich autonomous agent built in the prevailing style plateaued at 0.50x FlashAttention-3 after 133 model-driven steps; KernelSeek reached 1.13x. It then produced dependency-free FlashAttention forward kernels for Hopper and Blackwell, single CUDA files with no CUTLASS, CuTe, cuBLAS, or cuDNN, correct on all 182 scenarios per architecture and 1.16x FlashAttention-3 and 1.46x FlashAttention-4 (geometric mean). Across 162 human messages and 11.5K agent tool calls, every large gain followed a question that moved the agent to a different layer of the system or rejected a claimed limit, never one that supplied information. We release the kernels, harness, question bank, and all traces.
Machine learning supports materials design, atomistic simulation and experimental characterization, yet models and workflows remain fragmented. We present PaddleMaterials, an open-source framework integrating heterogeneous materials-learning tasks while preserving task-specific representations and scientific semantics. It unifies data processing, training, evaluation, pretrained-model management and workflows across six task classes: property prediction, structure generation, machine-learning interatomic potentials, electronic-structure fields, spectrum elucidation and microscopy image processing. Task interfaces record physical quantities, units, aggregation conventions and references, making cross-task composition verifiable. A shared compiler backend enables switching from dynamic-graph to compiled execution through one setting, using graph optimization, operator fusion and hardware-targeted code generation without changing model definitions. Across 65 model instances in 11 groups, compiled execution accelerated every instance in eight groups, with group median speedups of 11%-224% and group maximum speedups of 11%-336%. Case studies on LiMnO2 and toluene link structure generation or spectrum elucidation, interatomic-potential relaxation, property prediction and electron-density prediction, using same-geometry density functional theory (DFT) calculations as references. For LiMnO2, predicted formation energies for generated and relaxed structures differed from same-geometry PBE+U references by at most 7 meV per atom, while the computational cost of model steps was less than one-twentieth of a DFT single-point calculation. PaddleMaterials extends unified materials workflows beyond conventional property, potential and structure-generation tasks while supporting verifiable cross-task operations and compiled execution.
LLM agents are rapidly shifting software development toward intent-driven \emph{vibe coding}, yet their effectiveness drops sharply in high-level synthesis (HLS), where code optimization must jointly consider design structure, directive search, and downstream hardware constraints. Existing HLS optimization approaches, including both heuristic-based methods and recent LLM-assisted pipelines, typically operate at fixed abstraction levels or in weakly coupled stages, limiting their ability to leverage the strong interplay between source code restructuring and directive search outcomes. Our key insight is that HLS optimization can be reframed from a fixed, hardwired transformation pipeline into a skill-based agentic control problem: an LLM agent equipped with modular, executable optimization skills and a feedback-driven orchestrator can autonomously navigate the coupled space of code restructuring and directive search. This motivates the design of HLSClaw, the first HLS-specific, skill-based agentic harness for holistic HLS optimization. On real-world HLS benchmarks, HLSClaw achieves a 39.12x geometric-mean latency speedup, outperforming the compiler baseline ScaleHLS by 1.98x, and surpassing the commercial coding agents Claude Code and Codex by 1.86x and 1.70x, respectively, while maintaining strong timing robustness.
Approximate nearest-neighbor search on GPUs serves retrieval-augmented generation, recommendation, and deduplication over datasets of up to billions of vectors. The fastest GPU indexes at high recall, such as CAGRA, walk proximity graphs in chained rounds, each of which waits for the distances of the previous one. High recall therefore costs many dependent rounds per query. Our insight is to spend the HBM of the GPUs on precomputed neighbor lists, so that a query reads its candidates in one parallel step instead of a chain of rounds. This paper turns the above insight into a GPU vector search system, namely OctEx, which holds hundreds of randomly projected copies of the dataset in GPU memory. Every copy rounds its vectors to the nearest points of a regular grid whose cells are nearly round and stores for every populated cell the list of vectors nearest to its center. A query reads the lists of its cells in all copies at once and then, in a fixed number of waves, the precomputed neighbor lists of its best hits. OctEx ranks the union of all hits by an int8 sketch before one exact rerank. OctEx runs the whole search of a batch as CUDA graphs over all GPUs of a server and places every part of its index in HBM or host memory by its access frequency. On four H100 GPUs at recall 0.95 to 0.995, OctEx outperforms CAGRA by 2.0x to 46.4x on seven datasets of up to one hundred million vectors, and by 1.46x to 2.05x on the one billion vectors of Deep1B and SIFT1B, whose indexes it builds in 44 to 47 minutes. OctEx also outperforms GGNN, PathWeaver, FlowANN, and the cuVS inverted file indexes IVF-PQ and IVF-RaBitQ by 5.4x to 343x.
Local music applications estimate key during library import and benefit from an analyzer that is small, fast, and runs in-process. We start from a compact chroma-template key analyzer that fits this budget but lags behind learned references. Learned systems reach higher accuracy, but their installed model and runtime dependencies can occupy tens to hundreds of megabytes. We present key lime, a 21.5k-parameter DSP-guided transformer reranker with a 115 KB model artifact and a 76.7 MB installed footprint, including the extracted ONNX Runtime. It adds 0.32 ms per track to deterministic DSP scoring. Trained from labels on 4,040 FMAK tracks, with 449 tracks for validation and 1,000 disjoint FMAKv2 tracks held out for evaluation, it improves MIREX from 61.08 to 73.55 +/- 0.46 and exact accuracy from 52.30 to 67.07 +/- 0.65 over three seeds (mean +/- population SD). It approaches a same-manifest local madmom KeyCNN reference at 74.16 MIREX. A 12-parameter circular input stage matches the accuracy range of a 468-parameter linear input projection.
Continuous estimation of finger-joint angles from surface electromyography (sEMG) is a promising approach for controlling dexterous myoelectric prostheses. Deep time series models have been applied successfully in this context, typically relying on feature extraction rather than raw signals, given limited subject-specific data. Various time-domain and frequency-domain features can be used, extracted for a given window and step size. This study systematically investigates these three design choices for Long Short-Term Memory (LSTM) regression on the Ninapro DB2 dataset. Window sizes from 300 to 1200 ms and step sizes from 10 to 100 ms were evaluated in combination with six time- and frequency-domain features. Models were evaluated using regression performance metrics, inference time, and memory footprint to assess feasibility on embedded microcontroller hardware suitable for a prosthetic device. A 300 ms window with a 100 ms step, paired with either the Mean Absolute Value (MAV) or Root Mean Squared (RMS) feature, yielded the best performance (test PCC of 0.779 +/- 0.080 / 0.779 +/- 0.085). Window size, step size, and feature choice were found to interact non-trivially: jointly tuning them can improve both accuracy and computational feasibility for real-time, embedded myoelectric control.
This paper presents a scalable distributed acoustic surveillance system for real-time gunshot detection and localiza tion in large-area, GNSS-challenged environments. The system consists of multiple low-cost MEMS microphone array units (MAUs) deployed as distributed anchors. Each MAU performs edge-level sound-event detection using a lightweight Sound De tection and Echo Suppression (SDES) algorithm, followed by a TinyML-based classifier for gunshot identification. For detected gunshots, an enhanced Generalized Cross-Correlation (eGCC) al gorithm estimates intra-array time differences of arrival (TDoAs) with improved robustness under low signal-to-noise ratio and re verberant conditions. Angle-of-arrival (AoA) estimates obtained from individual MAUs are fused through a modified SCO-EV localization framework to determine the gunshot position in a global coordinate system. A misplacement calibration (MC) algorithm compensates for platform orientation errors, improv ing angular consistency and localization accuracy in practical field conditions. Extensive experiments conducted over coverage areas ranging from 4,985 m2 to 49,063 m2 demonstrate reliable operation across varying spatial scales. The proposed framework achieves a Matthews Correlation Coefficient (MCC) of 0.70 for gunshot recognition and average localization errors below 6.5 m over 25,000 m2 , while maintaining errors under 15 m at larger scales. The results demonstrate the practical potential of the proposed system for distributed gunshot surveillance in GNSS challenged large areas.
Generative AI is billed per token, yet tokens are not equally expensive. Each token enters the KV-cache and raises the cost of those that follow, so energy grows non-linearly along a sequence whose tokens look identical. No analytical model states what a token costs, and the most accurate SOTA model is a fit, not a law: it absorbs architecture into fitted coefficients, requiring per-LLM fitting. We derive closed-form expressions of the computational and memory-access complexity of LLM inference in terms of architectural hyperparameters, covering MHA and GQA, and compare a family of analytical models of increasing granularity. From these, we obtain the SIMPLE Inference Model to Predict LLMs Energy (SIMPLE), a four-parameter law that separates prefill from decode, and decode into KV-cache cost and next-token computation. Over traces of 13 dense LLMs from six families on NVIDIA H100 GPUs, it attains R2 = 0.969 and 11.97% MAPE, 3.1-3.5x below the SOTA, without per-LLM calibration. The SIMPLE law explains as well as predicts: prefill energy grows quadratically with the hidden dimension, so larger LLMs suffer most from long inputs; and more aggressive GQA KV grouping reduces the per-token energy by up to 3x relative to standard MHA. Increasing the reasoning token budget pushes output length beyond the analytically derived energy minimum, driving super-linear energy growth; although accuracy grows with the budget, energy grows faster, so Intelligence per Joule peaks at an intermediate budget and falls by up to 3.14x across a 4.0x reasoning-budget range. Agentic serving raises energy per request by 13.0x relative to non-agentic ones, driven by the growth of both the prefill and the KV-cache terms with input length; while prefix caching removes most of the prefill cost (-93%), the KV-cache term is unaffected, so even a cached agentic request costs 4.8x more.
Training models over data distributed across edge devices presents a trade-off: centralizing data simplifies training but raises privacy and regulatory concerns, while full local training may exceed the devices' compute and memory capacity. Split learning addresses this challenge by partitioning the model between clients and the server, allowing part of the computation to be offloaded. However, existing approaches typically rely on static split configurations or simplified assumptions that fail to capture the heterogeneity and complex dynamics of real deployments. We present DynSL, a framework that dynamically allocates model layers between clients and the server by continuously monitoring system metrics and periodically solving an optimization problem that accounts for client and server computation and communication. When beneficial, the system reconfigures the split, enabling the offloading strategy to adapt to changing device capabilities and system conditions. Extensive experiments on a testbed of 24 heterogeneous edge devices show that DynSL reduces training time by up to $\sim\!$30% over the strongest adaptive baseline and by more than 50% over a suboptimal uniform split, with no loss -- and sometimes a gain -- in accuracy.
Kinetic molecular memristors offer a compelling path toward high-density analog in-memory computing by achieving 14-bit precision, wherein discrete redox transitions and counterion displacements dictate non-volatile conductance across a deterministic energy landscape. However, integrating these nanoscale kinetics into macroscale circuit simulators has remained bottlenecked by derivative discontinuities and solver convergence failures. Here, we present a physics-inspired continuous kinetic model that bridges microscopic molecular transport with scalable circuit architectures. By casting discrete intramolecular states, counterion relaxation dynamics, and Marcus-Hush-Chidsey electron transfer rates into continuously differentiable state variables, we derive a robust, Spectre-compatible compact model and a matching vectorized Python solver. This work culminates in a unified computational platform to reliably predict the behavior of complex circuits built on these molecular memristors.
Modern FPGA SoCs are no longer just collections of accelerators on top of a programmable fabric. They combine programmable logic with hardened processors, NoCs, and components that dominate end-to-end performance. Evaluating individual pieces is insufficient; system performance depends on the interplay of SoC components. However, current FPGA SoC simulations provide limited support for this style of evaluation. Vendor simulations capture these interactions but are prohibitively slow for iterative design-space exploration. At the same time, generic software models overly abstract the hardware components that determine real behavior, resulting in less accurate simulations. Fast, easy-to-use simulators that accurately model real SoCs are critical for system debugging and design space exploration. We present Naviq, an open-source, cycle-level software simulator for FPGA SoCs, tuned to AMD Versal-class architectures. Naviq bridges accurate but slow vendor simulation and fast software modeling: it captures key behaviors of modern commercial FPGA SoCs while being fast enough for debugging and design-space exploration. Naviq models hardened SoC infrastructure in C++ and integrates accelerators through Verilator, allowing designers to evaluate custom hardware in a full-system environment. At the center of Naviq is an AXI4-compliant NoC model that captures transaction ordering, packetization, routing, backpressure, and memory interactions. Naviq provides CPU and memory integration, protocol checking, checkpointing, tracing, and performance-analysis tools for debugging and design-space exploration. Experimental results show Naviq closely reproduces the bandwidth and latency of Vivado's RTL NoC model while accelerating the evaluated simulations by 81x-182x. We present five case studies demonstrating Naviq's capabilities. Naviq enables practical full-system evaluation of modern FPGA SoC designs.
Spiking Neural Networks (SNNs) provide an event-driven paradigm for energy-efficient edge computing. However, operation-level sparsity in SNNs does not necessarily eliminate the hardware costs associated with redundant output neurons, including neuron-state maintenance, partial-sum (psum) accesses, and on-chip storage requirements. This paper presents SparseForge, a hardware-aware sparse SNN acceleration framework that jointly optimizes neuron pruning, sparse dataflow, and accelerator microarchitecture. At the algorithm level, we introduce a gated leaky integrate-and-fire (G-LIF) neuron to learn differentiable neuron masks with binary spike semantics. A dual regularization scheme is further developed to align pruning with both synaptic-operation reduction and psum buffer capacity reduction. At the dataflow level, we design a sparse Column-Wise Product (CWP) dataflow with a variant of incremental spike encoding to improve weight-access regularity and reduce sparse-index overhead under dynamic spike activity. At the architecture level, we develop a mask-aware sparse processing unit (SPU) that supports compressed spike decoding, sparse implicit im2col conversion, neuron-mask-gated psum updates, and decoupled sparse execution. Experimental results across multiple SNN backbones and datasets show that SparseForge retains 1.75%--77.99% of neurons and 0.16%--62.36% of synaptic operations, with a configurable accuracy--efficiency trade-off. The buffer-capacity-aware regularization reduces the required psum buffer capacity by up to 96.18\%. FPGA implementations on a Xilinx ZCU9EG achieve 296.94--619.49 GOPS/W across the evaluated workloads. Across matched network and dataset settings, SparseForge provides 1.24--4.17 higher FPS and 1.50--7.18 higher GOPS/W than representative FPGA-based SNN accelerators, while reducing EDP by 1.70--19.13, with benchmark-dependent accuracy--efficiency trade-offs.
Delta-rule linear attention underlies language models such as Kimi Linear and Qwen3-Next. Chunkwise implementations expose matrix parallelism, but their backward passes compute unnecessary auxiliary adjoints. We present ReWY, an algebraic reformulation that reduces this work before optimizing accelerator execution. Differentiating the chunk correction as one triangular residual system eliminates the inverse adjoint and reduces seven local WY matrix products to three. Two complementary transformations expose further reuse: centered matrix contractions replace pairwise vector-gate processing, and scalar-gate reductions reuse key-gradient products instead of forming a key Gram matrix. These identities simplify backward computation for DeltaNet, GDN, KDA, and GDN2 while preserving the recurrence and gradients in exact arithmetic, without changing the architecture or training objective. We implement the reduced graph on GPUs and Ascend NPUs. Under matched recomputation and full-saving policies, precision-qualified six-shape comparisons with frozen FLA references yield 1.15-1.81$\times$ geometric-mean complete low-level backward speedups on NVIDIA GPUs. All four operators achieve 1.18-1.70$\times$ on H100 and H200 while retaining original FP32 working matrices. A separate public-autograd evaluation on Ascend 950DT, including gradient conversions, yields 1.09-2.89$\times$ under available matched policies. Paired ablations establish further ReWY gains after optimizing the complementary contractions. Complete-gradient checks validate the recorded operator cases; supplementary full-model backward measurements assess integration into community FLA models and a GDN2 layer port. Code and reproduction materials: https://github.com/flashserve/ReWY.
Sparse Matrix-Dense Matrix Multiplication (SpMM) is a fundamental kernel in graph neural networks and scientific computing, yet its optimization on Ascend NPUs remains largely unexplored. The physically decoupled Cube/Vector architecture and single-threaded scalar instruction dispatch of modern NPUs introduce unique challenges. We present Ascend-SpMM, a novel and systematic framework tailored to the physically decoupled architecture of Ascend NPUs.We introduce a memory-centric cost model that quantifies hardware affinity and guides a conflict-bounded, densityaware task-mapping strategy. The strategy maps dense regions to the high-throughput Cube path and sparse residuals to the flexible Vector path, with a theoretical bound on cross-processor conflicts. To execute these workloads with high efficiency, we further develop highly customized sub-kernels. These kernels leverage hierarchy-aware pipelining and round-robin ring buffer-based implicit synchronization to effectively mask memory latency and reduce instruction pipeline stalls. Extensive experiments demonstrate that Ascend-SpMM achieves a geometric-mean speedup of 12.70-20.55X over the strongest evaluated NPU baseline for each matrix, accelerates end-to-end GNN training by up to 6.8X, and delivers performance competitive with state-of-the-art GPU implementations. The source code of Ascend-SpMM is publicly available at https://github.com/IDMG-Lab/Ascend-SpMM.
Machine learning for remote electrocardiogram monitoring must handle noise and diseases absent from training under tight resource constraints. We investigate whether upstream unsupervised anomaly detection can improve monitoring decisions more efficiently than classifier-derived uncertainty. We evaluate five detectors on PTB-XL and BUT QDB and compare a two-stage filter-classifier system with Monte Carlo Dropout and deep ensembles under unseen-disease and noise conditions. Performance is averaged over five training seeds. We measure single-pass energy, latency, and graph memory on a microcontroller's central processing unit (CPU) and neural processing unit (NPU). Per-score neural network costs account for repeated evaluations but exclude scoring outside deployed graphs. Energy equations relate filtering and rejection to avoided classifier execution. Across four disease scenarios, Deep Support Vector Data Description (Deep SVDD) achieves the highest mean area under the receiver operating characteristic curve within all tested finite CPU energy budgets. On the NPU, it leads at 1 mJ, while Monte Carlo Dropout has slightly higher means at 10 and 100 mJ. With a narrow CPU classifier, the Deep SVDD two-stage system increases the fraction of correctly classified or correctly rejected inputs by 15.0-19.4 percentage points over classification without rejection. Estimated energy per input is 1.93-6.99 mJ, compared with 52.53 mJ for Monte Carlo Dropout and 8.76-17.51 mJ for deep ensembles, at each method's retrospectively selected test accuracy peak. These results support upstream filtering for efficient remote monitoring, while independent validation and complete-system measurements remain necessary.
Attention kernels on Tensor Processing Units (TPUs) often underutilize their matrix units (MXUs) because of repeated matrix operand loads, vector-intensive computation, and dependencies that leave MXUs idle. We present FlyWheel, a set of fast TPU kernels for both softmax and linear attention. FlyWheel reformulates softmax and triangular inversion to reduce vector work, while coordinating data layouts to improve matrix operand reuse. It then overlaps independent computation to hide execution latency. We apply these techniques to both softmax attention and Gated DeltaNet on TPU v6e and v7x. FlyWheel speeds up softmax attention by up to $4.14\times$ over the state-of-the-art RPA kernel and $1.80\times$ over Splash Attention, reaching 81% of peak BF16 throughput on v6e. It also speeds up Gated DeltaNet prefill by up to $3.61\times$. Integrated into vLLM serving Qwen3.8-27B, FlyWheel improves prefill throughput by up to $1.56\times$ while maintaining comparable accuracy.
Private fine-tuning of large language models increasingly pairs DP-SGD with low-rank adapters on frozen, quantized backbones, and every such stack inherits one unstated premise: frozen layers keep each record's clipped gradient a function of that record alone. We show that LLM.int8 breaks this premise. It routes outlier activation columns to a higher-precision path using a union over the whole batch, so one record decides how every other record's activations are computed, even though the weights are frozen and the quantization scales are per row. We bound the resulting sensitivity inflation, state a record-isolation contract that restores the standard bound, and define record-owned precision dispatch that meets it with the same GEMMs as batched LLM.int8, so isolation needs no extra arithmetic. We then audit the native CUDA stack (bitsandbytes, PEFT, Opacus) against four matched controls. A constructed LoRA-Opacus witness reaches $15.0\times$ the add/remove bound. On pretrained OPT-125M with natural text, other records alter 20% of record-layer precision splits at batch size 16. An exploratory re-clipping analysis finds 10 of 64 candidate neighbors above the bound once every gradient is clipped; the top pair, revalidated natively, reaches $1.065C$ while every control stays within $C$, and forcing one layer's outlier mask removes the excess. A two-million-draw audit of the released adapters yields $\varepsilon$ lower bounds ($3.25$ at $δ=10^{-5}$) below the advertised guarantee ($4.38$). Resolving an excess this small at the release level would take about 200 times more draws, whereas checking the sensitivity contract directly exposes it in a few native runs, which makes contract checking the practical audit for private quantized training.
Hybrid quantum-classical workloads, particularly Potential Energy Surface (PES) scans in quantum chemistry using batched VQE and QAOA, involve many structurally similar Hamiltonians differing only in coefficients. Existing warm-start and meta-optimization strategies reuse parameters between instances but still consume fixed quantum-classical resources, ignoring cross-instance structural similarities. We present OptiPESS, a GPU-accelerated runtime system for adaptive resource allocation in hybrid quantum-classical execution. OptiPESS introduces the Locality-Occupancy-Variance (LOV) metric, a lightweight, geometry-aware measure of Hamiltonian similarity, coupled with efficient alignment and Pauli-classification passes. Using LOV, OptiPESS performs distance-aware scheduling of iteration and shot budgets, predicting which instances can be skipped, quick-solved, or fully optimized. Implemented atop CUDA-Q with persistent multi-GPU execution and near-linear scaling, OptiPESS reduces classical and quantum evaluations by up to 78.3\% and wall-time by up to 74.8\% while maintaining baseline accuracy. This demonstrates how lightweight static analysis and adaptive scheduling can transform quantum-classical co-execution from fixed-budget loops into a predictive, resource-aware runtime.
The rapid growth of electric vehicles (EVs) in urban environments demands intelligent, decentralised charging management strategies that operate under strict real-time constraints. This paper proposes an Edge-Intelligent architecture in which each EV selects a charging station (CS) by querying a shared, centrally trained Deep Q-Network (DQN) policy over a lightweight MQTT publish-subscribe backbone; independent per-EV edge inference is the deployment target, not the configuration evaluated here. The reported simulation evaluates that shared DQN policy applied sequentially to each vehicle's observations within a single SUMO run, using MQTT for KPI telemetry while observations are obtained through TraCI. The system is evaluated on two real-world road networks from OpenStreetMap (Rabat Agdal-Hay Riad and Tangier, Morocco); the primary network carries four geolocated public charging stations with heterogeneous power levels (50 kW and 150 kW). Three traffic demand profiles (morning, midday, evening peak) are simulated, and four decision policies are compared: an Adaptive Heuristic, a Tabular Q-Learning agent, the proposed DQN agent, and a Double DQN variant. A multi-seed statistical protocol (7 seeds, Wilcoxon signed-rank test, Cohen's d) evaluates the results. The DQN agent achieves a 6.8% relative improvement in charge rate and a 5.4% increase in delivered energy over the Adaptive Heuristic under Evening Peak demand ($p = 0.016$, Cohen's $d = 0.99$ and $1.51$), while showing no advantage in waiting time; on the second network the heuristic leads on waiting time, so the ranking is not stable across geographies. A scalability analysis shows linear state-space growth with station count and sub-millisecond inference; all runs include intermittent station outages under a Markov failure process.
Heterogeneous embedded boards pair a Linux application processor with a real-time microcontroller (MCU), which suggests moving a task between the two domains when its timing degrades. We present QFabric, a prototype that treats cross-domain placement as guarded recovery from a violated empirical soft real-time contract: a bounded task is compiled for both domains, monitored in windows with hysteresis, switched only after semantic and capacity admission at a quiescent epoch boundary, and committed or rolled back after probation. On an Arduino UNO Q (Cortex-A53 Linux and a Cortex-M33 running Zephyr), we characterize the stock Bridge RPC interconnect and measure five workloads in both domains under idle, CPU-load, and higher-priority SCHED_FIFO contention, with ordinary and real-time callers. Each MCU invocation costs about 19 ms of dispatch regardless of compute, the MCU runs the heaviest kernel 34x slower, and because every call traverses Linux userspace, Linux contention inflates the MCU path far more than the Linux path (p99 up to 400 ms versus 17.0 ms). Across 30 cells of a four-node chain, Linux meets every contract within 0.8% misses and QFabric correctly never migrates; a controller without hysteresis migrates in 7 cells, raising p95 by up to 177x, and probation rolls back 5 of those moves. On Linux-dispatched heterogeneous boards, migration therefore cannot isolate a task from Linux interference, and recovery guards matter chiefly by preventing harmful moves. This paper supersedes an earlier artifact release whose reported improvement arose from measurement defects that we identify. All source, raw traces, and generators are public.
Large Language Models (LLMs) exhibit powerful code generation abilities, fostering agentic systems to automate edge AI application development. However, existing agents typically operate via fixed agentic workflows and predefined design spaces, struggling to handle rapidly evolving demands of real-world scenarios (e.g., heterogeneous data modalities, diverse task pipelines). Unlike prior systems that treat LLMs mainly as code generators following rigid workflows, we fully exploit their reasoning capabilities to allow agents to determine their own workflows on the fly. To this end, we propose SE2, a self-exploring and self-evolving paradigm for adaptive edge AI development. Inspired by human learning processes, we distill development into four atomic operations (retrieve, reason, code, test), and empower agents to explore workflows by dynamically and adaptively selecting an appropriate next operation based on the current context. Extensive evaluations show that, with adaptive workflows, the synthesized programs achieve up to 26% higher accuracy and consume 28% lower resources than state-of-the-art baselines, while maintaining superior adaptability to unforeseen edge AI scenarios. This highlights the potential of reasoning-driven adaptive agentic workflow in edge AI development.
Microwave linear analog computer (MiLAC) has been proposed as an analog processing architecture to perform efficient computation in the microwave domain. Consisting of reconfigurable impedance networks, a MiLAC can extract many of the gains of large multi-antenna systems without costly digital processing, e.g., achieving capacity in narrowband multiple-input single-output (MISO). However, the behavior of wideband MiLAC, required to support high data-rate transmissions, is widely unknown. The impedance components of MiLACs have non-trivial frequency-dependent response, requiring specific implementations to characterize it. In this work, we consider MiLAC implementations based on LC resonators with tunable capacitance, which allow synthesizing near-arbitrary passive lossless impedance matrices at individual frequencies using inductors and varactors. We characterize the mapping between optimal MISO precoding at a reference frequency, and the vector of capacitances configuring the MiLAC. We derive the degrees of freedom (DoFs) remaining after fixing such optimum precoding at the reference frequency, and show that these DoFs can be exploited to tune the frequency response of the MiLAC. We propose two techniques to exploit these DoFs: a closed-form approach for frequency-flat MiLAC processing, and a numerical approach for multi-frequency MiLAC processing, optimizing processing at frequencies other than the reference one. The frequency-flat approach shows satisfactory performance in a simulated frequency-flat millimeter-wave scenario, approaching the performance of digital processing. The multi-frequency approach can adapt to the frequency-selectivity of a simulated sub-6GHz scenario, attaining gains over frequency-flat digital processing. This constitutes the first demonstration that realistic MiLAC implementations can be configured to perform independent computations at different frequencies.
Our study systematically evaluates the robustness of fine-tuned pre-trained models across different datasets and attack types, aiming to provide insights for researchers in selecting pre-trained models for image classification tasks while prioritizing robustness against potential attacks. Our work explores the influence of fine-tuning pre-trained models with different datasets and attack types on their robustness and overview of the appropriate pre-trained model for image classification based on factors such as computation resources, memory usage, and accuracy on clean data. We fine-tune the models with weights. Evaluate the models on data before any attack and after different attack cases. The study utilizes publicly available image datasets such as MNIST, CIFAR10, CIFAR100, and SVHN to assess the robustness of pre-trained models. The results indicate that the robustness of fine-tuned models varies across different data distributions and sample sizes, emphasizing the need for tailored selection based on specific task requirements. The findings also highlight the varying performance of ResNet50, DenseNet, and MobileNet in defending against attacks on different datasets. Overall, this research provides valuable insights into the robustness of fine-tuned pre-trained models and their suitability for real-world applications in image classification.
Analog In-Memory Computing (IMC) is a promising paradigm for accelerating GEMM-based AI workloads by executing computations directly within the memory array in the analog domain. While IMC delivers substantial energy and throughput improvements, it can suffer from significant task accuracy degradation due to analog-to-digital converter (ADC) quantization noise, which occurs at the GEMM output and is thus different from weight/activation quantization, which occurs at the GEMM input. Mixed ADC resolution allows IMC modules to choose between high- and low-resolution ADC modes with different energy costs, offering a compelling accuracy-energy trade-off. However, selecting ADC resolutions across IMC units while respecting hardware granularity constraints remains challenging. In this paper, we propose SMART-AIMC, a Sensitivity-aware Mixed ADC Resolution framework with fine-grained Tunability for Analog In-Memory Computing. SMART-AIMC correlates the algorithmic task loss to hardware ADC quantization noise in the form of an ADC Channel Sensitivity (ACS) metric. Using this metric, SMART-AIMC systematically, and in a disciplined manner, allocates ADC resources to critical computations. Furthermore, we introduce a sensitivity-based channel reordering scheme that overcomes the hardware granularity constraints with no runtime overhead and approaches the optimal ADC quantization possible for the model. Evaluated across diverse workloads, SMART-AIMC unlocks a superior accuracy-energy Pareto frontier, providing (1) higher task accuracy at matched energy budgets and (2) fine-grained tuning across efficiency targets, from aggressive energy savings to near-digital inference accuracy. At the latter operating point, SMART-AIMC achieves $6.99\times$ higher modeled system-level energy efficiency for linear projections than the INT8 digital baseline with negligible accuracy degradation.
Deep learning (DL) models for real-time wearable sleep staging often rely on long-kernel temporal convolutions, a requirement that current accelerators struggle to support efficiently. To overcome this, we present an FPGA-based CNN-BiLSTM accelerator that incorporates several algorithm-hardware co-design techniques to enable efficient edge inference. Central to this design is an FFT-based non-unit stride fast convolution, which performs accumulation in the frequency domain to enable a single IFFT per block, reducing post-processing complexity to $O(N\log N)$. On the target long-kernel convolution layer, this yields a 20.5$\times$ reduction in arithmetic operations. Power-of-two quantization is applied across the pipeline, including FFT-based convolution and BiLSTM activations, with an average accuracy degradation of only 0.11% relative to the floating-point baseline across all subjects on the public MASS-SS3 dataset. Additional optimizations, including block convolution, depth-first execution, and hardware engine reuse minimize memory access and enhance resource utilization. The accelerator, implemented on a Xilinx Zynq-7000 FPGA, operates at 10 MHz and consumes 11 mW of dynamic power including data handling and configuration overhead from a microcontroller. The design achieves a DSP power efficiency of 0.91 GOPS/W/DSP using only 4 DSPs, demonstrating scalable and energy-efficient DL acceleration for sleep staging that can be extended to other wearable applications.
The growing computational demands of deep learning (DL) have driven AI-optimized reconfigurable devices, such as the AMD Versal AI Engines (AIEs), the Intel Stratix 10 NX Tensor Blocks (TBs), and the AMD Ryzen AI NPUs. Their heterogeneous resources and novel dataflow models make efficient exploitation nontrivial, and FPGAs lack native support for structured sparsity. This dissertation optimizes dense general matrix multiplication (GEMM), the dominant DL kernel, on these devices, proposes FPGA architectural enhancements for structured sparsity, and extends the techniques to AMD Ryzen AI NPUs. First, MaxEVA optimizes GEMM on Versal AIEs, achieving up to 2.19$\times$ higher throughput and 20.4% higher energy efficiency than the state of the art. It is then extended to arbitrary matrix sizes, and a systematic framework is proposed for the Stratix 10 NX TBs. Both frameworks reach state-of-the-art int8 throughput of up to 77 and 68 TOPS, and energy efficiency of up to 0.94 and 1.35 TOPS/W, on Versal and Stratix, respectively. Next, Systolic Sparse Tensor slices (SSTs) are proposed as in-fabric hard blocks supporting dense, 2:4, 1:4, and a new 1:3 sparsity level. SST-based GEMM accelerators achieve up to 5$\times$ higher frequency and 10.9$\times$ lower area than conventional FPGA logic, and on sparse ViT and CNN models SSTs deliver up to 3.52$\times$ speedup over dense in-fabric acceleration with up to 13.3% area increase. Finally, a common systematic methodology is proposed to optimize GEMM across the XDNA and XDNA2 AMD NPUs, reaching state-of-the-art performance of up to 6.76 and 38.05 TOPS (int8) and 3.14 and 14.71 TOPS (bf16), respectively.
Image super-resolution is widely used on mobile devices for upscaling various types of image data. This creates a need for efficient deep learning-based models that are not only performant but can also be deployed on neural processing units (NPUs) present in the majority of mobile and IoT chipsets. In this Mobile AI challenge, we address this problem and propose the participants to design fast solutions that can upsample images from Full HD to 4K resolution while running on popular mobile NPUs from Qualcomm and MediaTek. For this, the participants were provided with the DIV2K dataset and trained models to do an efficient 3X image upscaling. The competition had two tracks: in the first track, the models had to be fully quantized (INT8), and the runtime of all models was evaluated on the Qualcomm Snapdragon 8 Elite Gen 5 NPU. In the second track, the participants were developing floating-point solutions that were evaluated on the MediaTek Dimensity 9500 NPU. Top solutions proposed in this challenge are able to reconstruct 4K images under 5 milliseconds while reaching a PSNR score of 30 dB. A comprehensive description of the models developed in the challenge is provided in this paper.
On-device language models can learn a user's preferences through adapters trained on private data. However, a base-model upgrade may change the layers or vocabulary, leaving the personal adapter incompatible. Replacing it with a non-personal adapter loses much of the personalization, while retraining requires retaining the user's labeled history. We present Heirloom, a system that transfers a personal adapter to a new base model on the phone without retaining labels. Heirloom uses the old personal model as a teacher, collecting its predictions on stored unlabeled inputs, vendor-provided prompts, or prompts generated by the new base. It then unloads the old base and trains an adapter for the new model to match those predictions, keeping only one base model in memory at a time. We evaluate Heirloom across three generation upgrades from two model families, an architecture-preserving refresh, and three personalization tasks. On movie tagging, migration from stored inputs maintains pre-upgrade accuracy and differs from labeled retraining by at most 1.7 percentage points across the three upgrades. On a phone's neural processing unit (NPU), a fixed-length migration for the 4B upgrade takes 38 minutes and consumes 20% of the battery end to end, making the stored-input setting feasible during overnight charging.
Accelerating modern foundational models has pushed domain-specific compilers toward increasingly complex, multi-tiered optimization pipelines. However, this heavy optimization focus introduces delicate tradeoffs between peak execution performance and correctness. Subtle codegen defects can trigger numerical degradation and memory or buffer corruption, invalidating multi-million-dollar training runs and degrading live serving quality without raising obvious compiler errors. In this paper, we present a battle-tested methodology for verifying and triaging Google's XLA:TPU compiler at hyperscale. We leverage fuzzing techniques, combining synthetic program generation with real production traces, while orchestrating AI agent swarms to autonomously generate stress tests, reproduce failures, and draft fixes. To pinpoint root causes in full workloads, we pair this with a hierarchical differential pipeline that bisects large graphs down to individual failing operations inside fused kernels. Deployed in production, our framework has already driven the resolution of over 142 compiler bugs (with 103 additional findings under review), from which we distill pragmatic lessons learned.
Persistent homology (PH) is a mathematically principled multi-scale topological descriptor of shape and connectivity in data. It is widely applied in image classification, segmentation, and shape analysis across medical and aerial imaging, providing structural information that pixel-wise deep learning may overlook. Standard cubical PH decomposes an image into cubical cells (vertices, edges, and squares) and reduces them sequentially in filtration order. This sequential reduction has $Θ(n)$ parallel depth for an $n$-pixel image, limiting within-image GPU parallelism and constraining its use in topology-aware deep learning. We introduce CuPH, a GPU-based parallel algorithm for computing cubical PH on 2D images that operates directly on pixels rather than cubical cells. CuPH buckets pixels by filtration values, activates them with a level-synchronous sweep, and tracks connected components with a lock-free union-find method. We prove that CuPH produces persistence diagrams identical to standard cubical PH on integer inputs, with a stability bound for floating-point inputs. CuPH reduces parallel depth from $Θ(n)$ to $O(L\log n)$ for an image with $L$ intensity levels. We implement the level sweep as a single CUDA kernel, eliminating cross-level kernel-launch overhead. Empirically, CuPH is $18.6$--$99.6\times$ faster than the existing GPU baseline (CMP). As a drop-in PH backend, CuPH accelerates topology-aware classification ($4.2$--$16.7\times$) and segmentation ($8.3$--$44\times$) while preserving accuracy. It also enables computing full-image PH losses on high-resolution images, improving segmentation Dice by $2.2$--$4.6$ points over patch-local PH baselines and accelerating training by $25$--$31\times$.
Diffusion large language models (dLLMs) generate tokens in parallel via iterative denoising, but their bidirectional attention prevents the KV caching used by autoregressive models and recomputes the full attention map at every step, incurring quadratic per-step cost. Existing sparse attention methods for dLLMs construct a sparse mask after an initial full-attention warm-up and reuse it for the remaining steps, while allocating the sparsity budget uniformly across heads. We argue both choices ignore key structural properties of attention in dLLMs: empirically, attention patterns are only locally consistent across steps and shift abruptly at certain transitions, and different heads concentrate their attention mass to very different extents. To this end, we design Step-dLLM, which periodically refreshes a block-sparse mask at anchor steps along the denoising trajectory and reuses it within each locally consistent interval, while employing a head-adaptive global budget that more aggressively compresses heads with concentrated attention. Backed by a fused IO-aware Triton kernel for block-sparse attention, Step-dLLM applies to pre-trained LLaDA-1.5 and Dream-7B-Instruct without retraining, achieving a 2.78x end-to-end speedup over dense FlashAttention on LLaDA-1.5 at 8K context with 256 denoising steps and 90% sparsity, while improving the accuracy-latency trade-off over prior sparse-attention methods.
Relative-position attention mechanisms offer compelling alternatives to RoPE across various transformer architectures, delivering superior performance in domains such as audio. These mechanisms typically modify attention scores by adding content-position or bias-position terms. Unlike RoPE, which encodes position via per-token rotations directly on the queries and keys, relative-position terms generally require pairwise offset-dependent computation or lookups, complicating their integration into fused attention kernels and introducing severe memory, bandwidth, and compute overheads as sequences scale. To address this bottleneck, we propose an exact attention algorithm that incorporates additive relative-position terms into FlashAttention-style tiled computation. Our method (1) uses offset-aware tiling to avoid redundant indexing and memory accesses, (2) computes only the necessary key/query-relation products and reuses them across token pairs in both forward and backward passes, and (3) uses offset-based gradient sharding and aggregation, eliminating atomic contention. We provide a unified, efficient Triton implementation across diverse relative-position attention mechanisms and evaluate it against naive and specialized implementations. For 8K-token sequences, our method speeds up the forward (backward) pass by 1.8-6.4$\times$ (2.0-21$\times$) and reduces peak memory up to 2.9-128$\times$ (1.4-54$\times$) over the strongest baseline for each mechanism, enabling long-context end-to-end transcription that is 1.8$\times$ faster and uses 2.4$\times$ less memory. Our method thus provides a drop-in, hardware-efficient solution for scaling relative-position attention transformers.
In recent years, the growing adoption of Edge AI systems has been accompanied by a surge in security threats. Edge AI deployments face an rapidly evolving spectrum of security threats, from side-channel information leakage to ad- versarial model manipulation and physical device tampering. Conventional security mechanisms such as malware scanners and resource-intensive firewall solutions are unsuitable due to edge AIs typically limited computational resources, strict energy budgets and thermal constraints especially in battery-powered and field-deployed environments. This paper proposes a stacked, hardware-based security mechanism that enables low-overhead, continuous detection of anomalous execution patterns. Anomaly detection utilizing hard- ware performance counters and physical sensors outperforms software-based alternatives by providing key capabilities such as tamper-resistant root-of-trust isolation, negligible runtime overhead, and resilience against compromised operating systems. Experimental analysis shows that adversarial or malicious ap- plications produce observable deviations in sensor and thermal readings on Edge AI systems and hardware-based anomaly detection provides effective and low-overhead anomaly detection.
The Citizens Broadband Radio Service (CBRS) shares the 3550-3700 MHz band between incumbent radar systems and commercial LTE/5G users, so radar activity must be detected accurately and promptly. Spectrogram-based YOLO detectors can localize coexisting radar and cellular signals, but running them on resource-constrained edge devices is challenging. Model-distributed inference (MDI) can increase throughput by partitioning a model across edge devices. Using an end-to-end pipeline from raw I/Q samples to on-device detection, we show that partitioning recent YOLO generations incurs high activation-transfer costs due to their multi-scale structures. To reduce these costs, we propose CEDAR, a communication-efficient, edge-distributed architecture for radar detection. Its RF-aware detector, CEDAR-YOLO, removes YOLO's multi-scale neck and two of its three detection scales, leaving a single-scale path with independently configured frequency and time resolutions, which lowers the communication cost of partitioning. Its compression-aware MDI framework, CEDAR-MDI, integrates lightweight activation compressors and selects, for each compressor, the partition that minimizes the pipeline bottleneck. On an NVIDIA Jetson Orin Nano testbed with real-world over-the-air data, CEDAR-YOLO reduces the parameter count by 35-57% and model inference time by 16-33%, and CEDAR-MDI achieves up to $3.27\times$ the throughput of single-device inference, with detection quality preserved in both cases.
Low Tucker rank tensors provide compact representations of multidimensional dependencies, but standard learning pipelines can incur substantial computational and memory overhead from redundant intermediate quantities. We reduce this overhead without modifying the underlying representation objective or parameterization. Our approach combines three complementary components: (i) sequential extraction of only task-required spectral components, avoiding unnecessary full decompositions; (ii) factor-wise multilinear contractions without materializing large Kronecker operators; and (iii) compact latent-state updates without storing full ambient tensor iterates. These components preserve the algebraic operations and learning trajectories of the original algorithms while reducing their computational and memory footprints. Across five low Tucker rank tensor estimators, our implementations match the original trajectories to numerical precision (relative discrepancies of $10^{-16}$ to $10^{-13}$), while reducing per-iteration runtime by up to $3.12\times$ and resident memory by up to $616\times$ ($99.8\%$) on the main order-3 benchmark at $n=100$, with memory reductions exceeding $900\times$ across the full evaluation suite. These results show that substantial computational redundancy can be eliminated from Tucker representation learning without changing the underlying model or learned representation.
Randomized caches protect against side-channel attacks by obfuscating the mapping of memory addresses to cache locations. Existing proposals typically describe architectural behavior, leaving practical hardware implementation challenges largely unexplored. We implement four prominent designs, CEASER-S, ScatterCache, SassCache, and MIRAGE, and integrate them as the last-level cache (LLC) of a multi-core out-of-order RISC-V system-on-chip (SoC). We identify common and design-specific implementation challenges and develop practical solutions. We evaluate complete FPGA implementations by running SPEC CPU 2017 rate and GAPBS workloads to completion, providing, to our knowledge, the first comparative evaluation of these designs on a common full-system hardware platform. Pipelining the index derivation function (IDF) alone is insufficient: in three of the four variants, other components become the frequency bottleneck, reducing frequency by up to 35% relative to the baseline set-associative cache. Redesigning the critical paths while preserving mappings, rekeying, and victim-selection policies brings the three affected variants to no more than 10% below the baseline frequency, at the cost of added latency or reduced concurrency. Increased LLC access latency, rather than additional misses, dominates IPC degradation, reducing instructions per cycle by up to 5% for SPEC and 11% for GAPBS. Hardware overhead depends on the design and implementation technology: FPGA lookup-table usage increases by 10-35%, while ASIC area increases by 1-14%.
Two Boolean functions are considered NPN equivalent if one can be transformed into the other by negating inputs, permuting inputs, or negating the output. Efficient NPN classification is valuable in fields such as logic synthesis, quantum computing, and information theory. However, exact NPN classification demands exhaustive enumeration of all possible input permutations and negations. Its complexity grows exponentially, making it impractical for larger functions with more inputs. In this paper, we introduce the first learning-based formulation for NPN classification by recasting it as NPN action prediction. Our method represents Boolean functions as AND-Inverter Graphs (AIGs) and applies a bidirectional heterogeneous graph neural network to predict input negations, output negation, and input permutations. To mitigate the non-canonicity of logic networks, we further introduce graph rewriting techniques that reduce structural ambiguity and recover useful symmetries before prediction. Experimental results show that NPN action prediction is learnable, that the resulting canonicalization quality improves over heuristic baselines, and that on practical benchmarks our method approaches exact methods on tractable cases while scaling to functions with up to 16 inputs.
Deep neural network (DNN) accelerators for edge inference must meet tight energy, latency, and area (EDP) constraints, and their optimal hardware architecture and dataflow configuration is typically identified through design space exploration (DSE). However, on-chip memory technology including cell type, bank organization, and process node, is rarely treated as a first-class citizen in DSE. For example, increasing memory capacity can reduce memory accesses through greater data reuse, but at the cost of higher area and access energy, thereby limiting the compute units budget. Consequently, memory technology can fundamentally shift the optimal accelerator architecture for a given area and EDP budget. Existing machine learning (ML)-based DSE methods require repeating simulation, data generation, and model training for each new memory technology. We present TADA, a memory technology-aware DNN accelerator performance and energy estimator that predicts system latency and energy for a given workload, hardware configuration, and memory technology. TADA is trained on a synthetic dataset that generalizes across memory technologies. The dataset spans current and future technology trends while enforcing physics-based guardrails to preserve realistic area-energy trade-offs. Each memory technology is represented by a technology signature, a vector of read energy, write energy, and storage capacity across a range of areas, provided as input alongside the workload and hardware configuration. Across 21 unseen technology curves spanning SRAM, STTRAM, ReRAM, PCM, FeFET, eDRAM, and CTT, TADA predicts system energy and latency with 1.5-4.4% and 0.7-1.2% mean absolute percentage error, respectively. In three pairwise case studies, it matches the simulator-selected memory technology at 98.7-100% of area budgets, with a mean EDP loss of at most 0.21%, while achieving over 10^4x higher evaluation throughput.
The computational cost of conventional Convolutional Neural Networks (CNNs) limits their deployment in low-power devices for high-speed edge AI applications. This paper presents an end-to-end cosine-domain CNN (CD-CNN) architecture for energy-efficient FPGA inference using high-level synthesis (HLS). The proposed architecture performs intermediate feature extraction using real-valued DCT-domain representations, avoiding complex-valued arithmetic and intermediate inverse transforms between spectral layers. The design is implemented on a Xilinx Kintex UltraScale+ XCKU5P FPGA and evaluated against a fully spectral Fourier-domain CNN (FD-CNN) under consistent network and implementation constraints. Experiments on the MNIST and 94-class ASCII datasets show that the CD-CNN reduces normalized resource utilization and improves inference throughput on both datasets. Its energy efficiency reaches 112.91 frames/J for MNIST and 101.27 frames/J for ASCII, corresponding to improvements of $1.51\times$ and $2.76\times$, respectively, over the fully spectral FD-CNN architecture. Furthermore, compared to the previous GPU-based fully spectral CD-CNN implementation, the proposed end-to-end FPGA implementation improves energy efficiency by $12.96\times$. These results demonstrate the potential of real-valued spectral processing in the DCT domain for energy-efficient edge AI deployment.
Register-Transfer Level (RTL) simulators are fundamental to modern IC design workflows, yet like any complex software, they frequently contain critical bugs that compromise simulation accuracy. To systematically stress-test these tools, we introduce GI-RTL-Fuzz, a grammar-informed fuzzing framework that leverages Large Language Models (LLMs) to generate high-quality RTL test cases. Our key innovation is guiding the LLM's generation process using Backus-Naur Form (BNF) grammar coverage information, enabling the exploration of complex and rarely-tested language features. By performing differential testing against different RTL simulators, our framework uncovered a significant number of previously unknown inconsistencies in both the linting process and simulation results of major Verilog simulators.
Although neural video codecs such as DCVC-RT demonstrate state-of-the-art coding efficiency, their practical adoption remains limited by intrinsic computational cost. One of the critical bottlenecks for real-time implementation lies in non-linear activation functions, specifically the Weighted Sigmoid Linear Unit (WSiLU). This paper proposes a hardwareefficient approximation for WSiLU based on Lookup Table (LUT) discretization, replacing complex arithmetic with precomputed values and simplified control logic. Experimental results demonstrate that the proposed solution achieves an average BDrate increase of only 0.40%, outperforming polynomial-based approximations while maintaining significantly lower hardware costs. ASIC synthesis results for a 40 nm technology show that the proposed architecture supports Full HD $(\mathbf{1 9 2 0} \times \mathbf{1 0 8 0})$ video in real-time at 30 fps, with a reduced area footprint and low power dissipation.
This manuscript presents a congestion-aware RTL-to-GDS-II physical synthesis of a RISC-V Execute (EX) pipeline stage at the 90 nm CMOS technology node using Cadence tools- Genus and Innovus. The EX stage integrates a 32-bit ALU, forwarding unit, branch comparator, barrel shifter, operand multiplexers, and EX/MEM pipeline registers, making it inherently prone to routing congestion due to converging datapaths. RTL simulation validates all functional blocks across four stimulus sets. Logic synthesis yields 1363 cells, 8256.26 μm2 area, and 0.625 mW power with a worst-case setup slack of +28 ps. Pre-route congestion prediction reveals hotspots at the ALU and operand MUX regions; mitigation via cell spreading, density screens, and congestion-driven placement re-optimization eliminates all routing overflow. Timing is verified across seven stages—from post-placement through post-route optimization—with final setup slack of +270 ps and hold slack of +55 ps. Postnanoroute verification confirms zero errors and zero warnings. A valid GDS-II output is generated, and a comparison with prior works confirms the novelty and completeness of the proposed methodology.
Modern trigger and data acquisition systems for high-energy physics experiments must process large data volumes while meeting stringent real-time performance requirements. Reconfigurable platforms based on FPGA (Field Programmable Gate Arrays) provide the adaptability needed to support evolving detector architectures and processing algorithms, but their development and validation grow increasingly difficult as system complexity scales. This paper presents the redesign of the ATLAS calorimeter back-end firmware using a High-Level Synthesis (HLS) approach, centered on a parallel switch matrix that streamlines data routing and improves architectural scalability. The framework accelerates the design and maintenance of large-scale FPGA firmware in modern trigger systems, enabling systematic exploration of implementations and reducing verification overhead through parameterized configurations.
AutoFV is a template-driven framework that automatically generates a standardized formal verification environment starting from a Register-Transfer Level (RTL) design directory. The framework addresses practical challenges associated with setting up formal verification infrastructures, including the management of compilation scripts, tool configuration, property organization, and environment constraints. Given an RTL design, AutoFV automatically produces a complete project skeleton including Makefile orchestration, formal execution scripts, compilation filelists, and structured directories for properties, assumptions, and coverage specifications. The framework promotes maintainability through macro-based SystemVerilog Assertions (SVA) conventions for assertions, assumptions, and coverage, while supporting non-intrusive verification through hierarchical binding that preserves the original RTL code. Additionally, AutoFV introduces a context-dependent specification mechanism, assume-guarantee (ROLE), allowing properties to change their verification role depending on the verification context, such as acting as assumptions during block-level verification and as assertions during higher-level integration. The resulting infrastructure reduces manual setup effort, improves reproducibility, and accelerates the bring-up of formal verification environments from simple RTL blocks to full system integration.
This paper aims to showcase an in-depth study of a stochastic system that could be used in complex systems rather than going for traditional methods. A stochastic system works with an approximation technique where exact accuracy may be moderately compromised compared with a conventional binary system, but even though we could receive an acceptable accuracy with low power, lower delay and lower area consumption VLSI metrics. Stochastic Computing (SC) has emerged as an effective approximate-computing architecture for energy- and area-constrained systems, especially in error-tolerant applications such as signal processing, error-correcting codes, and neural networks. This paper surveys recent advances in stochastic and hybrid SC architectures, focusing on accuracy, correlation management, latency reduction, fault tolerance, and FPGA/ASIC realizations. Key challenges—long bit-streams, SNG overhead, correlation-induced errors, and binary–stochastic conversion costs—are reviewed, along with techniques that exploit correlation as a design resource. Hybrid stochastic logic and multibit representations are analyzed for mitigating precision and latency limits while preserving robustness and managing the power-area-delay constraints. Application-driven designs, including LDPC decoders, stochastic inner products, FSM-based fault-tolerant SC, and fully hardware SCNNs, demonstrate significant gains in throughput, area efficiency, and transient-fault resilience. The study indicates that modern stochastic and hybrid stochastic computing architectures can achieve near-binary accuracy with substantially reduced hardware cost and power consumption thereby reducing the overall delay when carefully designed. The paper concludes by outlining open research challenges and future directions toward scalable, high-accuracy, and application-specific stochastic computing systems for emerging edge and near-sensor platforms. The survey further highlights emerging trends such as adaptive precision, reusable stochastic number generators, and correlation-aware synthesis for improved scalability. Comparative analysis reveals key trade-offs among accuracy, latency, and hardware efficiency. A brief on stochastic computing works, challenges and observations on future scopes are portrayed in this study.
Traffic signal controllers form a critical component of modern transportation infrastructure, where reliability, responsiveness, and efficient resource utilization are essential for safe and uninterrupted operation. This paper presents the design and FPGA implementation of an intelligent traffic signal controller based on an eight-state Moore Finite State Machine (FSM) architecture. The proposed controller incorporates emergency vehicle preemption, pedestrian crossing management, and adaptive green-time extension using traffic density information, enabling more realistic operation than conventional fixed-time controllers. To investigate the impact of state representation on hardware performance, three widely used FSM encoding techniques-Binary, Gray, and One-Hot-were implemented and evaluated on a Xilinx Zynq-7000 XC7Z020 FPGA using Vivado Design Suite 2016.4.A comprehensive comparison was performed in terms of resource utilization, power consumption, and timing performance under identical synthesis and implementation conditions. Experimental results show that Binary encoding achieves the lowest hardware utilization with only 60 LUTs and 44 flip-flops, while Gray encoding provides the highest timing margin with a Worst Negative Slack (WNS) of 13.687 ns. Both Binary and Gray implementations exhibit identical power consumption of 0.128 W total on-chip power, whereas One-Hot encoding increases dynamic power consumption by 50% and requires nearly twice the LUT resources. All implementations successfully satisfy timing requirements at 100 MHz with zero setup and hold violations. The results demonstrate that Gray encoding offers the most balanced overall performance, while Binary encoding remains an attractive option for resource-constrained deployments. The findings provide practical design guidance for FPGA-based intelligent traffic management systems and other embedded control applications employing finite state machines.
In general, traffic accidents are one of the main causes of death and financial loss. The majority of today’s car safety systems rely on microcontrollers or Internet of Things platforms, which have poor pre-crash data retention, low sampling rates, and limited parallelism. This work describes a deterministic FPGA-based vehicle accident detection and event data recording (EDR) system built using VHDL on the Nexys4 DDR board. Three levels of hardware parallelism are achieved by the system: pipelined frame formatting with CRC-8 computation and BRAM writes; concurrent severity classification with dual-phase buffer switching upon crash detection; and parallel multi-sensor acquisition across SPI, UART, and GPIO interfaces. Sensor data, severity classification bits, timestamps, and CRC-8 integrity verification are all integrated into a fixed-width, 96-bit structured Event Data Frame. The system stores 1024 frames of pre-crash data (10.24 s) in a circular BRAM buffer and 512 frames of post-crash data (5.12 s) in a linear BRAM buffer. Using peak acceleration, sustained duration, and delta velocity, a hardware-based three-parameter severity classification system yields four severity levels: No Crash, Low, Medium, and High. The results demonstrate accurate severity classification with validated 96-bit frames exported via UART to the PuTTY terminal interface.
This paper presents an RTL-level design and functional validation of the link layer in the JiaoXianjun open-source BTLE repository, establishing its correctness as a foundation for both FPGA-based hardware deployment and physical IC synthesis. The work is positioned strictly at the digital hardware design level, distinct from system-level BLE protocol analysis or IoT application studies. The analysis covers the module hierarchy, the bit-serial packet processing pipeline comprising a 24 -stage LFSR CRC24 module and a 7-stage LFSR data whitening module, and an eight-correlator parallel Access Address Detector. A key finding is the absence of a functional finite state machine (FSM) in the original link layer controller, which creates a structural asymmetry between a fully realized data plane and an unimplemented control plane. To address this gap, a sevenstate link layer FSM and an autonomous advertising engine with a 20 ms inter-packet gap and a 4 ms TX watchdog are designed and integrated into btle_ll.v without modifying any existing data plane submodule. Correctness is validated through three layers: module-level simulation against a Python golden model, end-to-end simulation using Icarus Verilog, and hardware-inthe-loop validation of the RX path on a ZedBoard platform using an external PHY via over-the-air packet capture. Physical synthesis via OpenLane on the SkyWater SKY130 PDK produced a GDSII layout of the link layer with a die area of $2,087,530 \mu \mathrm{~m}^{2}$, confirming physical synthesizability of the link layer logic. A twostage development roadmap toward a complete open-source BLE SoC is identified as future work.
Continuous, non-invasive monitoring of central hemodynamics is essential for advanced cardiovascular care. Although wearable ultrasound enables deep-tissue sensing, its widespread clinical adoption is limited by the inherent physical trade-off between driving voltage and acoustic bandwidth, alongside the challenges of scaling microelectronic ultrasound devices for medical imaging. Here, we report a wearable, 81.7% broadband piezoelectric micromachined ultrasonic transducer (pMUT) array driven at a low voltage of 5 Vp. To address these bottlenecks, this study introduces two parallel innovations. First, we introduce a novel frequency-merging design, quantified by an equivalent circuit model, to effectively decouple the voltage-bandwidth trade-off. Second, we provide a comprehensive framework directing a scalable silicon-on-insulator wafer-level fabrication process towards fully in vitro and in vivo experimental characterizations for medical imaging. Enabled by these advancements, in vitro characterization demonstrates 0.64 mm axial and 1.27 mm lateral resolutions while maintaining a penetration depth of up to 9 cm in wire phantoms. In vivo, the wearable patch achieves comparable high-fidelity B-mode and M-mode imaging of carotid, brachial, and radial arteries within 4 cm depth to commercial rigid probes. Crucially, the device enables continuous, stable hemodynamic monitoring under limited daily postural activities for up to 9 hours. Inter-subject validation confirms precise blood pressure measurements, with deviations of 2.2 mmHg compared to FDA-approved clinical monitors. Leveraging standard semiconductor batch manufacturing, this work addresses the technical mismatch between low-voltage wearable devices and high-performance ultrasound transduction, creating a feasible development path for highly integrated, mass-producible cardiovascular ultrasound systems-on-chip.
IMPORTANCE Artificial intelligence (AI) is increasingly integrated into health care delivery, but its influence on consumer perceptions of physical therapist services remains unclear. OBJECTIVE The objective was to examine how the use of AI influences consumer perceptions of in-person and telerehabilitation physical therapist services. DESIGN This study used a randomized 2 × 2 factorial experimental design comparing AI use and delivery mode. SETTING The experiment was conducted online in Australia. PARTICIPANTS The study included 710 Australian adults recruited through a paid panel. EXPOSURE Participants were randomized to 1 of 4 hypothetical physical therapist consultation scenarios (in-person vs telerehabilitation; AI used vs not used). MAIN OUTCOMES AND MEASURES Perceived warmth; competence; safety; provider trust; emotional trust; relational and functional service quality; working alliance; and intention to use were measured on 7-point Likert-scales. Composite mean scores were analyzed using ordinary least squares regression with heteroskedasticity-robust standard errors. RESULTS Mean scores in the no-AI, in-person condition ranged from 4.8 to 5.8. Among in-person scenarios, described AI use was associated with lower ratings across all outcomes, including warmth (b = -0.3, 95% CI = -0.5 to -0.1), competence (-0.4, -0.6 to -0.2) and safety (-0.8, -1.1 to -0.6), provider trust (-0.5, -0.7 to -0.3) and intention to use (-0.8, -1.0 to -0.5). Without AI, telerehabilitation also reduced ratings by 0.2 to 0.7 points across outcomes. Positive interaction effects indicated attenuation of AI-related reductions in telerehabilitation for relational (interaction b = 0.4, 95% CI = 0.1 to 0.8) and functional service quality (0.36, 0.02 to 0.70), intention to use (0.6, 0.1 to 1.0), and perceived safety (0.5, 0.1 to 0.9). CONCLUSIONS Hypothetical in-person consultations described as involving AI were associated with poorer consumer perceptions of physical therapists and services. Telerehabilitation was also perceived less favorably than in-person care. RELEVANCE Future studies should test whether transparency and reassurance mitigate the effects of AI-use on consumer perceptions of physical therapy.
Elliptic partial differential equations (PDEs) constitute a fundamental class of governing equations in physics and engineering, underpinning the mathematical descriptions of electrostatic potential distribution, steady-state heat conduction, and linear elasticity. The numerical solution of these equations on fine discretization grids is computationally intensive, as classical iterative solvers require numerous sweeps to achieve convergence. This computational bottleneck critically hinders real-time applications such as online process control, interactive design optimization, and inverse problem reconstruction, where rapid response to changing boundary conditions or source terms is essential. The traditional numerical method cannot meet the real-time application requirements because of its long calculation time when solving on fine grids. In this study, the solution framework of deep Q-network is constructed, and the elliptic PDE solution is modeled as Markov decision processes. The state space includes solution field, residual field and boundary mark on discrete grid. The action space is 11 discrete corrections (− 0.10 to 0.10 with an interval of 0.02). The reward function is weighted sum of PDE residual term (weight 0.6), boundary error term (weight 0.3) and smooth regularization term (weight 0.1). The feature extraction network uses three-layer convolution and two residual blocks to output 512-dimensional vectors. The proposed DRL solver achieves relative L2 errors of 0.28%, 0.39%, and 0.62% on Poisson, Helmholtz, and anisotropic diffusion test cases, respectively, outperforming the finite element method (0.32%, 0.45%, 0.78%) and the multigrid method (0.31%, 0.44%, 0.76%). In the inference phase, the DRL solver completes a single forward pass on a 256 × 2562 grid in 0.15 s, achieving a 9898-fold speedup over the multigrid method’s 14.8 s. The training process converges stably within 20,000 steps, with the residual norm decreasing from an initial value of 12.4 to a final value of 0.16. This method provides a feasible way for solving and controlling real-time physical fields.
As the deployment of small modular reactors (SMRs) and advanced reactors expands, physical protection system regulatory frameworks are shifting from prescriptive, compliance-driven approaches to risk-informed and performance-based methodologies. In Korea, the amendment to the Act on Physical Protection and Radiological Emergency, promulgated on March 31, 2026, establishes a legal basis for the institutionalization of Security by Design (SbD) by requiring physical protection regulations to be reviewed from the Standard Design Approval (SDA) stage onward. This study examines the applicable regulatory direction for the SDA stage of Light-Water SMRs based on an analysis of regulatory trends in major countries. It investigates alternative requirements reflecting SMR-specific characteristics and performance-based regulatory approaches and proposes a preliminary regulatory framework incorporating performance assessment feedback into design improvement. The findings serve as a technical foundation for developing review guidance at the SDA stage to advance SbD institutionalization in Korea.
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.
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.
This study addresses the air ticket sales forecasting problem for business-to-business (B2B) travel agencies, a largely unexplored area compared to airlines and business-to-consumer agencies. To meet guarantee deposit requirements and manage financial risk exposure, B2B agencies require accurate demand estimation for budget planning and revenue management. The study proposes a hybrid deep learning forecasting model integrating recurrent neural network (RNN), long short-time memory (LSTM) and bidirectional long short-term memory (BiLSTM) architectures to capture both short-term fluctuations and long-term temporal dependencies in ticket sales data. Approximately eight years (from 3 August 2014 to 2 June 2022) of daily search and ticket sales data from a B2B travel consolidator agency were combined with external variables, including holidays, weather conditions, exchange rates, gross domestic product per capita and COVID-19 cases. Two datasets (with and without pandemic data) were constructed. After preprocessing and Random Forest feature selection, seasonal autoregressive integrated moving average (SARIMA), Prophet, RNN, LSTM and BiLSTM models were implemented and compared with the proposed hybrid architecture using a sliding-window forecasting framework. Performance was evaluated using mean absolute percentage error (MAPE) and root mean squared error (RMSE). The proposed hybrid model achieved the highest forecasting accuracy across both datasets, outperforming statistical and individual deep learning models. It obtained MAPE values of 5.83% and 6.16% and RMSE values of 0.0379 and 0.0411, respectively. BiLSTM ranked as the second-best method, while SARIMA and Prophet showed substantially lower accuracy. Results also indicated that excluding pandemic variables did not significantly reduce predictive performance, but still, pandemic-related volatility increased the error rate of most of the methods and increased forecasting difficulty. This research is the first to examine ticket sales forecasting specifically for B2B travel agencies. It introduces a novel hybrid sequential deep learning architecture and evaluates the role of pandemic and macroeconomic variables in forecasting performance. The findings demonstrate the effectiveness of hybrid recurrent models for complex real-world demand forecasting problems.
Sentiment analysis is one of the most important branches of natural language processing. Sentiment analysis has become a key research area in natural language processing, driven by rapid advancements in deep learning architectures. Where It is used to extract opinions and emotional sentiments from texts about a specific topic or new product, as well as from political and financial news headlines. In recent years, recurrent neural network models, convolutional neural networks, and, notably, transformer-based architectures have significantly improved performance across a wide range of natural language processing tasks. This paper presents a systematic review of deep learning approaches used in sentiment analysis published between 2020 and 2025, following established systematic review guidelines, including a specific search strategy . After applying inclusion and exclusion criteria, a number of final studies were selected for analysis and comparison in terms of methodologies, accuracy, and data used.The study concluded that advances in deep learning models have significantly improved the accuracy of sentiment analysis. The review also recommends developing models capable of more effectively handling linguistic complexities, such as sarcasm, irony, regional dialects, and multiple dialects.
Cloud-edge collaborative inference, particularly the synergy between edge-deployed small models and cloud-based large models, demonstrates extensive application value in meeting the high-accuracy and low-latency demands of emerging artificial intelligence. However, existing dynamic offloading methods exhibit limitations in modeling the “many-to-many” high-order resource competition that arises when numerous small models concurrently access shared large model interfaces. They lack a unified representation capable of adapting to time-varying network congestion and often ignore the spatial exclusivity of concurrent tasks when sharing bandwidth and large model interfaces. To overcome these limitations, this paper proposes a spatiotemporal hypergraph-driven framework integrating high-order topological feature extraction with dynamic resource modeling. This framework introduces dynamic hypergraph sequences to naturally encompass local conflict domains, mitigating the topological blind spots and “over-smoothing” issues inherent in traditional pairwise graphs. In the first stage, a spatiotemporal hypergraph neural network (ST-HGNN) encodes complex spatial dependencies and predicts temporal resource bottlenecks. Subsequently, a weighted multi-dimensional hypergraph matching (WTHM) strategy transforms the NP-hard mixed-integer nonlinear programming problem into an efficient heuristic search. Trace-driven simulation results based on real-world urban trajectories (EUA and T-Drive) demonstrate that the proposed framework achieves a robust normalized VCI score of 0.64 even during extreme congestion peaks. By consistently achieving an absolute VCI improvement of 0.17 to 0.38 over traditional bipartite-graph and proximity-based baselines, the framework significantly mitigates cascading queuing failures. Moreover, the developed dual-timescale asynchronous architecture strictly confines online execution to the millisecond scale, validating its capability to balance inference accuracy and latency across dynamic cloud-edge environments.
Deploying 3D point cloud models on edge hardware such as the NVIDIA Jetson Orin Nano is severely constrained by compute and memory budgets. Existing compression methods require access to the model's original source code, rendering them inapplicable to the Open Neural Network Exchange (ONNX) binaries commonly distributed by vendors and model repositories. We present \textbf{H3DNAS}, a hardware-aware model compression framework that operates directly on ONNX computational graphs without requiring original source code, architecture class definition, or gradient access during search. H3DNAS makes three contributions: (1) a \textbf{Channel Dependency Graph (CDG)} that classifies ONNX operators into four constraint classes and formally establishes that the free parameter fraction $\rho_f$ is topological invariant, a provable compression ceiling computable in $\mathcal{O}(|V|+|E|)$; (2) a \textbf{Two-Stage Hierarchical Search} that prunes candidate architectures by $L_1$-importance channel selection, ranks them by output fidelity as a zero-shot label-free proxy, and applies GhostConv structural mutation to Pareto-optimal candidates; and (3) the \textbf{first source-code-free compression pipeline for 3D point cloud models}, operating entirely via ONNX graph surgery with no original architecture definition required. On ModelNet40, H3DNAS reduces the number of parameters in PointNet, PointNet++, and PointMLP by $65.5\%$, $43.2\%$, and $49.1\%$, respectively, while achieving $1.99\times$, $1.29\times$, and $1.67\times$ inference speedups with negligible loss in accuracy. The source code is publicly available\footnote{https://github.com/ClarityLab-Org/h3dnas}.
This paper presents an improved optimization methodology for thermoelectric microwave power sensors utilizing backpropagation (BP) neural network algorithms, addressing the critical issues of high return loss and insufficient sensitivity in conventional structures operating in the X-band (8-12 GHz). By establishing theoretical models incorporating key geometric parameters, including thermocouple height and substrate membrane thickness, comprehensive training datasets were acquired through full-wave electromagnetic simulations using High Frequency Structure Simulator (HFSS). A BP neural network was employed to construct the nonlinear mapping relationship between sensor geometric parameters and return loss. Global optimization search based on the neural network surrogate model enabled systematic analysis of the influence of structural parameters on microwave characteristics. The study reveals that collaborative optimization of thermocouple height and thermistor spacing effectively mitigates parasitic effects. Furthermore, the physical realization of these precisely optimized micro-structures relies heavily on advanced optical inspection and lithography techniques to ensure the dimensional accuracy of the substrate membrane and thermocouples. The optimized sensor architecture achieves a return loss of -24.49 dB at the center frequency of 10 GHz, representing a 4.7 dB improvement over the initial design, while maintaining structural stability.
Recognizing visual artworks is difficult because images of art combine large within-class diversity, ambiguous boundaries, and non-photographic textures with intertwined cues related to the artist, genre, style, and historical period. Standard convolutional or transformer models are commonly inherited from natural-image classification, and such backbones may fail to emphasize the visual statistics that distinguish paintings, drawings, prints, and other artistic media. This paper presents ArtNAS-Style, a differentiable neural architecture search framework that embeds artistic style priors into the architecture search stage for visual art recognition. Rather than optimizing architecture choices only through classification loss, the framework employs a style-prior-aware bilevel objective that jointly accounts for recognition accuracy, diversity of style-sensitive features, preservation of brushstroke-like texture, and robustness across stylistic domains. The proposed Style Prior Encoding Module derives Gram-statistic, spectral, and palette-distribution cues from each artwork, and the Prior-Conditioned Search Cell uses these cues to adaptively modulate candidate operations during the search. A Style-Balanced Architecture Regularizer further discourages the discovered architecture from specializing to dominant artistic categories and promotes discriminative features across heterogeneous art domains. The search model is compact, trainable end to end, and transferable to multiple recognition tasks. Our evaluation protocol covers style, artist, school, and object-type recognition, in addition to parameter-matched ablations, robustness, statistical testing, and computational costs.
Estimating whether a problem is too complex for a large language model (LLM) before the model attempts it can guide model routing, how to allocate test-time compute, and suggest the need for new tools to aid the agent. While existing estimators rely on the problem text, LLM-annotated task difficulty labels, or signals obtained after a model attempts to solve the problem, such as its hidden states, none explicitly leverage the constraint structure inherent in the problem: the variables to assign and how the constraints over them interact. We introduce CIRCE, which extracts a problem's structure by using a cheap LLM to construct a Constraint Intermediate Representation that deterministically compiles into the constraint graph. Across seven reasoning benchmarks (including SAT, logic, and math) and four LLMs, even a randomly initialized GNN encoder improves predictions of LLM response correctness, reducing negative log-likelihood from 0.465 to 0.423 and increasing AUROC from 0.691 to 0.727 relative to the strongest baseline for each metric.
Retrieval-augmented generation is becoming agentic. A search harness writes queries, reads results, reformulates, and decides when to stop. This largely solves accuracy: a harness with only BM25/grep and MiniMax M3 answers 85% of BrowseComp-Plus and 91% on multi-hop FRAMES in our runs, without compiled or vector indexes. It exposes a context problem: an undisciplined loop repeats queries, rereads documents, and fills the window with unused passages. The missing layer is knowledge compilation: a configurable ingestion pipeline where LLMs enrich documents with context no chunk contains alone--navigation hierarchies with claim-evidence leaves, entity graphs, timelines, mind maps, ontologies, wikis, and memory. Compilation is not a better retriever: under a competent harness, structural artifacts gave no detectable accuracy gain on FRAMES, and a PageIndex-style outline degraded at scale. It is a supply layer. A navigation tree whose leaves are claim-evidence pairs--the mechanism we measure--cuts query tokens by 91% when evidence suffices and adds 1.9% when it does not, a 38% average reduction; this implies a 2% break-even sufficiency rate and a measured rate near 43%. A decision gate cuts latency from 140 s to 30 s, though substitution's token saving is bought with retrieval rounds and carries its own latency cost. Other artifacts--ontology, graph, wiki, timeline, mind map, memory--pay off in deployment (governance, integration, onboarding, temporal orientation, continuity) where accuracy benchmarks cannot see them. We describe RAGFlow's pipeline and harness (open source at https://github.com/infiniflow/ragflow), which turns the loop into five disciplined stations and keeps the context entering the model the smallest sufficient set; together they form an agent-native database: a store that hands the agent compiled, governed context instead of a ranked list of passages.
Accurate prediction of molecular dipole moments (μ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu$$\end{document}) is essential for modeling electrostatic interactions in solvation, molecular recognition, and materials. While modern 3D learning models can approach saturation on low-noise quantum-chemical benchmarks, performance often degrades on experimentally compiled datasets where measurement noise, stereochemical ambiguity, and conformer variability introduce a substantial gap between idealized theory and practical data. Here, we present a progressive 2D–3D hybrid framework for experimental dipole-moment prediction under heterogeneous labels. The framework is constructed in two stages. First, a strong 2D tabular CatBoost model based on SMILES-derived multi-view fingerprints is enriched with physicochemical, dipole-related, and conformer-dependent 3D shape descriptors, introducing a first level of 2D+3D integration and improving performance over the 2D representation alone. Second, this enriched tabular predictor is linearly fused with a geometry-aware 3D graph neural network operating on conformer-derived molecular graphs, yielding a further gain through complementary structural learning. On the experimental test set, the resulting framework achieves R2=0.844\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}=0.844$$\end{document} with MAE=0.399D\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{MAE}=0.399~\textrm{D}$$\end{document}, outperforming a nonparametric fingerprint k-NN baseline (R2=0.65\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}=0.65$$\end{document}) and remaining close to the similarity-conditioned diagnostic reference range estimated from twin-pair dispersion (R2≈0.86\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{2}\approx 0.86$$\end{document}–0.91) under Tanimoto similarity thresholds of 0.95, 0.97, 0.98, and 0.99. The main scientific contribution of this study is the development of an experimentally grounded progressive 2D–3D hybrid learning framework that integrates enriched tabular chemical representations, geometry-aware graph learning, and twin-pair-based diagnostic evaluation to predict dipole moments under heterogeneous experimental labels. To contextualize performance in this noise-affected experimental setting, we further use twin-pair analysis within a test-to-train similarity-stratified evaluation protocol, while baseline comparisons and component-wise interpretability analyses help clarify the roles of representation diversity and learner complementarity. Relative to three recently published dipole-prediction studies, the proposed framework remains competitively positioned, although these comparisons are intended only for cautious literature contextualization because the underlying datasets, label sources, and noise regimes are not directly matched. In contrast to prior studies focused mainly on lower-noise theoretical benchmarks or narrower molecular domains, the present work targets the comparatively underexplored Stenutz experimental compilation.
Stress is an important physiological and psychological reaction which, when prolonged, has a negative impact on the health and well-being of humans. Traditional stress assessment methods tend to be subjective and are not scalable in real-time. In this research, a lightweight machine learning framework to predict stress from easily measurable environmental and activity-related features, temperature, humidity, and steps, is introduced. The proposed method is different from the approach that depends on complex physiological signals and uses minimal non-invasive inputs that can be applied in the wearable and IoT field. The stress level is considered as the target variable, and all the features used as input variables are independent from each other to prevent data leakage. There are several machine learning and deep learning models and their performances are measured, such as LightGBM, Support vector machine (SVM), AdaBoost, Gated Recurrent Unit (GRU). Bayesian hyperparameter optimization is used to improve the performance of the model and guarantee strong generalization. Experimental results show that LightGBM outperforms the other models, giving an accuracy of 97% and a balanced precision, recall and F1 score of 95%. The results demonstrate promising predictive performance on the evaluated dataset. However, since the model was validated on a relatively small publicly available dataset, further evaluation using larger and more diverse real-world datasets is required before practical deployment.
Two-dimensional edge plasma transport simulations provide a detailed description of divertor plasma behaviour, but their computational cost limits large parameter scans and iterative modelling. We present a database of 23,379 converged EDGE2D-EIRENE simulations for MAST Upgrade (MAST-U), spanning three divertor magnetic configurations, variations in separatrix density and input power, deuterium fuelling locations, and nitrogen seeding rates and locations. Independent multilayer perceptron surrogates are trained for twelve scalar outputs, including divertor front positions, target electron temperature and core-edge boundary quantities. On a held-out test set, the surrogates achieve $0.9462 \leq R^2 \leq 0.9994$, substantially outperforming linear regression baselines. Dense surrogate scans quantify the dependence of carbon radiation front motion on separatrix density, divertor geometry and fuelling location. Nitrogen seeding scans reveal a strong dependence on seeding location, with high-field-side seeding rapidly reducing core-boundary electron temperature, while divertor seeding produces little response. Batch inference for all twelve outputs at 786,432 operating points takes 1.76 s on GPU, compared with a median wall time of 2.61 hours using four CPU cores for separate EDGE2D-EIRENE simulations. These models enable rapid comparison of geometry and source-location effects on divertor and core-boundary conditions within the sampled simulation domain.
The explosive growth of the Internet of Things (IoT) across smart homes, healthcare, industrial systems, and critical infrastructure has dramatically enlarged the cyberattack surface, rendering traditional security mechanisms ineffective against evolving threats in highly heterogeneous and resource-constrained environments. This article presents a comprehensive and systematic review of machine learning (ML) and deep learning (DL) based intrusion detection systems (IDSs) tailored for IoT, with a particular focus on optimization algorithms for feature selection (FS), hyperparameter tuning, and model lightweighting. We critically analyze state-of-the-art methodologies, benchmark datasets, deployment strategies, and performance trade-offs, while uncovering previously hidden correlations between optimization techniques and ML/DL models that significantly enhance detection accuracy, interpretability, energy efficiency, and real-time feasibility on edge devices. The effectiveness of these approaches is systematically evaluated across diverse IoT applications, including real-time anomaly detection, lightweight edge deployment, and encrypted traffic analysis. Key challenges in integrating advanced ML/DL into constrained IoT environments are thoroughly examined, and a forward-looking research framework is proposed to overcome current limitations and drive the development of the next generation of intelligent, adaptive, and resilient IoT-IDS solutions.
A comprehensive analysis of the advantages and limitations of implementing intelligent control systems for domestic water heaters in Russian housing and utility services is presented. The relevance of the study is due to the high share of costs for heating and hot water supply in the energy consumption structure. An analysis of modern Russian and foreign solutions in the field of heating supply automation, the implementation of IoT systems, machine learning methods, and digital twin technologies was conducted. The architecture of an IoT system is described, including temperature and flow sensors, microcontrollers, and Wi-Fi modules for remote monitoring. Machine learning algorithms have been developed for predicting water consumption, allowing heating modes to be adapted to individual user needs. The methodology is based on mathematical modeling of thermal processes, taking into account Russian building codes and the climatic conditions of three cities: Moscow, Yekaterinburg, and Sochi. A comparative analysis of the energy efficiency of traditional and intelligent modes showed a reduction in energy consumption of 21.4% to 26% depending on the region; the greatest effect was recorded under conditions of a prolonged heating period. An economic assessment was performed depending on the region, tariffs, and equipment cost. The integration of intelligent boilers into the “smart city” concept and into a control system that ensures balancing of the grid load and participation in dynamic pricing programs was considered. The regulatory requirements of the Russian Federation were analyzed, identifying the main barriers to implementation: high initial costs, insufficient public awareness, and the absence of standardization of IoT devices. Ways to overcome the identified potential limitations are proposed: government subsidies, educational programs, and the development of national standards. The significant potential of intelligent water heaters as an element of digitalization of housing and utilities sector and for achieving national energy efficiency goals is shown.
Production continuity depends directly on the reliability, efficiency, and recoverability of connected physical assets, on which Saudi Arabia’s industrial diversification and digital transformation increasingly depend. This review critically discusses Internet of Things-enabled predictive maintenance for cyber-physical systems, centering on integrating sensing, edge and cloud computing, machine learning, remaining useful life estimation, digital twins, and maintenance decision support into an operational capability aligned with Vision 2030. We used a structured integrative review approach to synthesize thirty-seven peer-reviewed journal articles published between 2020 and 2025, prioritizing studies from Q1 and Q2 journals. The synthesis shows that predictive maintenance creates value only when accurate models are embedded in a trustworthy end-to-end loop linking condition data, prognosis, operational context, maintenance planning, execution, and learning. The evidence supports temporal deep learning, hybrid and physics-informed prognostics, digital twin-assisted reasoning, and edge analytics but also points out recurring weaknesses such as scarce failure labels, model drift, interoperability, cybersecurity, false alarms, and weak integration with work management. These limitations are additionally exacerbated in the Saudi industrial context by brownfield heterogeneity, severe operating conditions, large asset portfolios, and the need to localize advanced technical capabilities. The review proposes a resilience-oriented architecture and staged deployment roadmap in which asset criticality, secure sensing, model assurance, personnel supervision, enterprise integration, and lifecycle governance are considered interdependent design requirements. It concludes that predictive maintenance aligned with Vision 2030 should be evaluated in terms of operational reliability, maintenance productivity, cyber resilience, workforce capability, and measurable business outcomes rather than predictive accuracy alone.
Purpose: Non-invasive monitoring of blood glucose and cholesterol has become increasingly important due to the growing prevalence of diabetes and cardiovascular diseases and the limitations of conventional invasive measurement methods, which require blood sampling, cause discomfort, and increase the risk of infection. Although various non-invasive techniques have been proposed, most existing studies focus on a single biomarker or a single sensing modality, limiting their robustness and practical implementation for real-time health monitoring. This study aims to develop an Internet of Things (IoT)-based non-invasive monitoring system capable of simultaneously estimating blood glucose and cholesterol levels by integrating physiological multisensors with machine learning. The main contribution of this research is the development of a portable platform that combines Near-Infrared (NIR) sensors, photoplethysmography (PPG), and skin temperature sensing with a Random Forest regression model for simultaneous glucose and cholesterol prediction. Methods: The research method used is the experimental research and development (R&D) method with the ADDIE model (Analysis, Design, Development, Implementation and Evaluation). The system developed uses a combination of multi-wavelength near-infrared (NIR), photoplethysmography (PPG) and skin-temperature sensors and a Random Forest Regressor to estimate blood glucose and cholesterol levels. Sensor data were pre-processed by missing value handling, outlier removal using Interquartile Range (IQR) method and Min-Max normalisation. Hyper parameters of Random Forest were optimised using GridSearchCV and model performance was evaluated using R² , MAE and MSE. The optimized models were subsequently integrated into the non-invasive monitoring prototype and evaluated in terms of prediction error, signal stability, response time, and user perception. Result: In the experiment, the optimized Random Forest model achieved R² of 0.4515, MAE of 0.1386 and MSE of 0.0299 in glucose prediction, and R² of 0.4703, MAE of 0.0804 and MSE of 0.0150 in cholesterol prediction. The experimental results showed an average prediction error of ±8 mg/dL for glucose and ±10 mg/dL for cholesterol at system level. The standard deviation of the sensor signals was maintained at 0.03-0.05 and the average system response time was 2.8s. Moreover, the user evaluation has provided a usability score of 4.75 out of 5 which is a high level of usability of the system. The feature importance analysis of the optimized Random Forest model showed that age, nir680 and nir610 were the most important predictors for glucose estimation, while nir680, nir860 and nir610 were the most important predictors for cholesterol estimation. The results are consistent with the multisensor design of the proposed system, where PPG and NIR signals provide complementary physiological information for non-invasive estimation. The combination of physiological multisensor data with an optimized Random Forest regression model, supported by technical testing, fast system response, stable sensor signals, and high usability, looks very promising for the simultaneous estimation of glucose and cholesterol levels without touching the person and for the development of portable, real-time, IoT-enabled personalized healthcare systems. Novelty: The novelty of this research lies in the development of a portable IoT-based device that integrates physiological multisensors and machine learning for simultaneous and non-invasive glucose and cholesterol estimation with better interpretability.
Purpose: This research aims to develop a highly responsive Indoor Positioning System (IPS) to locate lecturers in real-time within educational facilities by mitigating Access Point (AP) Churn through decentralized data preprocessing. Methods: A Wi-Fi fingerprinting dataset comprising 1,428 observations from 40 Access Points (APs) was collected across lecturer operational areas. Missing RSSI values were standardized to −100 dBm on the Android client before transmission. Random Forest, Decision Tree, Support Vector Machine (SVM), and K-Nearest Neighbor (K-NN) were comparatively evaluated using stratified cross-validation and validated through real-time Android deployment. Result: The Random Forest algorithm achieved the best performance with 97.90% accuracy and 97.86% F1-score. The proposed Android-based workflow enabled reliable real-time lecturer localization while maintaining consistent RSSI preprocessing between the mobile application and the cloud inference model. Novelty: The contribution of this study is an Android-based edge preprocessing framework that standardizes missing RSSI values before cloud inference and demonstrates its practicality through comparative machine learning evaluation in a real-time lecturer localization system.
High relative humidity (RH>60%) significantly impairs the accuracy of low-cost optical particulate matter (PM) sensors due to hygroscopic aerosol growth and water droplet interference. Physical heating elements evaporate condensed water but introduce substantial power consumption that limits low-power IoT operation. This study evaluates a software-based relative humidity compensation approach designed to mathematically emulate the output of a physically heated optical sensor using an unheated sensor paired with ambient meteorological parameters (T, RH, P) across 48,860 10-min records in Upper Silesia, Poland. We benchmarked six machine learning algorithms, Polynomial Regression (PR), Decision Trees (DTs), Support Vector Regressor (SVR), Dense Neural Networks (DNNs), and Long Short-Term Memory (LSTM) networks, against a co-located, heated SDS011 target sensor across both predictive accuracy and full computational profiling (tuning time, training duration, and inference latency). Static non-linear models achieved superior performance, with sixth-degree PR, SVR and DNN reaching R2=94%, MAE=1.41–1.61μg/m3, and RMSE=3.69–3.79μg/m3 for PM2.5, whereas complex LSTM networks performed substantially worse (R2=88%, MAE=2.56μg/m3, RMSE=5.33μg/m3). Computational profiling revealed that static models require negligible inference latencies (0.0002–0.0011ms/sample) and low tuning times (0.004s/fit for Decision Trees), whereas LSTM incurs an inference delay of 0.1215ms/sample (>500× slower) and heavy tuning overhead (148.96s/fit). These findings demonstrate that instantaneous non-linear relationships dominate humidity artifacts, confirming that lightweight software compensation can emulate the effect of hardware heaters on resource-constrained microcontrollers without deep learning overhead.
Mammalian neurons exhibit diverse structural and physiological properties: a heterogeneity not present in current neuromorphic systems. Recent advancements have led to bio-realistic spiking systems which are based on the integration and long-term operation of organic electrochemical transistors (OECTs) with anti-ambipolar characteristics. However, the library of materials enabling anti-ambipolarity (AA) is restricted by the lack of stable, fast, and scalable organic mixed ionic and electronic conductors (OMIECs) exhibiting this specific OFF-ON-OFF behavior. This complicates mimicking specialized neuron types. Blending semiconducting with insulating polymers is a classic organic and OMIEC electronics strategy to improve device performance and enable tunability, yet unexplored for improving anti-ambipolar stability in non-ladder-type anti-ambipolar materials. Here, the n-type OMIEC, P(gNDI-gT2), is blended with the insulating polymer, PMMA, yielding stable anti-ambipolar characteristics. The miscibility in solution forms a mechanically stabilizing and ion-permeable framework in the solid state that considerably reduces swelling-induced structural damage, thus preventing dramatic failure after the first operation cycles without compromising the conductive polymer's electronic pathways. Modulating the polymer blend ratio enables low-power, fast-spiking Hodgkin-Huxley neurons with tunable and programmable dynamics, without optimized device geometries, demonstrating frequency deconvolution and reservoir computing. These two computing primitive building blocks are essential for designing efficient spiking neural networks.
Quantum algorithms for scientific computing are specified against access models such as block-encodings, sparse oracles, state preparations, and quantum random-access memories. Their concrete realization is deliberately left open, yet it is a research variable that changes both the numerical behavior and the resource cost of the program, so every reported conclusion is attributable to one realization. In this paper, we present oracq, a Python framework that makes this choice a managed and checkable decision. Our claim is linguistic rather than infrastructural. Algorithms are stored as persistent register-level programs whose oracle slots are declared and typed but unimplemented. A separate binding pass links compatible realizations, so one stored algorithm can be re-bound to a gate network, a QRAM-backed table, or compiled reversible arithmetic and compared across them. Typed QRAM resources, a compiler from pure Python numerical formulas to reversible fixed-point circuits, and block-encoding algebra with qubitization and QSVT-style kernels supply the constructions. Two end-to-end workflows exercise the framework, a quantum finite-volume pipeline for the compressible Euler equations and a quantum homotopy-analysis pipeline that lifts nonlinear PDEs to composable linear input models. The verification methodology separates implementation error from method error, and logical resource accounts cover programs before and after binding. The framework is openly available at https://github.com/IAI-USTC-Quantum/oracq.
Abstract This paper investigates the computational relevance of the Energy-Conserving Sampling and Weighting (ECSW) method for hyper-reduction in reduced-order models (ROMs), with particular attention to modern Graphical Processing Unit (GPU)-accelerated and deep-learning-based frameworks. While hyper-reduction techniques such as ECSW are traditionally introduced to enable tractable evaluation of nonlinear terms in Galerkin ROMs, their actual benefit for computable nonlinearities with traditional Proper Orthogonal Decomposition (POD)–Galerkin ROMs combined with deep learning and on deep-learning GPU hardware remains a question. We provide a full ECSW implementation for finite element Navier–Stokes solvers, compatible with automatic differentiation and batched tensor operations, and derive exact operation counts for both standard tensorial and hyper-reduced formulations. Theoretical analysis shows that the computational complexity scales as O r 3 $ \mathcal{O}\left({r}^3\right) $ Or 3 for tensorial ROMs and O rm $ \mathcal{O}(rm) $ Orm for ECSW, where r $ r $ r is the reduced dimension and m $ m $ m the number of retained quadrature points. Extensive CPU and GPU benchmarks are conducted on the fluidic pinball configuration to assess asymptotic scaling, memory effects, and hardware bottlenecks. Results confirm the linear complexity and efficiency of ECSW on CPUs but reveal that, on GPUs, the regular tensorial formulation can outperform hyper-reduction for common reduction dimensions due to highly optimized batched matrix kernels and limited bandwidth constraints. Overall, the study clarifies when hyper-reduction remains advantageous in practice and provides concrete guidelines for integrating ECSW within deep-learning-augmented ROM frameworks.
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.
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.
Looping re-applies a block of layers so that each token passes through the block more than once. Adding a loop to a trained model is attractive because it leaves the parameter count, and hence the disk size, the memory footprint and the training cost unchanged. Unfortunately, every generated token pays for the loop. Re-applying m of N layers makes each token about (N+m)/N times as expensive. Loops that a model learns during large-scale pretraining justify this price, but whether loops added after pretraining justify this cost remains unclear. We compare a looped model with two simple baselines. We first compare it with the same model trained identically without the loop (equal parameters). Secondly, we compare it with the best unlooped use of the same inference floating-point operations (equal FLOPs). Across Qwen3 and Qwen3.5 models from 0.8B to 9B parameters, only the looped 0.8B model beats both baselines. At 4B and 9B, the looped model does not beat the first baseline, i.e. the same model trained without the loop. This is not because the benchmarks are too easy, since on the same benchmarks the unlooped 9B model still scores up to 13 points higher than the 4B model, showing that there is ample room for improvement. Introducing loops in an already pretrained model therefore only pays for small models with a fixed parameter budget.
The accelerating demand for ultrafast, scalable, energy-efficient and miniaturized computing has exposed the physical limits of conventional CMOS electronics, prompting the transition toward optical and plasmonic architectures. Existing metallic (plasmonic) nanoparticles primarily gold and silver suffer from high losses, thermal instability, and incompatibility with CMOS fabrication, constraining large scale integration. To address this gap, we propose a dual band, CMOS compatible material platform, Insulator-Metal-Insulator (IMI) nanostructure leveraging alternative plasmonic materials – Indium Tin Oxide (ITO) and the dielectric Silicon Nitride (Si3N4) for ultrafast all-optical logic gate (AOLG) realization. The designed hybrid IMI waveguide-nanoring resonator demonstrates high transmission efficiency of 86.15% (λ = 1173.8 nm) & 80.69% (λ = 2054.7 nm) at two optical bands and low insertion losses. The compact (400 × 400 nm2) footprint supports all seven fundamental logic operations (NOT, AND, OR, NAND, NOR, XOR, XNOR) within a single architecture. Quantitative analysis reveals an extinction ratio exceeding 18 dB and ultrafast femtosecond response times, demonstrating competitive simulated optical performance subject to experimental validation. This CMOS compatible dual band architecture demonstrates a promising route toward energy efficient photonic integrated circuits (PICs) and opto-electronic integrated circuits (OEICs), enabling the next generation of AI driven, neuromorphic, and high-speed optical computing platforms.
Quantifying correlations is important for understanding how quantum states and channels support information processing. Quantum umlaut information $U$ connects such correlations to reliability in hypothesis testing and certain communication settings, but its known formulas require classical descriptions of the underlying states. To estimate correlations without these descriptions, we introduce measured quantum umlaut information $U^M$ by replacing Umegaki relative entropy with measured relative entropy. Its variational formulation leads to a quantum generative adversarial framework for estimating $U^M$ from measurement statistics of unknown bipartite states and classical-quantum channels. This framework optimizes reference states and measurements without reconstructing the target states or channel outputs. For classical-quantum channels, it also optimizes the input probabilities. We further analyze the relation between $U^M$ and $U$, deriving a state-dependent gap bound and characterizing equality for full-rank bipartite states. Numerical simulations show trained estimates close on average to independently computed numerical references for $U^M$ across the tested states and channels. These estimates could help assess how well quantum memories preserve correlations and how noise or drift affects input-output dependence in quantum transmitters and state-preparation devices.