photonics-computing
LLVM for the memristor and photonics compluting complier can do simulation and output the physically controlling paramters for the computing on next genenration hardware
5 repos · 35 papers · 62 scholar
Showing 102 of 102 items
LLVM for the memristor and photonics compluting complier can do simulation and output the physically controlling paramters for the computing on next genenration hardware
Developed and trained the differentiable logic-gate spiking neural network (Logic-SNN) framework for a black-phosphorus-based photonic neuromorphic computing architecture, integrating convolutional and dense logic layers with Leaky Integrate-and-Fire (LIF) neurons.
Open research laboratory for photonic AI accelerators, Mach–Zehnder interferometers, integrated photonic chip design, GDSFactory, LNOI, Meep FDTD simulations, and energy-efficient optical computing.
With the growing interest in using AI (Artificial Intelligence) for RTL (Register-Transfer Level) hardware development, robust and comprehensive verification has become more important than ever. As Large Language Models (LLMs) increasingly assist in creating and modifying RTL designs, ensuring that these changes preserve existing functionality is paramount for building trust in AI-driven workflows. We present SVApshot, a fully automated framework that leverages LLMs to generate and systematically attempt to correct SystemVerilog Assertions (SVA) for RTL modules, completely removing the human from the generation loop. SVApshot introduces a novel snapshot methodology that captures the current functionality of a design as a comprehensive set of formal assertions, creating a regression suite for validating future changes--whether made by humans or AI. The framework features automated duplicate detection, assertion set expansion, and iterative repair of failing assertions using LLM-guided debugging with formal verification feedback. Our experimental evaluation across diverse RTL modules demonstrates high coverage for complex modules, with successful detection of both manually-injected and AI-introduced bugs. Additionally, this framework opens the door for automatic generation of formal testbenches to benchmark LLMs for hardware design, moving beyond binary pass/fail metrics towards nuanced assertion-based scoring.
MCU-based embedded and cyber-physical systems increasingly require protection policies to change at runtime as resources, software, and system composition evolve. Yet widely used COTS MCU protection mechanisms are still largely designed around statically established policies. While they can restrict which domains modify protection primitives, they generally do not constrain what policy an authorized domain may create. As a result, domains may expand protection rules beyond their intended scope or interfere with boundaries established by other domains. We identify bounded configuration authority as a missing primitive for secure dynamic policy evolution and introduce an ownership-aware model in which explicit ownership and rule containment bound both configuration authority and recursive delegation. Changes to active protection-rule ranges are validated before commit, preserving protection boundaries without extending the normal access-checking datapath. We implement the model in an open-source RISC-V World Checker integrated into a CVA6-based SoC. Our FPGA evaluation shows no degradation in access latency or operating frequency, with 7.1% LUT and 1.2% FF overhead for eight rules. In an OTA update case study, direct reconfiguration under delegated authority reduces policy-transition latency by 89.5% over hypervisor mediation and 94.6% over trusted firmware mediation, while preserving established authority boundaries. All artifacts will be open-sourced to support the evolution of RISC-V specifications and system-level protection designs.
Non-linear functions are an important class of operations in modern large language models. However, chip vendors building for AI have so far relied on traditional vector or scalar processing units for non-linear functions. The performance of these units has not kept pace with matrix multiplication units (MMUs), which can be found in both high-performance and edge-computing devices from almost every vendor. Reconfigurable units such as FPGAs and CGRAs have been proposed in the literature, but have not been systematically compared with traditional units backed by MMUs. In this paper, we build two SoCs: one using an RVV vector processor and the other using a latency-insensitive (\emph{elastic}) CGRA, both coupled with an MMU and driven by a single-core CPU. We then propose mappings for common feed-forward networks on our SoCs and find that the CGRA reaches a maximum speedup in clock cycles of 7.12x over RVV on the activation stage of simple point-wise non-linear functions like SiLU, whereas the RVV reaches a maximum speedup of \textbf{9.8x} over the CGRA on complex non-linear functions such as $\mathrm{softmax}$. Over the enclosing feed-forward block the corresponding maxima are 3.16x and 4.24x.
Solving large sparse linear systems by Gaussian elimination is a challenging computational task due to the introduction of non-zero elements, or fill-in. Practitioners use heuristics such as the nested dissection reordering algorithm to try to reduce fill-in. In practice, reordering codes do not achieve the same parallel scaling as factorization, nor can they exploit AVX instructions or GPUs. Thus, reordering can be the primary runtime bottleneck in large-scale computer-aided engineering applications. We consider the use of a novel accelerator, the UpDown system, to accelerate the METIS multilevel nested dissection reordering algorithm. The UpDown system is designed for a broad class of graph computing and irregular algorithms. We report results for small-scale UpDown systems that will be available commercially in 2027, using validated simulation and FPGA emulation platforms. Our findings indicate that small UpDown systems can speed up METIS by two-orders of magnitude versus a single CPU. Compared to ParMETIS on hundreds of MPI ranks, we anticipate the same small UpDown systems offer speedups of 4 - 7, providing the potential to significantly speed finite element simulations
Large language models have improved Verilog generation, yet functional correctness remains difficult to optimize because simulation commonly provides a coarse module-level pass/fail outcome. Such feedback discards useful evidence from partially correct RTL modules and cannot determine whether a source statement is functionally responsible for an observed output behavior. We present a new method named Causal Signal Credit Assignment (CSCA), which is a counterfactual preference-optimization framework for Verilog generation. CSCA first performs continuous signal-level behavioral verification on sampled candidate modules. It then locates output-driving assignment spans and applies five interventions while holding the remaining module and testbench fixed. The resulting changes in signal-level errors are combined with behavioral agreement to form a continuous statement credit, which is mapped to response-token weights for strict weighted direct preference optimization. This enables CSCA to learn from locally supported statements without uniformly reinforcing an otherwise incorrect response. Built on Icarus Verilog and a CodeV-derived corpus, CSCA is evaluated on VerilogEval~2.0 and RTLLM~v2.0, including both code completion and specification-to-RTL tasks. With a Qwen2.5-Coder-7B-Instruct backbone, CSCA leads the evaluated 7B-scale models and remains comparable to the much larger GPT-5.4-mini on the reported Verilog-generation benchmarks.
Emerging memory modules of large language models, such as Engram in DeepSeek-V4.1-Flash, mix irregular integer work (a 64-bit n-gram hash and a sparse table lookup) with dense computation (a GEMM and a gate). When deploying on a GPU-FPGA system, we can disaggregate operations to different devices for Engram to speed up by 5.9$\times$. AWS Trainium provides a similar heterogeneity natively: each Trainium NeuronCore pairs dense engines with a customizable GPSIMD Engine, and all engines share the on-chip SBUF. However, the Neuron Kernel Interface (NKI) exposes the GPSIMD cores only through fixed built-in instructions, and the official GPSIMD custom-operator toolchain does not support Trainium 2 or 3 and cannot be used with NKI. In this work, we build a toolchain that compiles C++ custom operators for the GPSIMD cores and injects them into the compiled NEFF from the NKI kernel on Trainium 3, expanding the customizability of GPSIMD with the current software stack. Over a hand-optimized NKI baseline of DeepSeek-V4.1-Flash, two GPSIMD operators speed up the n-gram hashing by 1.79$\times$ at decode, the table retrieval by up to 4.95$\times$, and the whole Engram computation by up to 1.14$\times$.
Many digital systems courses teach on FPGA boards whose vendor tool is no longer maintained: the Digilent Basys2 and Nexys2 (Xilinx Spartan-3E) and Nexys3 and Atlys (Spartan-6) need Xilinx ISE 14.7, discontinued since 2013 and hard to install on the computers and tablets students use today. The author presents Silinx, a free (AGPL-3.0) HDL design platform for teaching that keeps ISE's Project Navigator workflow but runs in a web browser with nothing to install. A module can be written in VHDL or Verilog, in C or C++ through Silinx HLS, which is compatible with AMD's Vitis HLS and passes all 73 of its introductory examples, or drawn as a schematic, an ASM chart, a state diagram, a Boolean function with its Karnaugh map, or a Petri net, from which its HDL is generated. Designs are checked with explanations written for beginners, simulated, run on an emulated board, built, and programmed onto a real board over WebUSB. The implementation flow is open and runs in the browser: Yosys and nextpnr compiled to WebAssembly, with Silinx's own device data, timing model and bitstream writer for the Spartan-3E and Spartan-6, built by black-box observation of ISE's documented tools; Lattice iCE40 boards use the existing open iCE40 tools, and further families are planned. ISE is the oracle: Silinx's bitstreams are byte-identical to those ISE writes for the same placed and routed design, and pass ISE's design rule check. Against ISE at the same optimization goal, the open flow uses 15% fewer slices with 14% shorter critical paths when optimizing for area, and 21% fewer slices with 8% shorter critical paths when optimizing for speed (geometric means over 45 open-source designs built by both flows). The author also describes how the platform was developed with AI coding agents under test-first project rules.
Training Gaussian processes (GPs) at large scale is computationally challenging because evaluating the log-marginal likelihood and its gradients involves large, dense kernel matrices. We develop a scalable low-rank GP training approach that combines AnchorNet landmark selection with a fully independent training conditional (FITC) correction. The resulting objective and its analytic gradients can be evaluated using low-rank factors, diagonal quantities, and small dense matrices, without forming the full kernel matrix or its derivatives. The complete training procedure is implemented in single precision and evaluated on nine large-scale regression datasets. Across these problems, the method reaches the reference predictive accuracy while substantially reducing training time, although the required approximation rank varies across datasets. We also study the numerical behavior of the full-kernel predictive solve and show that single precision can become limiting for ill-conditioned problems. Finally, we investigate further reducing selected matrix operations to fp16 or TF32. These lower-precision variants provide only modest additional speedups and can become unreliable when the reduced system is poorly conditioned.
Training a neural network under homomorphic encryption requires transposes, batch reductions, optimizer updates, and repeated refreshes in addition to forward matrix products. We combine rectangular diagonal packing, feature-major SIMD mini-batch lanes, baby-step giant-step evaluation, and compact optimizer state in an MIT-licensed Rust CKKS implementation. Its CUDA path keeps training and bootstrap ciphertext arithmetic on the GPU, while retaining host key generation and client operations. In a matched 100-update experiment on the main non-interactive MNIST model, device-resident bootstrapping reduces training time from 14,268.295 to 293.075 seconds, a $48.68\times$ ratio against our CPU implementation, with exactly equal decoded parameters and predictions. Including setup and final evaluation gives a $40.67\times$ ratio. A separate GPU run completes ten passes over 4,800 images in 2,967.095 seconds and achieves 90.35% test accuracy versus 90.41% for its plaintext twin. These are single-initialization results; exact CPU/GPU parameter agreement is verified at 100 updates, not for the full 1,000-update run. Earlier completed CPU runs support a larger-batch key-switch advantage and baseline-batch accuracy within the registered tolerance, but the baseline-batch normalized key-switch target is missed. The released ReBoot baseline does not meet the common security-estimator threshold, and its full-setting CPU time is unmeasured. We therefore distinguish measured acceleration of our implementation from conditional layout comparisons and do not claim unconditional superiority over ReBoot.
We study how much the interaction strategy around a fixed, locally hosted 7B-parameter code model (Qwen2.5-Coder-7B-Instruct) affects its ability to replace PyTorch operators with custom CUDA kernels. On 30 KernelBench tasks (levels 1-3, first 10 each, evaluated on a V100), we compare four regimes of increasing structure: zero-shot prompting, profiling-guided prompting, iterative refinement on execution feedback, and a multi-agent loop (analysis, retrieval, generation, evaluation, feedback triage) in which the same model fills four roles around a non-LLM evaluator. Three findings stand out. (1) Correctness, not speed, is the binding constraint: even correct generated kernels average 0.97-0.99x the speed of eager PyTorch, because a 7B model cannot out-write cuBLAS/cuDNN on dense GEMM and convolution; the only real wins are fused memory-bound epilogues. (2) The best strategy depends on difficulty: profiling-guided prompting is strongest overall (43% correct, 27% faster than baseline), but the agent loop is the only method that solves any level-2 fusion problem (3/10 vs. 0/10 for all others). (3) Naive agentic scaffolding is actively harmful: our first agent scored 3% correct, below zero-shot's 30%, and the decisive repairs were matching the agent's ambition to the model's competence and spending compute on best-of-n sampling under a real execution oracle (3% to 33% correct from n=1 to n=8), not on more reasoning steps.
Sampling-based planners such as MPPI and CEM query a learned world model thousands of times per control step, so low-bit world models are an obvious way to make model-based control cheaper on a robot. With round-to-nearest activation quantization and no calibration, we find that planning can collapse, by up to 100 points of success, while the world model's latent one-step prediction error stays no larger than that of a stochastic rounding scheme that plans well. At the same bit width, what decides the outcome is not the size of the rounding error but whether it persists across planner iterations and is shared across candidates. Round-to-nearest leaves an error that persists, and an iterating planner drifts toward the candidates that this error favors; in TD-MPC2, holding the same dither fixed within a step or an episode brings the drift back. Dither shared by all candidates and redrawn at every iteration removes most of the damage at the same bit width and without training: at least half of it in all 19 affected Meta-World cells (task $\times$ bit width) of TD-MPC2 (median 95%), in 4 of 5 DINO-WM cells, and in all 8 affected tasks of a 317M-parameter multi-task TD-MPC2. Captured as CUDA graphs on an Ampere GPU, a W8A4 integer TD-MPC2 world model runs its planning rollouts 1.9-3.3$\times$ and the whole planning step 1.3-1.5$\times$ faster than FP32, though not faster than BF16 with pre-converted weights; shared dither adds at most 2.3% latency. In TD-MPC2, block-scaled NVFP4 removes most of the damage and MXFP4 part of it, which ties the failure to coarse grids with a shared scale. Hypotheses, thresholds and analysis code were fixed before the data that test them, and the hypotheses that failed are reported.
Deploying large language models in regulated industries requires more than picking the fastest inference engine: it requires evidence that a configuration meets contractual and legal obligations before it serves a single customer request. We describe Runway, an internal validation harness built at MasterControl that builds a governed decision pipeline around NVIDIA AIPerf. The key insight is separating concerns: AIPerf provides the trusted measurement, while Runway provides the experiment design that comes before it and the compliance, cost, and audit decisions that come after. We describe the architecture and the storage model that makes caching and audit one mechanism. We report three findings from Runway sweeps: TensorRT-LLM and vLLM rankings invert with concurrency; a newer, pricier GPU instance can be the cheaper one per request; and speculative decoding behaves differently across engines under load, so engine configuration can outweigh engine choice.
This paper presents a GPS-triggered geofencing and enforcement scheme that integrates real-time satellite positioning with Software-Defined Radio (SDR) based RF jamming to detect and neutralize unauthorized unmanned aerial vehicle (UAV) boundary violations. Conventional geofencing depends on onboard GPS and cooperative flight firmware. The proposed approach instead moves enforcement to a ground base station, so the response does not depend on the UAV. An ESP32- WROOM-32 microcontroller paired with a u-blox NEO-6M GPS receiver streams telemetry over Wi-Fi to a Python/GNU Radio base station. There, a lightweight Haversine boundary-violation algorithm computes the great-circle distance between the UAV and the geofence. This distance drives a piecewise-linear adaptive speed and altitude control law. The law progressively decelerates and descends the UAV inside a 1 km warning zone and brings it to a complete halt at the boundary. If the boundary is breached, a HackRF One SDR escalates to band-limited noise jamming of the 2.4 GHz control link. The hardware stack was integrated on a GoolRC S116 quadrotor with an end-to-end enforcement latency of about 1 s. Field trials achieved 95% disconnection accuracy within a 50 m radius, and u-center NMEA analysis confirmed sub-2-meter GPS positioning precision. Simulation over an 8-sector, 50 km geofence model confirms speed and altitude-invariant boundary enforcement for cruising speeds of 5 to 50 m/s and cruising altitudes up to 100 ft. A comparison against a PID baseline shows that the proposed law guarantees overshoot-free shutdown at the boundary without gain tuning. It also runs in constant time, which makes it well suited to low-cost, resource-constrained embedded platforms.
Connected physical systems fuse observations from robots, IoT sensors, and edge-AI modules into a shared account of the world, yet most evaluations score only the final answer or action and cannot tell whether a method attributed the evidence, updated its belief, handled contradictions, and justified its decision. We argue that shared reasoning in Connected Robotics should be evaluated as an evidence-to-belief trajectory, and introduce the Connected Robotics Evidence Graph (CREG) Benchmark to make that trajectory measurable. CREG comprises 350 deterministically generated scenarios across seven regimes (justified revision, belief preservation, relation association, contradiction onset and resolution, stale/delayed evidence, and source reliability), each exposing a normative posterior reference at every evidence prefix, validated by an independent log-odds checker and screened by a blinded expert audit. Nine representation conditions separate relation content from serialization, ordering, and metadata. Evaluating four open-weight instruction models and six transparent fusion baselines yields three findings. First, entity grounding accounts for a substantial part of the weak-model deficit. Second, deleting semantic relations costs 15.3 points of strict belief-state accuracy on relation-dependent scenarios (Holm-corrected p<0.001), but an information-matched structured list outperforms the full graph and shuffling the graph does not hurt: relation content matters where required, while graph serialization confers no consistent advantage. Third, belief accuracy is strongly non-monotonic along the evidence sequence, with mid-trajectory failures that endpoint-only evaluation cannot detect. All scenarios, references, prompts, metrics, and analysis code are released for offline, hardware-free reproduction.
Self-recurrence represents an important component for temporal processing in neural networks, both artificial and biological, where it appears naturally in the form of leaky integration. However, due to the complex temporal dependencies of the gradient, an efficient and effective method of online training is still lacking. Here we propose K-prop, a novel online learning method for leaky neural networks. It dramatically reduces the computational and memory cost of real time recurrent learning (RTRL) by propagating kernel gains. K-prop also outperforms existing RTRL approximations and achieves performance comparable to backpropagation through time in sequence learning, classification, and reinforcement learning tasks. We further analyze how its sole approximation affects gradient quality and identify conditions under which K-prop remains accurate and effective.
Lookup-table (LUT)-based neural networks (NNs) offer a compelling path to ultra-low-latency inference on FPGAs. However, existing LUT-based neural networks have largely been restricted to MLP-style topologies, leaving structured vision and audio tasks underexplored. We present ConvLUT, a family of fully unrolled LUT-mappable convolutional architectures targeting the MLPerf Tiny benchmarks. It is comprised of a standard convolutional model built around a hierarchical tree of LUTs and a depthwise separable model following the MobileNet design philosophy. These architectures are enabled by four key contributions: a unified LUT-based convolutional framework supporting standard, depthwise, and pointwise convolutions under explicit fan-in constraints; a LUT-tree that addresses the $k^2c$ receptive field dimensionality bottleneck; a channel-partitioned LUT pooling layer that replaces conventional spatial pooling; and a hardware-aware iterative pruning strategy for stable convergence at scale. Evaluated on CIFAR-10, Visual Wake Words, and Keyword Spotting, ConvLUT achieves competitive accuracy with fully unrolled inference latencies below $\sim 50$ ns. On CIFAR-10 it reaches at least a $8844\times$ latency reduction over non-logic-based prior implementations, and outperforms existing MLP-style LUT-based models on the same task by at least $20$ pp in test accuracy.
Face-to-face (F2F) hybrid bonds shorten interdie connections, but data, clock, and power-grid (PG) connections must be realized together. CU-3D is a RTL-to-GDS physical design flow based on commercial EDA tools. It derives role-compatible hybrid-bond (HB) sites from a 17-metal ASAP7-based technology model, complete HB footprints, and constructed PG geometry. To coordinate partitioning with placement flexibility, Scaled refines an inherited 2D placement, whereas Fresh places the receiving die first and uses its pin locations and clock loads to guide placement of the other die. Data and clock HB assignments are jointly optimized while retaining alternative data sites and reserving capacity for subsequent PG assignment. The selected HB sites and stacks are preserved during 3D clock-tree synthesis, supply-aware cell insertion, and routing. Across eleven designs, including a 133-macro processor, the flow achieves zero backend DRC violations and passes signal-connectivity checks. Nine paired designs admit lower-footprint, lower-power implementations; four improve setup WNS. On JPEG, early joint planning gains 87 ps and mobility-aware partitioning gains 346 ps. Path analysis explains why cell-delay savings can outweigh net-delay increases.
In expert-parallel Mixture-of-Experts serving, a layer finishes only when the most loaded GPU of its group does. Token routing shifts at every decode step and concentrates on a few experts, so most of the group's reserved GPU time waits on stragglers. Recovering that time means re-deciding, at every layer of every decode step, how each replicated expert's tokens divide between its copies. The exact linear-program solvers that make this decision optimally take longer than many of the layers they serve, and their working set outgrows GPU shared memory at deployment scale. We introduce Relief, a bidirectional pairwise relaxation that makes the linear program unnecessary, coming within a fraction of a point of the per-step optimum in a few passes of integer halving over the token counts. We build and measure three deployments, CUDA kernels on the serving GPU, the same verified engine on an FPGA over PCIe, and a fixed-function block beside the accelerator. On decode routing from three open MoE families, Relief holds the balance near the per-step optimum at every group size and layer time measured. Scored as effective utilization relative to that optimum, the kernel alone matches or exceeds every deployable scheme at every group size, from the hundred-microsecond layer on. The on-die engine widens that lead at every point, recovering up to 37.3 percentage points over the deployed static placement, while adding 0.02% of a current accelerator's transistors.
This paper introduces a budget-aware, multi-objective framework (called PRISM) for RTL optimization that improves power, performance, and area (PPA) while sharply limiting expensive commercial synthesis tool calls. Starting from a synthesizable baseline and functional testbench, the method iterates through LLM-driven rewrite generation, fast compile/simulation equivalence filtering, low-cost proxy feature extraction, and selective commercial synthesis tool (Synopsys Design Compiler (DC)) evaluation. To prioritize which candidates deserve high-quality synthesis, we train a preference-based reward model using pairwise ranking supervision derived from commercial synthesis tool outcomes, and combine ensemble uncertainty with a diversity-aware acquisition rule to avoid early bias and redundant evaluations. A key advantage is that PRISM maintains a best-tradeoff set G of non-dominated designs throughout the run: G preserves the best solutions found so far, is fed back into generation and refinement, and enables consistent, cumulative improvements rather than repeatedly restarting from the baseline. Across RTLLM v2.0 modules, the approach consistently yields post-synthesis gains under fixed commercial synthesis tool call budgets, and ablations confirm that reward-guided selection and candidate dropping are central to sample-efficient progress. PRISM efficacy is evaluated across multiple technology nodes (7 nm, 15 nm, and 45 nm) using identical pipelines without retraining.
Language-model agents increasingly run on workstations with a few consumer GPUs, where large mixture-of-experts (MoE) models keep part of their experts in host memory. With existing CPU-GPU engines, these agents get fast prefill or fast decode, not both: llama.cpp's expert cache skips forward passes of more than 32 tokens to keep the experts of decode cached, so an agent waits 2.1-2.8 s after each tool output, whereas its default layer placement shortens this wait but is 1.5-1.6$\times$ slower in decode. Our key insight is that a short prefill touches too many distinct experts to upload, yet sends 84% of its expert activations (routes) to a stable set of experts that fits in GPU memory. We present Ostinato, a CPU-GPU MoE inference engine that computes these experts on the GPUs in every forward pass without changing what the model computes. Short prefills split their routes, leaving fewer than one of a token's six routed experts per layer to the CPU; long prefills upload each distinct cold expert once; and each decode step runs as one fixed-shape CUDA graph per GPU. On real coding and customer-service agent sessions with DeepSeek-V4-Flash on four RTX 4090D GPUs, Ostinato is the fastest engine in both phases and finishes every session sooner than all six baselines. Against llama.cpp and KTransformers, it cuts the median wait after a tool output by 22-55% and finishes sessions 1.4-2.5$\times$ sooner.
This paper presents a unified transformer-based framework for object detection and dense region captioning across distributed edge networks. The architecture uses a shared visual encoder followed by detection, region extraction, and autoregressive language-decoding stages, allowing localization and region description to reuse the same image features. This design removes the second image-encoding path required by decoupled detection and captioning pipelines. To support execution on resource-limited devices, a workload-allocation formulation assigns model stages and input streams across connected edge nodes while accounting for execution time, memory capacity, processor utilization, power limits, and intermediate tensor transfer cost. The framework is evaluated on a six-device edge testbed comprising four NVIDIA Jetson Orin Nano and Orin NX inference nodes and two Raspberry Pi 5 auxiliary CPU nodes, with TensorRT FP16 and INT8 execution on the NVIDIA devices. On a Jetson Orin Nano, the unified INT8 model achieves 38.1 FPS with 3.5 GB peak memory use and 11.8 W measured power. Under four concurrent streams, the proposed allocation achieves 35.2 FPS and 28.4 ms mean end-to-end latency with an internode transfer rate of 6.8 MB/s.
This paper presents a unified framework for license plate detection and character recognition on low-power edge hardware. The proposed network uses a shared CSP-Darknet and PANet feature encoder with an anchor-free localization head and a RoIAlign-based sequence-recognition branch. Character decoding is performed directly on aligned feature maps, avoiding image-space plate cropping and a second recognition backbone. The model is trained on an AI-assisted dataset of approximately 37,000 manually verified images covering license plate layouts from all 50 United States. The dataset includes variations in illumination, viewpoint, weather, blur, occlusion, and plate design. The framework is converted to an INT8 TensorRT engine and evaluated on an NVIDIA Jetson Orin Nano with 8 GB memory. The INT8 model shows an mAP50 of 97.8% and a full-sequence accuracy of 90.7%. It records 147 FPS at 6.8 ms end-to-end latency, with 315 MB inference memory and 9.4 W average power consumption. Relative to the YOLOv8s+CRNN baseline, the proposed deployment provides a 7.1x throughput increase and a 77.8% reduction in allocated inference memory
Useful edge hardware can lose product support while retaining substantial computation. We treat this as systems design: compute regeneration here means re-productization, retaining the compute module while redesigning the surrounding product organization. Our case is GEdgeBot, a multilingual service robot built around a retired 32 GB Jetson AGX Xavier. A source-grounded inventory shows the box alone supplies inference, while the combined system exposes five on-device AI services, a seven-language speech chain, three sensing domains, and one deterministic seven-byte control boundary. AI and orchestration stay on the Xavier; finite motor-pulse generation on an STM32F407. Audits cover the six-mode control state machine, readiness barriers and sound orientation. Historical single-prototype measurements report 1,782 MB GPU memory with YOLO, Whisper-small, and NLLB-200 resident, a direct-control p99 deviation rising from 0.24 to 49.93 ms under concurrent AI load, and six of six trigger-to-follow demonstrations at 4.35 s mean latency, with conversational voice unavailable. A November campaign restores an isolated inference stack, re-measures AI workloads, and repeats control-frame runs. Driving two execution organisations with the same 6,000 measured 50 ms frame intervals shows the finite-pulse boundary removes host jitter from the pulse train but not the command start phase; its stopping bound is host-enforced, while the MCU 16-bit pulse field is unclamped. A six-check static audit exposes a host--MCU queue mismatch, and a 14-scenario replay executes the audited main() dispatcher across six modes with zero physical device opens and 54 contract-checked frames. Results support an existence claim and a reproducible architectural hypothesis, not fleet-level reliability. The 208 direct-control ratio remains unreplicated; MCU waveform capture and on-site fault injection remain open.
Cloud-native systems now extend to the network edge and host artificial-intelligence (AI) workloads, including AI agents, but evidence on their reliability and scalability is scattered and uneven. We report a PRISMA 2020 systematic mapping study with an evidence-graded synthesis. Five logged OpenAlex searches on 11 October 2026 covered cloud-native systems, edge AI, agentic AI for operations, reliability of agentic applications and physical AI; 189 records from an earlier exploratory search were added separately. After 301 duplicate removals and an automated peer-review venue rule, 4,611 records were screened against written criteria and a five-level relevance scale. The review includes 1,794 primary studies from 2017-2026 and 95 secondary studies, with every decision released in Dataset S1. Each primary study was assigned to a 13-category taxonomy and graded as simulation, prototype or testbed, production data, production deployment, or not determinable from the abstract. Of 1,003 studies with a determinable level, 66.2% used prototypes or testbeds, 18.0% simulations and 15.8% production data or deployments. A 93-study synthesis core, appraised from indexed abstracts, shows production deployments of cascading-failure prevention, trace- and graph-based root-cause analysis and predictive autoscaling, and production-trace evaluations of cold-start mitigation. Edge reliability and physical AI rely mainly on simulations and testbeds. AI-agent reliability uses tool-grounded diagnosis, explicit safety properties, trace-driven repair and workload-aware scheduling, while empirical studies show that agent frameworks remain fragile. Conclusions are limited to the documented corpus. The principal gaps are production evidence for edge-native systems, shared resilience benchmarks and dependable agentic operations.
Organizing control around events can reduce sensing and actuation demands while directly linking control decisions to the occurrence and timing of desired system behaviors. This paper proposes a novel event-centric model predictive control (MPC) framework that unifies event sensing, event prediction, and event regulation. Events define the feedback information, prediction horizon, and event-timing objectives, placing them at the center of predictive control design. The framework is developed for a scanning mirror system equipped with a baseline neuromorphic controller composed of spiking neurons that supports sparse actuation. The control objective is to achieve prescribed zero-crossing times and a desired scanning range while minimizing the mechanical energy transferred between the actuator and the mirror. A sampling-based algorithm is presented to obtain suboptimal solutions of the event-centric MPC optimization problem without requiring an explicit event-to-event prediction model. By grounding predictive optimization in events, the proposed framework opens a new direction for MPC design and its integration with neuromorphic control.
In edge intelligence systems, computation offloading and dynamic voltage and frequency scaling (DVFS) are widely used to improve energy efficiency. However, as applications like deep neural network (DNN) inference evolve in complexity, their demands for memory resources have progressively increased, establishing memory as a critical bottleneck that impacts both latency and energy consumption. In this paper, we propose a joint optimization scheme integrating memory, computing and offloading for energy-efficient DNN inference. We formulate the energy minimization problem under deadline constraints as a mixed-integer nonlinear programming (MINLP) problem, where the main challenge arises from the coupling between offloading and resource optimization. To address this issue, we decompose the original problem into two subproblems under a fixed offloading policy. The local subproblem is transformed into an equivalent geometric programming (GP) problem to derive the optimal memory frequency and computing frequency, while the edge subproblem is solved via convex optimization. An efficient coordinate descent (CD) algorithm is further developed to jointly optimize all variables. Simulation results demonstrate that our proposed algorithm achieves near-optimal performance with the gap from the optimal policy being at most 9.3%, and reduces the energy consumption of devices by up to 16.1% compared with benchmarks without memory frequency optimization.
Exploratory Data Analysis (EDA) aims to discover informative patterns in large datasets. For transactional data, Association Rule Mining (ARM) is an established approach that facilitates EDA by efficiently generating candidate patterns. Beyond efficient mining, identifying informative rules has also been extensively studied through both objective and subjective interestingness measures. Recent advances in Large Language Models (LLMs) offer new opportunities for assessing subjective interestingness using world knowledge. Nevertheless, relying solely on LLMs raises concerns about scalability and statistical reliability. Building on prior ARM research, we propose a hybrid framework, DiSARM, that leverages the world knowledge of LLMs to assess the subjective interestingness of rules, while maintaining statistical reliability and promoting diversity among the selected rules. Across three transactional datasets, experiments demonstrate that DiSARM effectively combines statistical reliability with surprise and diversity, offering a promising approach to ADE from transactional data.
High-level synthesis (HLS) can translate C/C++ kernels into register-transfer level (RTL) designs, yet generating high-quality hardware still relies heavily on manual pragma tuning and repeated synthesis evaluation. We propose AutoC2V, a C++-to-Verilog generation framework driven by real synthesis feedback that enables large language models (LLMs) to directly generate RTL with improved quality of results (QoR) from synthesizable C++ kernels. We construct training data from real synthesis results, including synthesizable C++ source code, the corresponding Verilog RTL, QoR metrics, and optimization objectives. Using this dataset, we align the model toward synthesis-aware RTL generation through Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO). To reduce the high cost of repeatedly invoking real synthesis during GRPO, we design the QoR Prediction Reward Model (QoR RM), which predicts latency and resource metrics by encoding RTL into long-context semantic representations, thereby providing low-cost and continuous QoR reward signals for policy optimization. Experimental results demonstrate that AutoC2V not only substantially improves RTL functional pass rates, but also achieves superior latency and resource optimization compared to the unfine-tuned base model, representative HLS optimization approaches (HLSPilot, LLM-DSE, and AutoDSE), and advanced general-purpose commercial models (Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, DeepSeek v4 Pro, and GLM-5.2). These results indicate that combining real synthesis data with efficient QoR reward modeling can further improve the ability of LLMs to generate high-quality RTL.
Advances in large language models and semantic IDs (SIDs) have brought generative retrieval (GR) into industrial applications. These systems must return items that satisfy request metadata constraints as well as relevance requirements. A catalog trie ensures SID validity, but filtering metadata only after generation wastes computation on rejected candidates and may leave too few eligible results. We propose RoutingTrie to incorporate metadata constraints before beam selection. RoutingTrie maintains one shared SID trie and builds a compact index for each metadata field using Dense Bitset or Sparse List. We derive a storage cost formula to guide the choice between these index representations for different field distributions. During decoding, field probes determine whether requested values occur below candidate edges and produce a combined mask that guides beam selection. Experiments on ShopTrie, our configurable synthetic benchmark, demonstrate compact storage and efficient CPU and GPU probing across diverse metadata distributions. In a production replay, RoutingTrie raises Hit@100 from 67.5\% to 96.1\%, a gain of 28.6 percentage points, with only 1.1 ms latency increase. We will release the ShopTrie generator, benchmark scripts, and RoutingTrie CPU and Triton GPU implementations to support reproducibility.
Reproducibility is important for LoRA-adapted language models used in repeated evaluation, auditing, and decision-making. However, Split-K kernels used to introduce parallelism in low-rank LoRA multiplications can have the unintended consequence of numerical nondeterminism when partial results are combined using floating-point atomic additions. We investigate this non-determinism in vLLM across GSM8K mathematical reasoning, MBPP code generation, and Jev-style structured toxicity decisions on Civil Comments and introduce a deterministic Triton Split-K implementation. Under identical-input replay, Atomic Split-K=8 produces different FP32 shrink outputs in 97.61% of comparisons, while our Deterministic Split-K=8 shows no observed variation. Repeated atomic inference changes generated outputs in 17.30% of GSM8K and 14.96% of MBPP comparisons, sometimes altering task correctness. Structured-decision probabilities vary in 81.47% of atomic comparisons, with decision flips. Controlled tracing demonstrates how shrink-output differences propagate into the KV cache and how accumulated cache differences can change token decisions and final answers. Our deterministic Triton Split-K=8 implementation preserves parallelism, achieving 16.8-21.1% latency-based speedup over unsplit reduction with only 1.02-2.62% additional latency relative to Atomic Split-K=8.
Electric vehicle (EV) charging stations are increasingly operated with an aggregate power limit below the sum of their port ratings, and require scheduling algorithms to divide the available power among the connected EVs. Charging standards permit a current that is either zero or at least a minimum rate, typically 6 A, which makes this allocation problem non-convex. Existing methods either solve a mixed-integer program with a commercial solver, which can be computationally expensive, adds software dependencies, and incurs licensing costs, or apply heuristics without optimality guarantees. This paper proposes Bisect-Index-Choose (BIC), a fast and solver-free scheduling algorithm that computes a globally optimal allocation among EVs under mild assumptions on their utilities. For $n$ EVs, BIC runs in $O(n \log n)$ time when all EVs share one maximum rate and in $O(n^2)$ time for two, times a logarithmic factor in the numerical precision. In tests with up to 1000 EVs, BIC attains global optimality and is up to three orders of magnitude faster than the solver. On real-world charging data, it needs less capacity and delivers more energy than existing heuristics. We implement BIC on a low-cost microcontroller and show that it solves instances with 1000 EVs in less than 90 ms.
This study investigates sequential multi-task learning using Parameter-Efficient Fine-Tuning (PEFT), LoRA, and 4-bit quantization on Llama 3.1 8B Instruct for mental health problem detection, emotion detection, and sentiment analysis. Using 4,243 English social media posts, it compares two task sequences and demonstrates that task ordering significantly affects knowledge transfer and task interference. The MHP-first sequence achieves the best mental health problem detection (Micro-F1: 0.7624), while the Emotion-first sequence performs better in emotion (F1: 0.6052) and sentiment analysis (F1: 0.9459). A Joint-MTL baseline further highlights the challenges of simultaneous learning, including task interference and generative format collapse. Overall, the findings demonstrate the effectiveness of sequential PEFT for mental health NLP while showing that task order is critical to balancing performance across related tasks. The integration of LoRA and 4-bit quantization also reduces computational and memory requirements, supporting the adaptation of large language models in resource-constrained environments.
Embedded AI is moving from perception to prediction. Robots now use world models$\unicode{x2014}$learned simulators that predict the outcome of candidate actions in a compressed latent space before any is executed. Every predicted step must be computed on board within a tight power and thermal budget, yet the energy these steps fundamentally require, and how far real hardware sits from it, have not been characterized. We derive lower bounds on the arithmetic operations and the CMOS dynamic energy of a latent state transition from information theory and Boolean circuit complexity. Profiling a DreamerV3-style Recurrent State Space Model (RSSM) shows that every rollout operation is memory-bandwidth bound and that a GPU consumes at least $5.9\times10^{10}$ times the CMOS energy floor. Transient 3D thermal simulations show that this inefficiency is also a thermal one, since sustained back-to-back planning heats a millimeter-scale silicon tile to 556 K even behind an ideal copper heat spreader. Guided by this chain from bound to bottleneck, we propose the World Model Planning Accelerator (WMPA), a chip architecture for embedded world model planning that combines near-memory computation, real-time entropy estimation, variable-precision rollout units, and copy-on-write latent scratchpads. With a design target of 40x lower power, the same tile stays near ambient at 306 K in simulation, pointing toward practical planning on thermally constrained robots.
LoRA offers a parameter-efficient adaptation strategy, supporting multi-user workloads via customized adapters on a shared backbone. However, serving concurrent requests with different adapters introduces substantial computation and memory-management overhead. We ask the research question: Do all requests need their adapters throughout the whole generation process? Our motivational study reveals that employing LoRA adapters at every decoding step is unnecessary, releasing them early has limited impact on accuracy while substantially accelerating generation. Although skipping adapters early offers theoretical workload reduction, translating these computational savings into concrete throughput gains requires tailored GPU kernels. We present LoRazor, a multi-LoRA serving framework that combines early adapter release with active-row aware GPU kernels. LoRazor determines when to permanently release a request's adapter and continue decoding with the backbone alone. By dynamically resizing LoRA kernel launches as adapters are released, LoRazor trims idle GPU blocks. Across two backbones and four tasks, LoRazor achieves up to 3.02x request throughput and 2.53x output-token throughput over vLLM multi-LoRA, with a mean relative task-accuracy decrease of 2.89%.
We introduce Micius-Flash, a 27-billion-parameter model for engineering with native multimodal input, deep reasoning and a 256K-token context window. We design the Micius data pipeline as a knowledge-tree-driven data flywheel comprising three stages: knowledge tree construction and expansion, data synthesis, and quality validation with hard-example augmentation. From this data, we train expert models for semiconductor knowledge and reasoning, hardware coding, and semiconductor multimodal understanding, along with agentic experts. Subsequently, we employ Multi-Teacher On-Policy Distillation (MOPD) to supervise training on sequences generated by Micius itself, thereby integrating the capabilities of these expert models into a unified model. On more than 20 datasets with about 15K records, Micius-Flash averages 64.2% on hardware code benchmarks, 15.0 points above a same-size baseline. It scores 62.4% on a semiconductor-specific agent benchmark, exceeding the same-size baseline by 17.2 points. We further develop Micius Harness for industrial applications, connecting the model to specialized tools, including EDA tools and KiCad, as well as execution environments. Case studies in PCB design, chip design workflow automation, and software development demonstrate the model's ability to use tool execution feedback for multi-step planning, code debugging, and iterative verification. We release Micius as an open-source model to advance the development of agents for science and engineering.
Neural network (NN)-based transistor models have proven to be a promising solution to accelerate device modeling. Although these models demonstrated remarkable speedup in circuit simulations, they were not rigorously tested in complex tasks within standard EDA tool flows for chip design, which require very high transistor model precision to provide accurate results and converge properly. To investigate their full potential and promise, we develop NN-based transistor models that accurately predict the drain current and various charges for a 3nm nanosheet, and employ the models for a standard cell library characterization. This enables us to evaluate the accuracy and speedup under a very wide range of SPICE simulations, benchmarking our models against an industry-standard implementation. Since standard cell library characterization involves thousands of SPICE simulations, it serves as a great benchmark for transistor model quality and consistency. To further evaluate model accuracy, we also perform analog- and digital-level simulations for various complex circuits. Our experiments reveal that NN-based models are a valid alternative to existing transistor compact models, capable of producing accurate delay and power estimates with a much higher speed and convergence time compared to the conventional Verilog-A-based industry-standard (BSIM-CMG) transistor compact model. Our NNs achieve sub-0.11% and 0.76% errors for delay and total power evaluation on a wide range of circuits, while taking up to 8 times less time during the standard cell library characterization. Additionally, we provide comprehensive information and guidance on NN accuracy, unveiling the relationship between NN size and its effect on circuit simulation precision.
We present the framework of TACS-RF - a Tool-driven Agentic Circuit Synthesis of RF and mmWave modules. In this framework, a large-language-model (LLM) coding agent (Claude Opus 4.8) directly drives industry-standard electronic-design-automation (EDA) tools such as Cadence Virtuoso/Spectre, Calibre, Cadence EMX, and Ansys HFSS to design RF and mmWave modules. We validate the framework through the agentic design of a 240-251 GHz sub-THz transmitter in a 90 nm SiGe BiCMOS process. The agent semi-autonomously designed the sub-blocks including a 60 GHz push-push Colpitts VCO with second-harmonic extraction, multistage differential cascode power drivers, a push-push frequency doubler, a differential bowtie antenna, and all the interstage matching networks. The transmitter chain with parasitic-extracted devices and EM-verified passives delivers +5.1 to +6.3 dBm across an 11 GHz varactor-only tuning range, phase noise of -80 to -88 dBc/Hz at 1 MHz offset while consuming a DC power of 361 mW. We describe the agent-driven design flow and the resulting division of labor between human and agent, taking a step toward a"text-to-ASIC"future for designing RF/mmWave hardware.
One of the obstacles limiting the fabrication, modeling, and large-scale adoption of memristors is its relatively large variability in electrical characteristics. To address this challenge, we utilize a kinetic Monte Carlo (KMC) simulation tool to illustrate and quantify structural design changes that can lead to a significant reduction of electrical variations in metaloxide (MOx) thin-film memristors, while also enhancing their switching performance. In particular, we explore various MOx interface layer options through which oxide diffusion occurs, and where subsequent filamentation for state change typically starts. Both chemical (MOx interface layer type) and physical (thickness and shape) properties of interface layers are investigated via KMC simulations. It is found that a 2 nm $\mathbf{T a O}_{\mathbf{2}}$ interface layers in $\mathbf{H f O}_{\mathbf{2}}$-based memristors offer the best improvement in simple layer architecture. Inclusion of a cubic or conical seed layer in the interface layer reduces Ron and Roff resistance variations from 6% to 3% range in the KMC simulations. The electrical properties of the proposed devices are then captured into the compact (SPICE and Verilog-A) models to extract circuit performance in crossbar devices, essential for neuromorphic applications and AI/ML accelerators. Their performance is assessed as a function of crossbar size and parasitics, showing that delay, power, and multiply-and-accumulate errors remain well below 8% for n = 16.
This paper presents the design, simulation, and ASIC-focused implementation of a Binary Convolutional Neural Network (BCNN) tailored for energy-efficient edge AI applications. The proposed architecture utilizes binarized weights and activations, substituting standard multipliers with XNOR-popcount operations to reduce hardware complexity and energy usage. The system utilizes modular Verilog HDL, combining convolution and thresholding units in a pipelined architecture to achieve high processing efficiency. The design features parallel processing lanes and clock-gating methods to improve throughput and lower dynamic power usage. Functional verification validates the accurate of binary inference, while post-synthesis findings from Cadence Genus show decreased area and power relative to traditional convolution neural network implementations. The proposed BCNN architecture shows the feasibility of binary neural networks for real-time, low-energy edge intelligence and offers a scalable and reusable framework for future ASIC-driven deep learning accelerators
This work presents a suitable design as well as implementing a reconfigurable 32-bit Multiply-Accumulate (MAC) unit based on a hybrid approximate adder architecture aimed at energy-efficient VLSI applications. The proposed design partitions the adder into three segments according to bit significance: an 8-bit approximate adder for the least significant bits (LSB), a 16-bit carry select adder (CSLA) for the middle region, and an 8-bit Brent–Kung prefix adder for the most significant bits (MSB). This hybrid approach achieves a trade-off between computational accuracy and hardware efficiency by lowering complexity in non-critical regions while maintaining accuracy in critical paths. It operates in both reconfigurable and non-reconfigurable modes, and it can perform both accurate and approximate computations, which is a precious feature in error-tolerance-based applications like signal processing, machine learning etc. The logic of the proposed architectures has been designed through Verilog HDL and the correctness of the proposed technique has been established through the simulation. Physical design flow (from netlist to tape-out ready) is carried out in 14nm technology using SAED 14nm standard cell library at worst SS corner, 0.585V and 125°C temperature, using Synopsys ICC2 tool for placement, clock tree synthesis and routing. Various checks, such as Design Rule Check (DRC), Layout Versus Schematic (LVS) and formal equivalence checking, have been carried out for ensuring correctness of final layout. In summary, the proposed architecture achieves approximately a depletion of area with 14.36% and 22.47% cutdown in delay if comparison takes place with the exact MAC implementation while maintaining acceptable computational accuracy.
Semiconductor systems have become more complex with the advent of AI applications, especially those using AI chiplets at the highest technology nodes like 3 nm.The proliferation of AI applications has made semiconductor systems more complex, especially for AI chiplets at advanced technology nodes like 3 nm. The development time, cost of design and productivity in traditional RTL-to-GDSII design methodology is high because of the dependency on human expertise and iterative optimization. In recent years, the rise of Agentic Artificial Intelligence (AI) and foundation models offers a promising path to redefine semiconductor design automation, offering the capability of making autonomous decisions throughout various phases of the electronic design automation (EDA) process. A hierarchical multi-agent foundation model framework for the autonomous RTL-to-GDSII design of chiplet-based AI accelerator for the RISC-V architecture in 3 nm FinFET/GAAFET technology is presented. The proposed framework is an orchestration, hierarchical structure, composed of specialized agents for specification analysis, RTL generation, verification, design for test (DFT), physical design and timing closure. The floorplanning, placement, routing and congestion prediction are enhanced by AI-driven optimization techniques such as graph neural networks and reinforcement learning, and foundation models are used to convert natural language specifications into synthesizable Verilog/SystemVerilog descriptions. The proposed framework is evaluated using a case study of a chiplet-based RISC-V AI accelerator that consists of compute, memory, interconnect and security chiplets. The experimental validation is done with OpenROAD, OpenLane, Yosys, Vivado and FPGA emulation platforms. Results indicate that the proposed approach enables to save 6 weeks of RTL development time and 100% to 25% manual verification efforts. In addition, the 40% run-time reduction, 27% reduction in routing congestion and 3x faster convergence to timing closure provide a 38% improvement in power-performance-area (PPA) metrics that are achieved by the AI-assisted physical design flow. The results suggest a viable and scalable approach to achieving fully autonomous design of semiconductor devices at high technology nodes for future (UTeM) chiplet-based AI systems using hierarchical multi agent foundation models
Advanced Encryption Standard (AES) is the most prevalent scheme applied in the implementation of a secure transmission of data. However, conventional programs have harsh power and resource demands in the Internet of Things (IoT) environments. The present paper has provided a better AES-256 architecture in Verilog HDL and geared towards secure and energy efficient implementation of distributed IoT image processing systems. This design will be pipelined and clock gated to improve the throughput and save great power. Two new fundamental management schemes, i.e. Lightweight Secure Session Key Generation (LSSKG) and Key Generation Design (KGD) are combined so as to enhance security and to decrease the key-handling overhead. The different versions of the AES are compared and analyzed. It is demonstrated that it does not affect the cryptographic security, area, and timing performance but decreases the power by 99.89 percent (2698.505 mW to 2.866 mW). The in-time safe combination of image processing is used to verify the suitability of the design to the resource-constrained IoT applications.
Convolutional Neural Network (CNN) accelerators rely heavily on Multiply-Accumulate (MAC) operations, which dominate the overall power consumption. An adaptive truncated MAC architecture is presented to achieve an efficient balance between computational accuracy and energy efficiency at the hardware level. The proposed design integrates an exact Booth based MAC unit and a truncated approximate multiplier within a unified configurable framework. A sensitivity aware layer selection mechanism governs the operating mode, assigning exact computation to early convolutional layers to preserve low level feature integrity, while the truncated approximate computation is employed in deeper convolution and fully connected layers to exploit neural network error resilience. A programmable threshold T = αN, where N denotes the total number of convolution layers and α (typically 30–40%) represents a predefined sensitivity factor, which determines the transition between the exact and approximate computation. This is ultimately incorporated to minimize unnecessary switching activity in inactive data paths, thereby reducing dynamic power. This architecture is modeled and functionally verified using Verilog HDL and the system level accuracy and efficiency evaluation using CNN simulation in Colab. Experimental results demonstrate reduced switching activity and improved energy efficiency with bounded accuracy degradation compared to a fully exact MAC configuration, indicating strong potential for FPGA and ASIC based energy-aware CNN acceleration.
Spiking neural networks (SNNs) promise energyefficient temporal processing, but the choice of neuron model affects both task accuracy and hardware cost. We compare three mechanisms for enriching temporal processing in SNNs- heterogeneity, adaptation, and synaptic delays-on auditory benchmarks (SHD, SSC) under controlled conditions, analyzing per-neuron arithmetic cost, state storage, and total energy. Our key finding is that constrained adaptive LIF (cAdLIF) neurons in a feedforward topology outperform recurrent heterogeneous LIF networks while eliminating recurrent connections, the dominant contributor to synaptic energy. Parameter-efficiency sweeps show that cAdLIF networks with $4 \times$ fewer parameters exceed the LIF accuracy ceiling at any size, producing 30% fewer total spikes at matched accuracy. Post-training quantization confirms that 10bit fixed-point incurs <1% accuracy loss on both benchmarks, and RTL synthesis (Yosys + SKY130) shows that, at matched 256$\times$2 size, feedforward cAdLIF is $2.4 \times$ smaller in silicon area than recurrent LIF, despite its higher per-neuron cost.
Contemporary embodied multimedia systems demand agents capable of seamlessly bridging high-dimensional sensory streams, complex natural language instructions, and precise physical interactions. While Vision-Language-Action (VLA) models show promise in this direction, their deployment in safety-critical, real-time scenarios like autonomous driving is constrained by inefficient processing of redundant visual tokens, weak 3D spatial grounding, and an opaque semantic-to-control gap. We propose HiST-VLA, a hierarchical and spatio-temporal VLA framework specifically designed to handle the rigorous multimodal requirements of embodied systems. Our core contributions are: 1) A dynamic token sparsification mechanism that dynamically prunes redundant visual tokens based on self-similarity attention, serving as an efficient cross-modal alignment layer for high-throughput multi-view video streams. 2) A multi-stage, semantics-aligned hierarchical planner that bridges granular language commands with precise motion primitives, incorporating confidence-aware regularization for transparent uncertainty quantification. Extensive evaluation on the NAVSIM v2 benchmark demonstrates state-of-the-art performance, achieving an EPDMS of 88.6 on the open-loop Navtest set—surpassing prior SOTA methods by 1.5 points. Crucially, we provide in-depth analyses of cross-modal token efficiency, interpretability, and real-world deployment feasibility, establishing a robust paradigm for trustworthy embodied multimedia intelligence.
Imagined handwriting decoding offers a promising pathway toward non-invasive brain-to-text communication without requiring overt speech or physical movement. However, learning reliable neural representations from limited multimodal brain-signal data remains challenging. Existing methods typically rely on convolutional filters to discover task-relevant frequency patterns directly from the data, without explicitly incorporating the characteristic neural oscillations associated with motor imagery. Moreover, conventional symmetric or tightly coupled multimodal fusion may allow less informative fNIRS signals to interfere with the more discriminative EEG pathway. To address these challenges, we propose a unified framework for EEG--fNIRS imagined handwriting decoding. First, we introduce Motor Imagery Spectral-Prior Modulation (MISM), which extracts spectral features from the motor-imagery-related alpha and beta bands and uses them to adaptively modulate early convolutional representations. This design explicitly injects neurophysiological frequency priors into temporal feature learning. Second, we develop Confidence-Gated Asymmetric Fusion (CAF), which treats EEG as the primary prediction stream and selectively incorporates complementary fNIRS evidence when the EEG prediction is uncertain. Third, we propose NLL-Guided Ensemble Composition Optimization (NECO) to improve generalization under limited-data conditions. NECO ranks candidates by out-of-fold NLL-fitted weights and selects a compact top-N subset by validation accuracy, outperforming equal aggregation over all candidates. On the EEG--fNIRS Handwriting-Trajectory Classification dataset released for the ACM MM Challenge, our framework achieves state-of-the-art performance with an accuracy of 82.61%.
Meteorological reanalysis encodes extreme weather through continuous, physically constrained fields, posing a fundamental challenge for vision–language models (VLMs) whose perceptual assumptions are shaped by natural images. Tropical cyclones exemplify this mismatch: critical properties such as intensity extrema, asymmetry, spatial extent, and physical impacts arise from field-level organization rather than object-centric visual cues. Existing approaches address this gap through text alignment or annotation, treating the problem as multimodal supervision rather than representation design. We introduce Physics-Aware Visual Abstraction (PAVA), a plug-and-play physics-aware representation and annotation interface that maps physical reanalysis fields to visually identifiable and semantically grounded perceptual abstractions for supervision and evaluation in vision--language reasoning. Building on PAVA, we construct OmniClimate-TC, a benchmark for tropical cyclone analysis spanning five classes of reasoning and nine tasks, with 243,890 physically grounded instruction-tuning pairs. Using PAVA-aligned supervision, we adapt VLMs and provide evidence that this representation design improves reasoning over tropical cyclone hazard fields. Our results position OmniClimate-TC as a benchmark for multimedia reasoning over structured geophysical fields, and highlight representation design as a key ingredient for physically grounded reasoning in scientific media.
Multimodal 3D object detection is a cornerstone of autonomous driving perception, yet its adversarial robustness remains poorly understood. Due to the lack of a unified cross-modal representation, existing adversarial attacks manipulate image and point cloud inputs separately, leading to limited physical realizability and suboptimal attack effectiveness. To address this problem, we propose MiuGS, a modality-integrated unified Gaussian representation framework for generating 3D adversarial objects with strong attack effectiveness and physical-world potential. For physical realizability, we design a multimodal 3D Gaussian coordination strategy to address the imprecise cross-modal fusion issue. Specifically, MiuGS models the adversarial object as a shared 3D Gaussian representation and instantiates it into consistent image and point cloud observations, enabling the two sensing streams to be coupled at the representation level rather than optimized independently. To further strengthen attack effectiveness, we introduce a cross-scenario guidance rebalancing strategy to mitigate cross-modal guidance drift across scenarios and the digital-to-physical domain gap. Extensive experiments show that MiuGS achieves strong attack effectiveness, reaching an average ASR of 78.8% on nuScenes and 44.3% on CARLA, supporting the effectiveness and physical-world potential of the proposed unified representation.
Lensless imaging has emerged as a promising paradigm for ultra-compact, low-cost, and privacy-preserving visual sensing. However, extending recent advances in 2D lensless reconstruction to 3D novel view synthesis remains challenging. Existing methods mainly rely on neural-radiance-field-based representations, which suffer from slow rendering and limited practicality. In this work, we take the first step toward high-fidelity lensless novel view synthesis using 3D Gaussian Splatting (3DGS). This setting poses three major challenges: maintaining cross-view geometric consistency, suppressing ghosting in occluded regions, and restoring high-frequency textures under efficient optimization. To address these issues, we propose Uncertainty-Aware Consensus 3D Gaussian Splatting (UACGS), a framework that combines structure-aware consensus preprocessing with mask-guided optimization. Specifically, our framework incorporates: (1) a structure-aware consensus module that leverages adjacent camera poses to filter out occlusion-induced erroneous projections; (2) an uncertainty-aware mask-guided supervision mechanism that effectively corrects reconstruction errors caused by physical limitations; and (3) a contextual loss design that restores fine-grained high-frequency textures while preserving geometric stability. Extensive experiments show that our framework achieves superior reconstruction quality across datasets, while also delivering strong cross-dataset generalization, real-time rendering, and highly efficient training. The code is publicly available at https://github.com/YUNYU7472/UACGS.
A primary challenge in deep-sea 3D reconstruction is the strong coupling between active illumination and the participating water medium. Existing underwater 3D Gaussian Splatting (3DGS) methods typically model backscattering along pixel rays as a fixed constant, neglecting the distance-dependent attenuation of active light sources. In terms of optimization, the densification process often confounds view-space positional gradients with medium noise signals, leading to systematic errors in radiance measurements and redundant primitives. To address these challenges, we introduce AbyssSplat, the first 3D reconstruction framework specifically designed for complex deep-sea optical environments. AbyssSplat incorporates two core innovations: a Hybrid Light Field Illumination Attenuation Model, which explicitly models active light source scattering to achieve physical-domain decoupling between non-uniform illumination and scene-intrinsic radiance; and a Medium-Guided Densification Modulation Mechanism, which integrates medium transmittance and Gaussian attributes to construct a geometric confidence, applying adaptive gradient amplitude modulation to low-confidence regions to suppress overfitting of medium noise by Gaussian primitives. Evaluations on the Eiffel Tower deep-sea and SeaThru-NeRF shallow-water datasets demonstrate that AbyssSplat achieves consistent state-of-the-art performance across all scenes, with average PSNR improvements of 0.67 dB and 0.39 dB respectively over the best competing methods, alongside up to 49.5% LPIPS reduction—while requiring fewer Gaussian primitives and less training time.
Predicting accurate spatiotemporal sequences in complex spatiotemporal systems is a fundamental challenge in machine intelligence. A critical aspect lies where real-world systems always appear critical transitions can be anticipated through early-warning signals. However, existing predictive models primarily learn visual representations from historical sequences, overlooking dynamic transient patterns governing behind system evolutions in several inherent challenges, including physical concealment, causal interpretability and hotspot effect. Thus, we propose a novel predictive framework called transient patterns discovered for spatiotemporal sequence forecasting (TransientCast), seamlessly integrating transient pattern discovery with spatiotemporal predictive learning. In order to quantitatively capture transient patterns, disentangled transient pattern modeling is proposed to extract transient patterns using transient pattern discovery and then disentangled as transient embeddings together with visual embeddings. Interpretable future transient patterns and sequences are jointly predicted by a multivariate causal transformer, ensuring causal dependency. To further enhance the predictive accuracy of identified critical regions, hotspot refinement is designed by focusing on regions with high transient modularity. Extensive experiments on several scenarios show significant improvement in predictive accuracy and plausible predicted future transitions. (https://github.com/miaomiaoToday/predictiveModel.git)
Large Language Models (LLMs) incur severe memory-access overheads due to the cost of loading large weight matrices and key-value (KV) caches. Analog In-Memory Computing (AIMC) is an emerging solution that enables computation directly on static weights stored in memory, substantially reducing energy and latency. 3D-IMC further stacks memory arrays into vertical tiers, increasing capacity but revealing a new challenge: how weights are placed in memory determines which matrix-vector multiplications (MVMs) can execute concurrently, affecting latency and throughput. The concurrency aspect of 3D-IMC weight-mapping remains underexplored in the literature, leading to sub-optimal utilization of compute resources on 3D-IMC chips. In this context we propose CAMP (Context-Aware Mapping and Placement), an inference-context-aware weight mapping algorithm for 3D-IMC-based LLM accelerators. CAMP captures the potential resource contention resulting from the architecture, targeted network, and runtime context in a conflict matrix, enabling throughput-optimized weight mapping. This mapping is refined through a profiling-driven feedback loop that efficiently prunes the search space and returns optimized mappings within practical runtimes. CAMP is an offline, software-based solution requiring no hardware changes and improving 3D-IMC compute utilization from 42.4% to 60.3%, delivering up to 2.0x higher 3D-IMC throughput and 1.74x higher end-to-end throughput over state-of-the-art mapping algorithms.
Diffusion models have become the dominant paradigm for high-quality image generation, but their iterative inference across tens of timesteps incurs substantial computational overhead. Prior accelerators exploit temporal similarity in diffusion models via differential quantization or output sparsity, yet suffer from increased memory footprint, design complexity, and image degradation. In this work, we present Phantom, an algorithm-architecture co-designed accelerator built on temporal caching. We view temporal similarity as differential sparsity: between adjacent timesteps, only a small fraction of tokens (outliers) change significantly. Phantom recomputes only these outlier tokens and reuses the rest from a temporal cache. Phantom runs on standard systolic arrays with lightweight additions: a Phantom Engine at the global buffer that detects outliers and updates the cache as activations stream through, and a condensed pipeline that maps the resulting sparse updates onto dense systolic computation for FC, convolution, and attention layers. Evaluations on modern diffusion models show that Phantom achieves an average 4.0x speedup and 5.5x energy saving over GPUs and prior accelerators while preserving generation quality.
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.
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.
Male infertility contributes to approximately half of all cases of couple infertility worldwide, yet its diagnostic cornerstone, conventional semen analysis anchored in World Health Organization (WHO) criteria, suffers from inter-observer variability that can reach 30–40% for morphological parameters even among trained andrologists. Computer-assisted sperm analysis (CASA) reduced some sources of error but failed to eliminate the subjectivity inherent in morphological classification, particularly for the acrosome, midpiece, and cytoplasmic droplet defect categories. Convolutional neural networks and ensemble deep learning architectures have since demonstrated high classification accuracy on selected benchmark datasets (detailed under Results), prompting fundamental questions about the clinical readiness and translational utility of AI-assisted andrology. This is a critical review that synthesises evidence from peer-reviewed studies identified through targeted searches of PubMed, Scopus, and IEEE Xplore (through May 2025), covering AI applications in sperm morphology classification, motility analysis, non-obstructive azoospermia (NOA) outcome prediction, and sperm selection for intracytoplasmic sperm injection (ICSI), supplemented by hand-searching the reference lists of key articles. Search terms combined controlled vocabulary and free-text terms for “sperm morphology,” “deep learning,” “convolutional neural network,” “azoospermia,” “testicular sperm extraction,” and “ICSI,” restricted to English-language, peer-reviewed publications. Studies were selected for methodological relevance, recency, and clinical significance rather than through a pre-registered, reproducible systematic protocol; no formal risk-of-bias or quality-appraisal instrument (e.g., QUADAS-2 or PROBAST) was applied. This is a deliberate scope decision, declared here rather than implied: the review is a critical, evidence-synthesising appraisal of a fast-moving field, not a systematic or scoping review, and quantitative performance figures drawn from heterogeneous studies (summarised in Table 1) should be read as illustrative of the state of the field rather than as directly comparable, pooled estimates. Deep learning models, including ensemble and attention-augmented convolutional architectures, have reported accuracies of 94% to 96.77% and F1-scores of up to 98.2% for sperm morphology classification, principally on the HuSHeM benchmark dataset; the morphology-classification studies summarised in Table 1 report accuracy, precision, or F1-score rather than AUC (area under the receiver operating characteristic curve). Machine learning models integrating hormonal, genetic, and histopathological features achieve AUC values of 0.807 to above 0.90 for predicting sperm retrieval success in microdissection testicular sperm extraction (m-TESE). For AI-assisted real-time sperm selection during ICSI, a prospective sibling-oocyte study of 646 metaphase II oocytes found only a non-significant trend towards improved fertilisation, whereas a separate, smaller sibling-oocyte comparison, reported to date only as a conference abstract, described a statistically significant improvement in fertilisation; live birth rate data remain scarce. However, the predominance of single-centre, algorithmically homogeneous training datasets, absence of prospective multicentre validation, and the opaque decision logic of most deployed models collectively constrain clinical adoption. AI-assisted sperm morphology assessment has attained technical maturity but not clinical readiness. The transition from benchmark accuracy to actionable diagnostic utility requires prospective multicentre validation using clinically diverse and ethnically representative cohorts, live birth rate pre-specified as the primary endpoint, with fertilisation and blastulation rates as secondary endpoints, and the development of explainable AI frameworks compatible with regulatory standards including the EU AI Act. The Middle East and North Africa region, characterised by high rates of male-factor infertility, distinct genetic aetiologies, and constrained laboratory infrastructure, represents both a critical validation environment and an equity-relevant deployment target for these technologies.
Differentiable logic networks (DLNs) are interpretable, edge-friendly classifiers that learn both their logic operators and their connections through gradient descent, achieving accuracy competitive with multi-layer perceptrons at orders-of-magnitude lower inference cost. Their accuracy, however, is bounded by the problems of vanishing gradients, scalability, and lack of an ecosystem, leaving a gap relative to conventional neural networks on several datasets. We introduce knowledge distillation for DLNs to narrow this gap. We distill from a trained multi-layer perceptron teacher into a DLN student, integrating the distillation objective into a two-phase training loop without altering the architecture or the inference procedure. The teacher is discarded after training. Hence, the distilled model retains the interpretability, sparsity, and inference cost of a conventionally trained DLN. We investigate this approach through four variants: a joint hyperparameter search that optimizes the distillation weight and temperature alongside the architecture, a grid search over these two hyperparameters with the architecture fixed, and two single-phase variants that employ distillation during only the neuron-function or only the connection phase. On 20 public tabular classification tasks, distillation improves balanced-class accuracy at negligible cost to gate-level operation counts and model size, though the benefit varies by dataset. Performance improvements of up to 15% are observed, in some cases surpassing the teacher itself.
BACKGROUND Neurodegenerative diseases (NDDs), including Alzheimer's disease (AD), Parkinson's disease (PD), and Amyotrophic Lateral Sclerosis (ALS), present a growing global health challenge, with traditional diagnostic and therapeutic approaches facing significant limitations in early detection and mechanistic resolution. OBJECTIVES Innovations in transcriptomic technologies-spanning bulk RNA sequencing, single-cell RNA-seq (scRNA-seq), and spatially resolved transcriptomics-provide unprecedented insights into cellular heterogeneity and microenvironmental dysregulation. This systematic review synthesizes recent applications of artificial intelligence (AI), machine learning (ML), and deep learning (DL) across transcriptomic tiers to benchmark computational models for biomarker identification, disease diagnosis, patient subtyping, and drug repurposing. METHODS & ELIGIBILITY Following PRISMA 2020 guidelines for qualitative systematic reviews, literature published between 2016 and 2025 was systematically searched across PubMed/MEDLINE, IEEE Xplore, Google Scholar, and bioRxiv/medRxiv. Of 1058 records identified, 19 core empirical studies met all eligibility criteria and were synthesized in depth alongside contextual reference benchmarks. SYNTHESIS & KEY FINDINGS Traditional ML algorithms (SVM, RF, Elastic Net) demonstrate high diagnostic accuracy (AUC 0.72-0.98) in blood-based biomarker selection, while deep architectures (autoencoders, LSTMs, and Graph Neural Networks) excel at modeling continuous disease trajectories and resolving spatial transcriptomic niches (e.g., STAGATE, SpaGCN, cell2location). Emerging spatial deep learning models successfully localize disease-associated microglia (DAM) around amyloid plaques and trace axonal degeneration pathways. LIMITATIONS & CONCLUSIONS Key translational hurdles include data scarcity in rare NDD subtypes, lack of cross-platform spatial benchmarks, and the black-box nature of deep neural networks. Integrating explainable AI (XAI) and prospective multi-modal cohorts is essential for translating computational transcriptomics into clinical diagnostics and targeted therapeutics.
Unidirectional Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) models are types of Recurrent Neural Networks (RNNs) capable of modelling long-range dependencies in financial time series, such as stock prices. However, the performance of LSTM and BiLSTM is highly sensitive to hyperparameter tuning, and there is limited guidance on systematic tuning, particularly for nonlinear stock-price series with varying historical lengths. Hence, this study aims to optimise the forecasting performance of LSTM and BiLSTM by tuning the hyperparameters through grid search and examining their behaviour. Daily Google stock prices were collected from Yahoo Finance, covering 3 January 2005 to 25 May 2022, and the adjusted closing price was selected as the target because it reflects corporate actions and offers more stable price movement. Two datasets were analysed, Sample 1 (25 May 2017–25 May 2022) and Sample 2 (3 January 2005–25 May 2022). Each dataset was split using an 80:20 train–test design. Min–Max normalisation was fitted on the training set and applied to the test set using the same parameters. Both architectures were implemented with a single recurrent layer and a single output neuron for regression, trained using the Adam optimiser and Mean Squared Error (MSE) loss. A grid search considered hidden neurons (32, 64, 128), epochs (20, 50, 100, 150), and batch sizes (32, 64, 128). The results show that in Sample 1, the best LSTM achieved MSE 8.4748 and MAE 2.2192, while the best BiLSTM achieved MSE 7.4766 and MAE 2.1031, both at 128 neurons, 150 epochs and batch size 32. For Sample 2, the best LSTM achieved MSE 4.0303 and MAE 1.4567, and the best BiLSTM achieved MSE 3.9929 and MAE 1.4373 under the same hyperparameters. In conclusion, increasing the number of hidden neurons and training more epochs generally reduced errors, while larger batch sizes tended to increase errors, and BiLSTM yielded lower average MSE and MAE than LSTM across the two samples.
Adult spinal deformity (ASD) demands precise radiographic quantification of spinopelvic parameters to guide corrective surgery. Artificial intelligence (AI) and machine learning (ML) tools offer automated alternatives to manual measurement. This systematic review evaluates the diagnostic accuracy, failure rates, and inferred surgical planning implications of AI-based spinopelvic measurement versus manual assessment, with particular focus on postoperative and instrumented spines. This Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020-compliant systematic review searched MEDLINE (PubMed), ScienceDirect, Cochrane Library, and Google Scholar (first 200 results) through June 2026. Studies comparing AI/ML tools, including convolutional neural networks (CNNs), U-Net, Mask R-CNN, and Keypoint R-CNN, against manual clinician measurement on radiographs or EOS biplanar systems were included. Quality was assessed using QUADAS-2. Narrative synthesis was performed given heterogeneity across architectures, populations, and outcome metrics. Sixteen studies (sample sizes ranging from 21 to 1,810 patients per study) compared machine vision against manual measurement across preoperative, postoperative, and instrumented cohorts. Compared to manual measurement, automated machine vision demonstrated agreement statistically comparable to manual measurement for sagittal vertical axis (SVA) (intraclass correlation coefficient (ICC) 0.99-1.00) and pelvic tilt (PT) (ICC 0.96-0.99), matching experienced clinician inter-rater reliability. For lumbar lordosis, automated systems achieved acceptable accuracy in preoperative populations (ICC 0.91-0.97) but exhibited directional biases absent in manual measurement (up to 6.31° overestimation). Thoracic kyphosis and sacral slope agreement was highly variable and dataset-dependent (ICC ranging from 0.58 to 0.97 across studies), generally among the least consistent parameters rather than uniformly weak. Unlike manual measurement, automated accuracy declined significantly in postoperative and instrumented spines. Machine vision failure rates ranged from 0% to 76.2%, versus near-zero manual failure rates. Processing speed was up to 238-fold faster than manual workflows. Automated-versus-manual discrepancies for LL and PI occasionally exceeded the 10° PI-LL surgical correction threshold used as a proxy for surgical-planning impact, since no included study measured a surgical outcome directly. Machine vision matches manual speed and accuracy preoperatively but is not yet reliable enough postoperatively to replace manual clinical assessment without standardised validation.
This paper presents an artifact-complete evaluation of Pareto-balanced three-parent genetic neural architecture search (PBO-3PGA) for compact maize leaf disease classification. The official PlantVillage raw-color maize subset was locked by source group and SHA-256 hash into 3,052 training, 400 validation, and 400 independent test images; validation and test each contain 100 images per class. Four controllers were compared with identical budgets over three seeds, producing 864 traceable candidate records. PBO-3PGA attained $0.9283 \pm 0.0014$ test accuracy, 0.9277 ± 0.0015 macro-F1, and 0.9917 ± 0.0002 macro-AUC with an 82,244-parameter, 164,416-FLOP searched head. Accuracy-only 3PGA-NoMO was slightly more accurate (0.9300 ± 0.0050) but used, on average, $8.07 \times$ more head parameters and FLOPs. Thus, PBO-3PGA is an efficiency-oriented Pareto choice rather than the raw-accuracy winner. Validation-only temperature scaling reduced PBO-3PGA ECE from $0.0348 \pm 0.0064$ to $0.0200 \pm 0.0033$, Brier score from 0.1105 ± 0.0039 to 0.1059 ± 0.0017, and NLL from 0.1920 ± 0.0166 to 0.1761 ± 0.0040. On a four-thread Apple M5 CPU, the searched head required 0.0092 ms/image, whereas the normalized-tensor-to-logit path required 98.04 ms/image, showing that the shared encoder dominates end-to-end latency. Exact objective weights, normalization, chromosome decoding, generation examples, logits, and raw timing traces are retained for reproduction.
Oral squamous cell carcinoma (OSCC) is associated with considerable morbidity because of occult lymph node metastasis (OLNM). Conventional imaging has limitations in detecting OLNM in oral cancer. AI-based technologies have been proposed as potential aids in diagnosis. This systematic review was carried out following PRISMA 2020 guidelines and registered in the International Prospective Register of Systematic Reviews (PROSPERO) (February 2025). PubMed, Scopus, Embase, Web of Science, and Google Scholar were searched. Studies utilizing artificial intelligence (AI)-based techniques, such as machine learning (ML), deep learning (DL), and convolutional neural networks (CNNs), in the detection of lymph node metastasis in oral cancer with reported results of diagnostic accuracy, sensitivity, specificity, or area under the curve (AUC) were included. Methodological quality was assessed by Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2); evidence certainty was assessed by Grading of Recommendations, Assessment, Development, and Evaluation (GRADE). A total of 15 articles (n = 5,758) fulfilled the inclusion criteria. The AI-based models yielded AUC values of 0.729-0.961, sensitivity of 0.65-1.00, and specificity of 0.576-0.98. DL-based CNN architectures and radiomics-DL hybrid models showed the highest diagnostic accuracy. Moderate overall risk of bias was reported with the QUADAS-2 tool, while GRADE provided certainty of evidence as low to very low due to study heterogeneity and lack of external validation. The techniques employing AI appear to have great potential for diagnosing OLNM in patients with mouth cancer because they have more accurate capabilities than imaging. There is a need for future multicenter validation, standardization of algorithms, and ethical frameworks for their application.
Clinical data are split across incomplete cohorts with partially overlapping variables. Causal discovery under missing data and multi-source integration have developed separately, and tabular foundation models cannot take a causal graph without retraining. We make two contributions. First, causal graph fusion learns a partial ancestral graph on the shared variables, fuses it into each cohort's graph under temporal background knowledge and lets Markov blankets guide re-imputation, so missing data are handled inside discovery. Second, causal attention steering compiles the learned graph into the attention of a frozen tabular foundation model at inference. Against known ground truth, fusion dominates pooled and per-cohort baselines in every missingness and heterogeneity condition tested (p<0.001), and steering improves discrimination under distribution shift (p<10^{-5}), a gain that a random graph does not produce. In two spinal cord injury cohorts, fusion orients age, albumin and hemoglobin as ancestors of length of stay, a set no single cohort orients in full, and yields compact, temporally ordered predictor sets that match or exceed automatic feature selection on four of five outcomes.
The National Institutes of Health (NIH) fund several programs (aka consortia) that support the construction of organ-specific or whole person atlases. These initiatives are designed to build comprehensive open reference datasets while advancing knowledge graphs, analytics, visualizations, and user interfaces needed to explore multiscale human atlas data. Here, we compare the Human Tumor Atlas Network (HTAN), Human BioMolecular Atlas Program (HuBMAP) and Cellular Senescence Network (SenNet) to examine collaboration by consortium component (CC) and researcher role and its relation to scientific impact. We retrieved 884 NIH award records and identified 1,255 focal publications acknowledging these awards, involving 8,952 researchers; compiled 99,108 citing publications covering 93% of the focal publications; constructed co-authorship networks; and as single-project publications or internal collaborations (IC), including intra-CC and inter-CCs collaborations. within each atlas program. Results show that IC publications are rare but involve many funded projects and principal investigators (PIs). Technology-oriented CCs tend to engage in intra-CC collaboration, whereas experimental-oriented CCs often collaborate across CCs of the same type. Compared with other researchers, PIs exhibited a significantly stronger association with inter-CCs than with single-project publications. IC publications have higher mean Relative Citation Ratios than single-project publications. Inter-CCs publications have significantly higher cumulative and annual citation levels in the first five post-publication years. Study methods and results can be used to inform the funding, management, and impact reporting for future Big Science programs that bring together 300+ experts from different disciplines to address major research challenges.
Molecular dynamics predicts molecular motion from atomic forces and trajectory integration. Density functional theory (DFT) provides first-principles forces, but its $O(N^3)$ cost limits simulations to thousands of atoms; empirical force fields scale to millions but miss bond breaking and charge transfer. Equivariant message-passing Transformers learn the ab initio potential at linear cost. Stacked layers propagate information beyond a single cutoff, giving four layers a receptive field up to 23{\AA}. At this frontier, scale is limited by communication: layers exchange boundary features across devices, force backpropagation repeats the exchange, and moving atoms change neighbor graphs and tensor shapes. We present Tessera, a distributed runtime that fixes workload shapes between rebuilds. It sorts atoms along a space-filling curve into equal-count shards, ensuring identical atom counts and tensor shapes per rank. Peer sets scale as $K^{0.5}$, not $K$. A conservative neighbor skin preserves topology across updates, enabling each rebuild interval to replay as one CUDA Graph. Fused kernels, three-level neighbor search, and overlapped All-to-All complete it. On 512 GPUs, Tessera sustains 30.7M atoms at 89.7% weak-scaling efficiency (77--90% across three per-GPU loads); per-step time rises 1.11--1.29x as machine and problem grow 16x. Graph reuse cuts full-step time $1.87\times$. Across an $8\times$ device increase, neighbor-search cost grows $1.7\times$, versus $3.0\times$ for Morton order. In a live $10.24$M-atom salt solution, NVE energy drift is $0.42\,\mu$eV/atom/ps over $10$ ps, and structure matches a serial reference. A 10.06M-atom Norovirus capsid sustains 100 ps of NVT at $\approx 8$ ps/day on 256 A100 GPUs. Over 200,000 steps, it has no broken bonds or anomalously close pairs and uses three compiled variants, with no graph breaks or eager fallbacks.
Proprotein convertase subtilisin/kexin type 9 (PCSK9) plays a central role in cholesterol homeostasis by promoting the degradation of low-density lipoprotein receptors (LDLR), making it an attractive therapeutic target for cardiovascular disease. Although monoclonal antibodies and RNA-based therapies targeting PCSK9 have shown clinical success, developing orally available inhibitors remains a major challenge, largely because of the complex protein-protein interaction landscape of PCSK9. In this study, we present an integrated artificial intelligence-driven drug discovery framework that combines a graph neural network (GNN) model with structure-based molecular modeling to identify and characterize novel allosteric PCSK9 inhibitors. A curated dataset of 1,039 PCSK9 ligands with experimentally measured IC₅₀ values was compiled from multiple public databases and used to train an optimized message-passing GNN. The resulting model demonstrated excellent predictive performance (R² = 0.949 on the test set) and robust generalization within a well-defined applicability domain. To translate predictive insights into mechanistic understanding, top-ranked compounds were subjected to a multistage structural workflow including blind and targeted docking, pharmacophore modeling, molecular dynamics simulations, MM/GBSA binding free-energy calculations, and protein-protein interaction analysis of the PCSK9-LDLR interface. Among the identified candidates, Ligand40 exhibited the most favorable binding energetics, while Ligand42 demonstrated stable interaction patterns and enhanced dynamic restraint, supporting its prioritization for further investigation. Overall, this work demonstrates that coupling deep graph learning with physics-based simulations can facilitate the discovery of candidate PCSK9 modulators. The proposed framework is scalable and transferable, enabling AI-assisted discovery of small-molecule and peptide modulators for challenging protein-protein interaction targets.
Optimizing accelerator code across hardware architectures is difficult. ML compilers search predefined spaces of schedules and micro-kernels, while LLM-guided optimization can explore alternative algorithms, operator decompositions, and hardware-specific techniques. However, existing LLM agents typically target a single programming interface, toolchain, or accelerator family, focus on kernel-level optimization without model-level transforms, and offer limited reuse of experience across hardware targets. We present ModelBlaster, a Memory-Augmented In-Context Reinforcement Learning (MAIC-RL) framework for optimizing data-parallel kernels and whole models across heterogeneous SIMD and SIMT accelerators. Its mutable policy is a persistent knowledge base updated at inference time by Actor-Critic Textual Gradient Descent (AC-TGD). The knowledge base maps performance states to actions, outcomes, and constraints, enabling experience reuse across kernels, models, and hardware targets. A model-level optimizer uses kernel-level knowledge to inform whole-model decisions. The same optimization loop and knowledge schema support CUDA on NVIDIA GPUs, OpenCL on Qualcomm GPUs, and RVV intrinsics on RISC-V vector cores, with only lowering and measurement layers specific to each target. Compared with PyTorch, ModelBlaster achieves geometric-mean speedups of 1.85x, 5.58x, and 5.18x on KernelBench Levels 1, 2, and 3, and 1.69x on SOL-ExecBench Levels 1 and 2 on a B200 GPU. On Qualcomm Adreno GPUs, it achieves 5.8x over the vendor SDK and 1.7x over Apache TVM; on RISC-V vector cores, it achieves 2.7x over ExecuTorch with XNNPACK across 70 KernelBench problems.
The construction sector still faces a high incidence of fatal accident and ergonomic injury risk, largely due to the inherent limitations of traditional manual safety auditing. While computer vision (CV) provides a viable solution to this problem, traditional deep learning architectures have struggled to balance the parametric requirements of micro-object detection, such as safety masks, with the stringent hardware limitations of traditional computer vision solutions. Furthermore, traditional spatial detectors are unable to assess active biomechanical safety risk and manage automated incident reporting without compromising inference latency. To address this critical gap in safety auditing, this research introduces a novel unified cyber-physical system using a Tri-State Routing Engine. By stratifying a dataset in a deterministic manner to prevent micro-object parametric starvation, the system dynamically routes inference using a hardware-aware Hybrid Ensemble of YOLO11s and YOLO11m backbones. The spatial inference of overlapping regions is mathematically fused using a custom-designed Expert-Calibrated Weighted Boxes Fusion (EC-WBF) algorithm to reduce semantic inconsistencies and deliver a peak mAP@50 of 0.910. The system also bridges passive spatial detection and active kinematic tracking using a novel routing mechanism to dynamically crop worker tensors and route them to a lightweight YOLO11n-pose network. This allows simultaneous multi-target Rapid Entire Body Assessment (REBA) ergonomic tracking and extrapolates Implicit Geometry failsafes to account for missing PPE even in severe visual occlusion. The system also introduces a novel natively coded asynchronous Robotic Process Automation (RPA) daemon. This daemon runs completely decoupled from the GPU-bound visual inference loop and sustains 14.2 FPS on severely constrained local edge hardware with 4 GB VRAM, rapidly sending Tier 1 life safety alarms for kinematic falls and compiling Tier 2 admin compliance digests, thereby successfully closing the cyber-physical loop.
Artificial intelligence (AI) is increasingly being explored as a tool to support food risk assessment and broader food safety activities in response to the growing complexity of global food supply chains, emerging foodborne hazards, and evolving food fraud practices. Unlike previous reviews that have largely focused on individual AI technologies or specific applications, this review provides a comprehensive and critical synthesis of machine learning (ML), deep learning (DL), transformer-based models, explainable artificial intelligence (XAI), multimodal learning, and emerging AI paradigms within a unified framework relevant to food risk assessment and food safety decision-making. The review examines their applications in hazard identification, contamination detection, predictive microbiology, risk classification, food fraud detection, surveillance, traceability, and decision-support systems. Particular attention is given to the strengths, limitations, interpretability, validation requirements, and implementation challenges associated with different AI approaches. The integration of AI with multimodal data sources, Internet of Things (IoT) technologies, spectroscopy, biosensors, and advanced analytical platforms is also discussed as a means to improve risk prediction, monitoring, and early-warning capabilities. Key challenges include data quality and availability, model generalizability, limited external validation, computational requirements, interpretability, and regulatory acceptance. Emerging developments, including foundation models, federated learning, blockchain, digital twins, edge AI, and multimodal frameworks, are highlighted as promising directions for future research. Overall, this review provides a critical comparative framework for understanding the evolving role of AI in food risk assessment and related food safety applications, while identifying key scientific, technical, and regulatory priorities for future development and implementation.
Machine learning from local atomic environments offers an efficient route to electronic structure, but evaluating global electronic properties in large systems remains computationally demanding. Here we combine real space charge density learning with stochastic propagation using Kohn-Sham Hamiltonians constructed from the predicted densities. The framework avoids target system self-consistent iterations and explicit diagonalization, while hierarchical blocking enables dense grid inference within limited GPU memory. We demonstrate density of states calculations for one million Si atoms and optical and transport calculations for over one hundred thousand atoms, without training separate models for each property. This framework extends the practical scale of electronic structure calculations within Kohn-Sham DFT, opening opportunities to investigate how atomic structure governs electronic, optical and transport properties at length scales relevant to real materials and devices.
The rapid growth of cloud computing, Internet of Things (IoT) devices, enterprise networks, and interconnected digital services has increased the complexity of cybersecurity threats. Traditional signature-based security mechanisms often struggle to identify zero-day attacks, evolving malware, and previously unseen malicious activities. This review paper examines the application of Artificial Intelligence (AI), Machine Learning (ML), and Deep Learning (DL) in intelligent cyberattack detection and dynamic security monitoring. It reviews commonly investigated algorithms, including Random Forest, Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and anomaly detection techniques. Existing research is compared in terms of detection capability, computational requirements, adaptability, false-positive rates, and deployment challenges. The review also examines network traffic analysis, behavioral monitoring, real-time alert generation, and automated incident response. Major challenges include imbalanced datasets, concept drift, adversarial attacks, limited generalization to unseen threats, and the difficulty of deploying computationally intensive models in resource-constrained environments. The review identifies opportunities for hybrid AI architectures, explainable detection, privacy-preserving collaborative learning, and feedback-driven security monitoring. It concludes that combining accurate threat classification with continuous monitoring and carefully controlled response mechanisms represents a promising direction for improving cybersecurity resilience across modern digital infrastructure.
We Continuous physiological monitoring through healthcare system with Internet of Things (IoT) technologies is useful for early detection of any impending instability, however, the reliability, strong communication, efficient computation and integration into workflows of such systems are all key factors for their effectiveness. This study created an IoT-based machine learning framework for future next window physiological-instability assessment and healthcare services prioritization using the BIDMC PPG and Respiration Dataset that contains data from 53 subjects. Physiological sensors used were heart rate, pulse, respiratory rate, SpO₂, respiration waveform, photoplethysmography and ECG Lead II. Five fold subject-grouped cross validation was performed for six machine-learning models. Random Forest achieved the strongest overall performance, with accuracy of 0.8706, sensitivity of 0.8061, specificity of 0.9223, F1-score of 0.8471, ROC-AUC of 0.9366, PR-AUC of 0.9110, and MCC of 0.7384. Model stability was confirmed through bootstrap and calibration analyses, and permutation importance showed that oxygen-saturation, heart-rate and respiratory features were important. Preserved discrimination was demonstrated through robustness testing in the case of packet loss and sensor noise. Favorable latency-bandwidth trade-offs and 38.0% reduction in mean waiting time for future-unstable episodes was further demonstrated in simulated Edge–Cloud–Hybrid deployments, as well as risk-prioritized scheduling. Clinical validation (outside the lab) is still required prior to deployment in the real world.
Tiny Machine Learning (TinyML) enables resource-constrained devices to perform local inference in the Internet of Vehicular Things (IoVT), potentially reducing communication overhead, latency, and dependence on cloud services. This survey systematically reviews 57 primary studies covering vehicular applications, embedded hardware, lightweight learning models, optimization techniques, and deployment practices. The reviewed studies are organized using a seven-dimensional literature taxonomy covering deployment location, data modality, learning task, learning paradigm, optimization strategy, application domain, and system objective. Separately, their evaluation practices are critically examined through predictive performance, memory usage, inference latency, energy consumption, communication requirements, trustworthiness, and deployment evidence. The taxonomy characterizes existing research, whereas the benchmarking dimensions support comparison of reported experimental results and deployment limitations. The survey also examines computation–communication trade-offs, V2X interaction, collaborative learning, security, privacy, robustness, and functional safety. The findings reveal substantial heterogeneity in datasets, hardware platforms, evaluation protocols, and resource measurements, limiting direct cross-study comparisons. Moreover, experimentally demonstrated embedded feasibility does not necessarily establish reliability under realistic vehicular conditions or compliance with functional-safety requirements. The survey concludes by identifying research gaps in standardized benchmarking, hardware-aware optimization, adaptive learning, real-world validation, and trustworthy distributed vehicular intelligence.
The rapid development of Artificial Intelligence (AI) is reshaping pedagogical practice in Physical Education (PE), yet evidence on how AI functions pedagogically and what impedes its implementation remains fragmented. This study conducted a Systematic Literature Review (SLR) following the PRISMA 2020 guidelines using the Scopus database. Of 2,096 records initially identified, 400 were screened and 27 peer-reviewed studies published between 2020 and 2025 met the eligibility criteria and were synthesized thematically. AI applications in PE clustered into four categories: activity monitoring through wearable and IoT sensors, motor skill assessment using computer vision and machine learning, personalized learning through adaptive systems, and AI-enhanced gamification. These applications were associated with higher student motivation, more objective assessment, and more responsive instruction. However, the review identified four specific and interrelated implementation challenges: (1) limited teacher readiness, reflecting gaps in data literacy and in the pedagogical competence needed to interpret algorithmic feedback; (2) infrastructural and cost barriers, including dependence on connectivity, sensors, and technical maintenance, which disadvantage resource-constrained schools; (3) ethical and data-governance risks related to the collection of physiological and behavioral data, consent, and algorithmic bias; and (4) an evidence gap, as most studies were short-term, conducted in higher education, concentrated in a few national contexts, and rarely examined long-term learning outcomes. The review offers an integrative account of AI-driven pedagogical transformation in PE and argues that its effectiveness depends on the alignment of technological capability, teacher competence, and institutional support. Practical implications include content-specific teacher professional development, equitable infrastructure provision, and explicit data-governance policies.
Background: The rapid proliferation of Internet of Things (IoT) devices over MQTT-based infrastructures has significantly expanded the attack surface of cyber–physical systems. Reliable intrusion detection systems (IDS) are required to distinguish benign traffic from diverse attack patterns under realistic network conditions. Problem Statement/Research Gap: Most existing MQTT-based IDS studies rely on simulated datasets, limited attack taxonomies, or conventional machine learning models that are evaluated only under clean conditions. The robustness of deep architectures, particularly under noisy or perturbed inputs that approximate real deployments, remains insufficiently explored. Methodology: This work investigates IoT intrusion detection on the real-world MQTTEEB-D dataset, comprising 634,644 records and 6 traffic classes derived from MQTT/TCP flows. After SMOTE-based class balancing and Z-score standardization, the data are split into 70% for training, 15% for validation, and 15% for testing. The study benchmarks Gaussian Naive Bayes, a regularized Decision Tree, Random Forest, XGBoost, a ReLU-based multilayer perceptron (MLP), and a deep neural network with Stochastic Local Winner-Takes-All (S-LWTA) layers. In addition to standard test performance, robustness is assessed by injecting Gaussian noise with increasing standard deviation into the test features. Results: On clean data, XGBoost achieves the best performance with test accuracy 0.8858 and Macro-F1 0.8806, closely followed by Random Forest (accuracy 0.8827, Macro-F1 0.8809). The proposed Deep S-LWTA network outperforms the plain MLP (accuracy 0.6080 vs. 0.5770; Macro-F1 0.6012 vs. 0.5807), but remains below the tree ensembles in absolute accuracy. Under strong noise (σ = 0.10), Deep S-LWTA becomes the most robust model, retaining accuracy 0.4661 and Macro-F1 0.4532, whereas XGBoost and Random Forest drop to accuracies of 0.2606 and 0.2955, respectively. Conclusion: Real-world MQTT traffic analysis confirms that tree-based ensembles are the most accurate models for clean IoT intrusion detection, while the proposed S-LWTA architecture provides a competitive deep baseline with markedly better robustness to feature perturbations. These findings position S-LWTA as a suitable neural backbone for future work on adversarially robust IoT IDS across multiple datasets.
One of the most common chronic illnesses is diabetes, which requires ongoing observation,individualized care, and quick action to avoid complications. The real-time monitoring ofphysiological indicators has advanced thanks to recent advancements in Internet of Things(IoT) technology, including wearable sensors, smart insulin pumps, continuous glucosemonitoring devices, and mobile healthcare apps. Large language models (LLMs) are now apotent tool in the artificial intelligence toolkit for decision support, clinical reasoning, naturallanguage comprehension, and healthcare communication. Thanks to the power of both IoTand LLMs, there are plenty of possibilities for smart diabetes care, and much of the data fromsensors can now be used to inform healthcare insights. The authors have conducted aliterature review on the topic of diabetes monitoring systems with IoT technologies, as wellas machine learning in predicting glucose levels and the growing role of LLMs in healthcare.The goal of this paper is to provide a summary of the state of the art in machine learningmethods for glucose prediction, IoT-based diabetes monitoring systems, and the expandingrole of LLMs in healthcare. The most recent developments in diabetic treatment using LLMare covered in this article, including patient interaction platforms, clinical summaryautomation, conversational healthcare assistants, and tailored recommendation systems. Anexamination of existing research is done to identify trends, strengths and weaknesses. Inaddition, issues concerning with privacy, security, explain ability, model reliability andethical aspects are discussed. The research gaps are identified, and the future directions areproposed so that motivating intelligent, scalable and patient-centric diabetes managementsystems are developed. Overall, the findings indicate that the potential for IoT and LLM towork together in healthcare is enormous, with the ability to improve treatment adherence,proactive management of diseases, and better healthcare outcomes.
Falls can cause serious injury when assistance is delayed. This paper presents a hardware demonstration prototype using an ESP32 DevKit V1 and MPU6050 motion sensor to identify motion patterns associated with a possible fall. The implemented algorithm applies sequential thresholds to resultant acceleration and gyroscope angular-velocity magnitude; no trained machine-learning model is claimed. Hall-effect and capacitive inputs gate monitoring. When the programmed conditions are satisfied, the controller updates an LCD, activates a buzzer, checks for a recent valid GPS location, and attempts SMS transmission through an A7670C cellular modem. Functional development tests observed ESP32 serial output, Wi-Fi connection, GPS data, fall-sequence status, SMS submission, and an SMS-received indication. Detection accuracy, false-alarm rate, and delivery timing have not been quantified through repeated controlled trials.
UrbanGlow is an intelligent, AI-powered streetlight management system designed to improve urban lighting efficiency, optimize energy consumption, and support smarter streetlight operation by combining environmental data, traffic patterns, weather conditions, and machine learning. Conventional streetlight systems often operate at fixed brightness levels, regardless of changes in traffic, ambient light, or weather conditions. This can lead to unnecessary energy consumption, inefficient lighting, and limited visibility into streetlight performance. UrbanGlow addresses these challenges through an integrated approach that uses data-driven predictions to support adaptive lighting decisions and energy-efficient management. The system incorporates five major modules: Traffic Prediction, Brightness Optimization, Weather-Aware Lighting, Energy Consumption Forecasting, and Monitoring and Dashboard. The Traffic Prediction module analyzes historical traffic patterns and time-related information to estimate traffic levels. The Brightness Optimization module determines suitable lighting levels using factors such as ambient light intensity, traffic density, and time of day. The Weather-Aware Lighting module considers weather conditions when recommending lighting adjustments, while the Energy Consumption Forecasting module estimates energy usage based on brightness levels, operating hours, and traffic patterns. These modules are integrated through a Python-based backend that loads trained machine learning models and processes input data, while a React-based web dashboard presents current conditions, predicted traffic, recommended brightness levels, energy forecasts, and system status in an accessible interface. The IoT-oriented architecture supports the integration of environmental sensors, traffic detection devices, and connected controllers for streetlight monitoring and control. By combining predictive analytics with centralized monitoring, UrbanGlow aims to help improve lighting decisions, reduce avoidable energy consumption, support maintenance planning, and provide a scalable foundation for intelligent streetlight management in smart cities.
Implementing analog reservoir computing (RC) at the edge is fundamentally constrained by conflicting hardware requirements for the reservoir (volatile, nonlinear) and readout (nonvolatile, linear) layers. Here, we resolve this dilemma by monolithically integrating a multi-timescale photoelectric reservoir and a linear ferroelectric readout within a single vertical organic field-effect transistor (Fe-vOFET). This architecture functionally decouples photocarrier dynamics from ferroelectric polarization, allowing orthogonal optimization of reservoir dynamics and synaptic weights. In the optical mode, the Rubrene/F16CuPc heterojunction underpins reservoir dynamics, exploiting wavelength-dependent relaxation kinetics to construct multi-timescale grouped reservoirs that project inputs into high-dimensional spatiotemporal states. Conversely, in the electrical mode, the metal–ferroelectric–metal–insulator–semiconductor (MFMIS) stack governs readout functionality by mitigating interface trapping, translating remnant polarization into linear, symmetric, nonvolatile weight updates (endurance >106, retention >3000 s, asymmetry nonlinearity 0.31 ± 0.05). The vertical-channel geometry is pivotal, simultaneously amplifying optical feedback and boosting the on/off ratio to ~107, which is essential for high-fidelity weight mapping. Leveraging this dual-mode operation, we demonstrate a compact RC system achieving >90% accuracy in dynamic human-action recognition. This work establishes a scalable device paradigm that unifies dynamical richness with programming precision, paving the way for energy-efficient, all-in-one dynamic machine vision systems. A dual-mode Fe-vOFET integrates multi-timescale photoelectric reservoirs and linear nonvolatile ferroelectric readout, enabling compact, efficient dynamic machine vision with high recognition accuracy.
Although time-series foundation models (TSFMs) achieve strong zero-shot transferability across high-information-density physical systems, applying overparameterized TSFMs to quantitative finance is fundamentally constrained by low signal-to-noise ratios, non-stationarity, and prohibitive autoregressive decoding latency. To reconcile representation learning with institutional operational constraints, this study presents TailDART, a dense 2.68 million-parameter foundation model engineered for financial time-series forecasting. By integrating grouped-query attention into a lightweight backbone, the architecture drastically reduces Key-Value cache memory bandwidth overhead, achieving major inference acceleration (up to 48.1%) over prevailing state-of-the-art TSFMs to enable low-latency, high-throughput streaming execution. Furthermore, multi-asset benchmarks across varying temporal resolutions show that a compact model can remain competitive for market forecasting: TailDART matches the predictive fidelity of substantially larger baselines while demonstrating superior error resilience in high-volatility regimes.
3D spatial reasoning from video remains challenging for VLMs because videos provide sequential, view-local observations rather than persistent metric scene records. Many questions must be answered in a query-specified perspective rather than the observed camera viewpoint, and prompting with a fixed set of RGB views may miss sparse or occluded visual cues. To address these gaps, we introduce BEV3SR, a training-free and backbone-agnostic framework for RGB-D videos with camera parameters. BEV3SR turns video observations into queryable metric bird's-eye-view (BEV) scene memory and formalizes each query into an evidence-aware spatial task. The Spatial Task Formalizer specifies the query's reference frame, spatial constraints, objective, and evidence requirement. The evidence requirement determines the appropriate reasoning pathway. A query treated as requiring geometric evidence is handled by the Geometric Reasoner, which deterministically computes the answer using the metric BEV records in the query-specified reference frame. Otherwise, a query treated as requiring visual evidence is passed to the Visual Evidence Aligner, which uses BEV-indexed object records and their associated supporting views to retrieve object-consistent and scene-context RGB frames for VLM grounding. Within a standard 16-frame budget, BEV3SR consistently improves different VLM backbones across multiple 3D scene QA benchmarks, delivering large gains on spatially intensive questions and significant improvements on open-ended scene QA. These results demonstrate that BEV3SR offers an effective, efficient, and interpretable framework for explicit 3D spatial reasoning from video.
Zero-knowledge proofs allow one party to prove the correctness of a statement or execution of a computation while revealing no more than the statement of correctness itself. Recent use cases, such as AI datasets, suggest the need for faster proof generation over increasingly large constraint sets. This paper explores efforts in optimizing zero-knowledge proof generation, specifically PlonK, using GPUs, for applications that are not amenable to solution via incrementally verifiable computing (IVC). Our intention with this work is to improve ZK performance by introducing fundamental changes in the low-level architectural structure of proof generation. These changes affect performance curves for most current proof generation algorithms. In particular, this paper explores the trade-offs between increased parallelization on GPUs and the resultant memory management problems that occur due to this parallelization. We implement a parallelized PlonK in order to analyze these trade-offs and make arguments concerning the nature of data flows within the proving protocol. Our results lead us to consider multi-GPU environments to maximize performance for large circuit sizes. Such environments introduce issues of communication, efficient hardware management, data partition, and data transfers. This paper further discusses the use cases that motivate multi-GPU implementations, their cost models, and the use of remote direct memory access (RDMA) for data transfer among GPUs.
Quantization-aware training (QAT) is the most reliable way to obtain accurate low-bit large language models (LLMs), but training with AdamW keeps two dense FP32 states for every weight matrix. We find that low-rank memory-efficient optimizers such as GaLore and LDAdam degrade substantially in QAT. A considerable part of the QAT momentum lies outside its top-r singular subspace, and projection-based methods discard this part at every step. We further observe that this residual remains useful when only its sign is kept. Based on these observations, we propose ResQAdam. It keeps a single BF16 momentum, applies Adam only to the r x r coordinates of the principal subspace, and adds a normalized sign update computed from the residual. We evaluate ResQAdam with two from-scratch QAT methods and three QAT fine-tuning pipelines, with weights quantized down to 2 bits. It reduces optimizer-state memory by 71% to 74% relative to AdamW, stays close to AdamW in most settings, and lowers training throughput by less than 2%.
Efficient encoding of data into quantum states remains a fundamental bottleneck in many quantum computing protocols, which would otherwise offer advantages over their classical counterparts. Here, we introduce deterministic and scalable amplitude-encoding methods with low two-qubit gate costs and tunable trade-offs between circuit depth and encoding fidelity. Our approaches require no amplitude amplification or mid-circuit measurement and are capable of encoding arbitrary complex data or queryable functions by using tensor-networks to go beyond what can be held in classical memory. Our principle contributions are two complementary algorithms for simplifying fully controlled multiplexer circuits to one- and two-qubit gates. In the first, we formally map the elimination of redundant gate controls to a Boolean logic minimization problem, while in the second we utilize a single ancilla qubit to map the multi-controlled rotation gates to a sequence of one- and two-qubit operations via the Walsh-Hadamard spectrum of the gate angles. We benchmark these methods across a broad range of function classes, characterizing their total depth in an all-to-all connectivity graph against system size and target fidelity. Hybridizing our methods gives up to orders of magnitude improvement in circuit depth at equivalent state fidelity compared to the default state-preparation method in the Qiskit package. We furthermore demonstrate our method on IQM hardware with a square lattice connectivity, finding excellent fidelity, even with minimal noise mitigation. By significantly reducing gate depth for encoding, this work enables the experimental execution of applications on current and near-term hardware which would otherwise be reduced to noise before the first calculation step, shortening the time horizon to robust quantum utility.
A game may have, for each target below its value, a finite controller meeting that target, yet no finite controller attaining the value itself. We study finite stochastic games with a hard specification that must hold on every supported execution, including paths of probability zero. Max uses a pure finite controller; Min may use any pure public-history strategy. We distinguish three numerical questions: can a supplied target be met, can the unspecified optimum be attained, and what is one exact value? For explicit hard and Boolean soft Streett conditions, the strict and non-strict supplied-threshold problems are coNP-complete, attainment of the unsupplied supremum is $\mathrm P^{\mathrm{NP}}$-complete, and one exact reduced fraction is $\mathrm{FP}^{\mathrm{NP}}$-complete under polynomial metric reductions. The proofs preserve both the value and its finite attainability when fixing a single opponent, and encode adaptive queries without computing unknown subgame values. The same construction yields matching value-comparison and exact-output lower bounds for ordinary stochastic Streett games with unrestricted randomized history strategies. We also bound the memory needed for optimal finite implementations for arbitrary Muller hard conditions and rational rewards in $[0,1]$ determined by infinitely recurring states. One controller is sure from the entire original sure region and attains at every finitely attained start, with no attainment premise on the other states. For $n$ states in that region it uses $2^{O(n\log(n+1))}Q(1+\log R)$ modes, where $Q$ is the product of the least common denominators of full original chance rows sourced in that sure region and $R$ clears all reward values. Its joint dependence on $Q$ and $R$ is worst-case sharp at each fixed $n\ge6$ when $Q,R\ge2$. Optimal dependence on arena size and efficient general synthesis are not resolved here.
Test-time adaptation (TTA) for vision-language models (VLMs) is usually studied in a fully unsupervised regime, where adaptation relies on pseudo-labels, confidence scores, or entropy computed by the model itself. This is a severe restriction on hard shifted samples: the model is asked to repair its own uncertain predictions without any external evidence. At the other extreme, supervised online adaptation assumes labels can directly train or correct the model. We study the intermediate deployment regime between these two extremes. In Active Test-time Adaptation for VLMs (Active VLM-TTA), a pretrained VLM may spend an explicit labeling budget on selected test samples, but the prediction for the current sample must be emitted before its label can be used; queried labels are therefore costed, delayed evidence for future samples rather than free correction for the present one. Under this future-only constraint, the central question is how to reuse sparse labels conservatively. We instantiate this setting with ACTive VLM-TTA for Online Retrieval (ACTOR), a training-free plug-in module that routes uncertain samples among base prediction, label query, and memory-based correction using a consistency-and-reliability check over previously queried labels. Across ten cross-domain benchmarks and a domain-transfer benchmark based on ImageNet, ACTOR consistently improves strong VLM-TTA baselines on ResNet and ViT backbones with negligible inference overhead. Our implementation is publicly available at https://github.com/kaist-dmlab/ACTOR.
Attention is a major scalability bottleneck in large language models (LLMs), with its computational cost and key-value (KV) memory traffic becoming increasingly prohibitive as context length grows. Sparse attention addresses this bottleneck by restricting each query to a subset of the context, but its effectiveness hinges on identifying relevant regions both cheaply and reliably. Existing training-free methods primarily rely on runtime relevance estimates, with selection dimensionality and granularity typically governed by fixed heuristics rather than the intrinsic interaction structure of pretrained attention heads. We take a different perspective: the query-key operator of each pretrained head defines a characteristic response structure that determines the resolution needed for sparse selection, both across feature directions and across context tokens. Building on this insight, we propose Head-Adaptive Sparse Attention (HASA), a training-free sparse attention framework that jointly adapts selection resolution along these two dimensions. Along the feature dimension, the singular spectrum of each head's query-key operator determines an effective rank, yielding a compact, head-specific response space. Along the sequence dimension, neighboring regions are progressively merged according to the directional coherence of their token responses in this same space, while heterogeneous regions remain separate. Exact attention is then computed using the original full-dimensional KV states of the selected regions. HASA captures 88.1% of dense-attention mass and, compared with UNIQUE, a strong recent sparse-attention baseline, improves accuracy by up to 3.47 and 2.57 points on RULER and LongBench-Pro, respectively, and accelerates self-attention by up to 1.32x, while achieving up to 14.07x speedup over dense attention.
Graph similarity computation is an important task in graph mining with many real-world applications, such as chemical compound analysis, program dependence comparison, and social network analysis. The problem is computationally challenging: Graph Edit Distance (GED) and Maximum Common Subgraph (MCS), which quantify the distance between graphs and identify their largest shared subgraph, respectively, are NP-hard to compute exactly. To address these computational challenges, Graph Neural Networks (GNNs) have been applied to this problem. However, existing GNN-based approaches have primarily focused on small-scale graph datasets, while scalability to larger scales remains underexplored. In this work, we propose FiLSim, a scalable GNN-based framework, and investigate whether a simplified GNN pipeline can maintain competitive performance on standard benchmarks while scaling to large graphs. We additionally propose a sequential graph-editing strategy to construct large-scale graph datasets with reliable GED labels to evaluate scalability. The proposed framework is evaluated on four benchmark datasets and four newly created large-scale datasets, demonstrating its effectiveness and scalability. FiLSim remains accurate on graphs with more than 150{,}000 nodes, where the baseline methods run out of memory or fail to learn, and scores all test pairs within seconds on a single GPU.
Designing power electronics converters traditionally requires intensive human expertise across topology sizing, commercial component selection, circuit simulation verification, and schematic drafting. While Large Language Models (LLMs) show promise in electronic design automation (EDA), applying standalone LLMs directly to board-level power electronics schematic synthesis suffers from severe engineering hallucinations, including inaccurate analytical sizing, fabricated or mismatched component part numbers, violated voltage derating margins, and broken topological connectivity. In this paper, we propose a Bounded Dual-Loop Automated Schematic Synthesis Framework that translates natural language specifications into verified, editable KiCad schematics and manufacturing deliverables. Our workflow integrates six stages across three functional columns: (1) requirement parsing and knowledge-driven power stage synthesis covering non-isolated (Buck, Boost) and isolated/resonant (Flyback, DAB, half-bridge LLC) topologies; (2) contract- and evidence-coordinated component selection grounded in distributor APIs and OEM datasheets, coupled with closed-loop ngspice transient verification and stress feedback; and (3) modular control hardware integration and multi-format deliverable generation featuring native KiCad netlist round-trip verification. End-to-end case studies on a 24-36V to 12V,2A Buck converter (29 positions, 83 verified terminals) and an isolated 48V to 24V, 0.5A, Dual-Active Bridge (DAB) converter (87 positions, 341 verified terminals), alongside comparative evaluations against a Vanilla LLM baseline, demonstrate that our framework improves component verification coverage from 22.2 to 98.6, nominal value consistency from 0 to 97.7, and SPICE simulation pass rates from 0 to 100.
In contrast to a classical bit, which can only take the value $0$ or $1$, its quantum counterpart---a qubit---can exist in a superposition of $0$ and $1$. This is a superposition of data values, and it is then tempting to superpose programs as well. Indeed, a particular superposition of programs, known as the quantum SWITCH, has attracted considerable attention. A natural way to superpose programs is conditional branching on quantum data, and several studies have pursued this idea. However, there is a well-known semantic difficulty with this approach, and the fundamental question of whether the languages proposed in prior work can be physically realised remains open. This paper proposes a physically realisable quantum programming language with quantum branching over quantum channels. We first analyse the semantic difficulty from an implementation perspective and show that the difficulty is caused by arbitrariness in the implementation. We then design a language that is free from this arbitrariness. As expected from its implementation-oriented design, our language is physically realisable. We argue for this from two directions: first, we give a method for compiling programs into quantum circuits; and second, we show that our programs stay within the physically admissible range, namely, that they neither yield probabilities greater than one nor violate causal inequalities. The absence of such violations is established through the inclusion of our language in QC-QC, a notion of realisability studied in physics.