On-Device Inference

Activation memory, not compute, limits CNN inference on constrained hardware such as microcontrollers. Direct in-place convolution removes the dual-buffer cost, but the memory-optimal formulation of Gural and Murmann assumes valid padding, unit stride, unit dilation, and odd square kernels, and needs a non-sequential traversal costing $2\times$ inference time in transposes. We identify two regimes in which their published closed form does not hold: (1) an under-allocation of exactly $(k-1)C_{in} \bmod (C_{out}-C_{in})$ scalars, active on every convolutional layer of their own deployed network and manifesting as a silent corruption of still-live input; (2) an unbounded overestimate, up to $2{,}432\times$, once the critical leg leaves the output grid. We correct both and generalize to arbitrary stride, dilation, padding, and rectangular kernels. We then propose Right In-Place (RiP) convolution, a bit-identical operation in which every layer reads its input right-aligned in a shared workspace and writes its output left-aligned from index zero. The debt is piecewise affine in the output pixel index, so evaluating its breakpoints in $O(1)$ yields the minimum safe gap without enumerating the output grid, with row-major access preserved. Across $10{,}000$ random layers RiP produced no corruption, and across 84 convolutional layers from 25 architectures it matches the herringbone workspace exactly on 58 and within 5% on 81, using 24.8% less memory than dual buffering on average. Written into TinyEngine's kernels and deployed to a Raspberry Pi Pico 1 and Pico 2, it cuts peak activation memory across eleven MCUNet models by 12.5 to 33.3% at unchanged cycle counts and bit-identical outputs, raising the number of models that fit the Pico 1's 256 KB SRAM from six to nine.

Passive Acoustic Monitoring of bats generates massive ultrasonic datasets (>27 GB/night per node), straining edge storage and battery life. Legacy triggers fail against acoustic confusers, while deep models exceed microcontroller limits. We present a hardware-aware Bat Activity Detector (BAD) specifically designed to discriminate bat calls from hard biological and environmental confusers across variable sampling rates (192-384 kHz). Tailored for the Silicon Labs EFM32PG26 (MVP) in 8-bit integer precision, our model achieves 100 percent hardware offload across all 14 layers (17.2 KB Flash, 73.1 KB RAM). End-to-end preprocessing (74.00 ms for 76 frames) and inference (30.00 ms) of 100 ms clips at 192 kHz require 104.00 ms per clip. On spatially out-of-domain recordings under a realistic low-prevalence regime (r_pos = 0.05), BAD achieves an AUC-ROC of 0.9748 and suppresses 99.4% of non-target noise frames while retaining 65.3% of bat calls - delivering a >33x precision gain over classical Goertzel baselines.

Video diffusion models are rapidly scaling and exhibiting enhanced generation capabilities. Among these recent advancements, MiniMax-H3 stands out as a highly capable, production-level open-source model. However, its 33-billion parameters and multi-step iterative denoising process introduce substantial computational overhead. Consequently, their practical production is hindered by generation latency in the cloud deployment like NVIDIA-GB200, alongside strict memory limits that pose further challenges at the edge device like DGX-Spark. To address these diverse hardware bottlenecks from cloud to edge device, we present a full-stack inference pipeline that integrates efficient algorithmic design with optimized operator implementations. Algorithmically, we introduce a cross-resolution two-stage generation scheduler that exploits the step-wise nature of diffusion: early low-resolution steps rapidly establish the global layout, while later high-resolution steps focus refinements of local and perceptual details. These stages are connected by a learned latent-to-latent mapping module, completely eliminating the computationally expensive VAE decode-reencode cycle for resolution transferring cross different resolutions. For operator implementation, we deploy a Recursive Self-Improvement (RSI) loop that searches kernel fusions and memory layouts, evaluating latency together with numerical agreement. Together, these optimizations deliver up to 30x end-to-end speedup and 20% lower memory: a 5-second 1344x768 video with audio is generated 3.5x faster than real time on an 8xGB200 node, and in under a minute fully memory-resident on a single DGX Spark.

Keyword Spotting (KWS) is becoming increasingly important as voice-controlled devices grow more widespread. While voice interaction with smartphones and smart TVs is already common, deploying KWS on heavily resource-constrained edge devices such as wearables remains challenging. These systems must meet high accuracy requirements while operating under strict constraints on computational power, memory footprint, and real-time latency. In this work, we present an application of the STMC (Short-Term Memory Convolutions) framework to adapt a modular CNN model for online, LSTM-like inference. Our approach reduces power consumption and redundant computations while maintaining the stability and simplicity of training CNNs. We achieve up to 82% and 46% MCPS reduction compared to equivalently frequent standard CNN execution and vanilla STMC, respectively. The best configuration achieves 93.8% accuracy on the 11-class Google Speech Commands task and 97.1% on the same task with zero-padded data.

Deep learning-based speech enhancement is increasingly deployed on-device in hearing aids, headsets, and earbuds. Most of these devices, however, can only accelerate static int8 graphs, so a depth-varying network must be implemented as several graphs, orchestrated by a policy. In this paper, we supervise every intermediate depth of one causal model, then we fine-tune its output heads to guarantee that deeper outputs are never worse than shallower ones. Using this training protocol, we can derive a family of static models that are more Pareto-efficient than their equivalently-sized counterparts trained from scratch on the same budget. Specifically, we achieve up to 0.11 higher PESQ for equivalent compute, and match the best PESQ at 30% less compute. We then quantize the models to int8 and measure the latency-quality frontier on an STM32N6 microcontroller. On VoiceBank-DEMAND, the dynamic enhancer lies on the same frontier as the static models, rather than trading quality for dynamic execution. Running the policy on the companion Cortex-M55 takes only 26 $μ$s per frame, while splitting the enhancer into separate NPU graphs adds 2.2% latency overhead. The cost of dynamic execution is therefore small.

This paper presents an Edge AI-based system for detecting sleep and wake states in non-stationary mobile environments using resource-constrained embedded hardware. Conventional approaches relying on accelerometer-based activity metrics are highly susceptible to motion and vibration artifacts and are limited by strict compute and energy budgets of wearable and IoT devices. To address these challenges, a multimodal pipeline is designed and implemented on an ESP32-S3 microcontroller. The system combines inertial sensing for head movement analysis and visual pose classification. A dual-core architecture with FreeRTOS enables parallel execution of real-time data acquisition and on-device inference. Sleep detection follows a two-stage strategy: low-movement detection over a temporal window, followed by visual validation of poses. Experimental results show accuracies of 96.5% for motion-based detection and 89% for pose classification, yielding robust binary sleep-wake classification. Field tests confirmed feasibility in representative mobile scenarios. The results demonstrate that privacy-preserving, local sleep detection is achievable on edge hardware through careful co-design, while highlighting limitations in sensing intrusiveness, dataset scale, and system integration.

Data acquisition (DAQ) systems at future particle physics experiments stand to benefit from the extremes of AI/ML development: large-scale foundation models can enhance the performance of feature extraction algorithms, and small-scale on-detector deployments can enable real-time intelligent data handling. This work provides the first fine-tuning of an industrial foundation model for particle physics DAQ. Starting from the backbone of Google Research's TimesFM (Time Series Foundation Model), we demonstrate fine-tuning on real-time regression tasks for drift chamber trackers and dual-readout calorimeters. Furthermore, the fine-tuned TimesFM model is distilled into a student and co-designed with FPGA implementation to enable these models to run in real-time at future colliders. The fine-tuned distillations meet or exceed the performance of previously published AI/ML solutions for each task. Further, the pipeline of distillation and model compression from TimesFM is generic and can be easily adapted to a variety of 1D waveform tasks across domains.

Billion-parameter vision--language--action (VLA) policies demand substantial onboard power, while communication delays in remote inference hinder timely responses. We propose VLA-ULAP, which interleaves remote VLA calls with an Ultra-Lightweight Local Action Predictor (ULAP). With approximately 7.4M parameters including the frozen vision encoder, ULAP combines current views, proprioception, and executed action history to predict chunks in one pass. Trained independently, it requires no VLA hidden states, online verification, or server round trips. On Jetson Orin Nano, ULAP takes 19.9 ms and 0.183 J per inference, compared with 284.3 ms and 50.55 J for GR00T on RTX A6000. Across three simulated base-policy/benchmark pairs, selected operating points remove 48.8--76.7% of VLA calls while retaining 95.0--97.5% of the baseline success rate. Against local VLA-acceleration alternatives on VLA-JEPA, ULAP uses an estimated 49.2% less inference time and 51.0% less GPU energy per successful episode than ACT at comparable success rates, and 77.1% less time and 79.9% less energy than SP-VLA at equal success rates. Physical SO-101 experiments retain 95.2--100% of the baseline success rate across seen and held-out placements while reducing inference time by an estimated 47.9--58.0% and inference-device energy by 52.1--62.5%, based on successful-episode call counts and measured device costs. Faster responses also improve dynamic-task success rates: in latency-aware LIBERO-Safety simulation, VLA-ULAP exceeds $π_{0.5}$ by 11.0 and 15.5 percentage points on two tasks while approximately halving VLA calls.

As deep learning models are increasingly deployed on resource constrained devices, the demand for efficient model optimization techniques continues to grow. Effective deployment of AI models on edge and low power platforms requires optimization methods that reduce model size and computational cost while maintaining high accuracy. This paper presents Quantization Analysis Tool, a practical system designed to streamline quantization workflows and support performance efficient model deployment. Built on the ONNX framework for broad interoperability, the tool provides detailed layer-wise sensitivity analysis, visualization of weight and activation distributions, and insights to guide precision selection. By identifying layers that are resilient or sensitive to reduced precision, the tool enables developers to make informed trade-offs between model size, latency, and accuracy. Experimental evaluations across multiple neural network architectures demonstrate that the tool effectively improves the quantized accuracy, leading to improved efficiency in real-world deployment scenarios. The tool also provides developers valuable insights into the effects on quantization on the model and its accuracy. This work highlights the tools capabilities, practical applications, and its role in enabling efficient AI model deployment through robust quantization analysis

Affective computing has advanced wearable state inference, but on-device reasoning about whether, when, and how to intervene remains challenging. We present Affective Agent, a three-layer reference architecture for personalized intervention reasoning under uncertainty on wearable-class hardware. It combines a compact sub-billion-parameter language model with physiological evidence, context, and user history to decide whether, when, and how to intervene, without cloud dependency or per-user retraining. The architecture is organized into three interacting layers (perception, personalization, and reasoning), adapting to individual users through host-managed structured memory evolution rather than per-user weight updates. We instantiate Affective Agent in indoor environmental quality control and evaluate it on held-out, simulator-generated longitudinal scenarios spanning physiological variation, context, signal quality, and intervention history. Results show that memory-driven personalization and two-pass structured reasoning improve intervention decisions within this synthetic evaluation. By moving the decision layer on-device, this work demonstrates a path from wearable state inference toward closed-loop, personalized intervention on wearable-class hardware.

Foundation models, including large language models, vision-language models, and time-series foundation models, are increasingly deployed on embedded and edge platforms for CPS and IoT applications, where energy, latency, and memory are as critical as task accuracy. Existing benchmarking tools evaluate model capability in isolation, reporting accuracy assuming sufficient compute, while hardware profiling tools remain platform-specific and mutually incompatible. As a result, users lack a unified workflow for making deployment decisions across heterogeneous devices. We present HoliBench, a modular benchmarking and deployment toolkit that jointly characterizes accuracy, latency, and energy across platforms from single-board computers to GPU servers. Its platform abstraction layer calibrates cross-device measurement, and the toolkit supports multiple model modalities, inference engines, concurrencies, and existing evaluation harnesses. An interactive interface exposes constraint-aware configuration selection over a design space that is profiled once and reused across studies. Using HoliBench, we characterize 20 models across 7 device types, 3 quantization levels, 8 inference backends, and over 30 tasks, surfacing tradeoffs that existing tools miss: quantization reduces latency only on hardware with low-precision support, accuracy gains show diminishing returns relative to energy, and for autoregressive workloads, average inference power is approximately constant across output lengths. We further find that single-model profiles compose under sequential co-resident execution. In a multi-model CPS deployment, standalone profiles predict combined-pipeline latency and power within 1.2% and 2.5%, enabling deployment exploration without exhaustively profiling every pipeline configuration. We release HoliBench as open-source infrastructure for deployment-aware evaluation of foundation models.

Vision models deployed on microcontrollers (MCUs) are quantized to integer-only arithmetic and run in inference-only runtimes that do not carry the machinery backpropagation needs: the standard tool for adapting a model to the distribution shift (sensor noise, blur, lighting) it meets in the field. Existing forward-only test-time adaptation (TTA) methods either run only on server- or edge-GPU-class models (not true microcontroller integer execution), or require the batch-normalization (BN) layers that integer deployment fuses away. We present a forward-only TTA method that operates on deployed, BN-folded, integer-only convolutional networks. The key observation is that fusing BN into the preceding convolution, a mandatory step for integer inference, destroys the statistics that normalization-based adaptation relies on. We restore adaptation by re-normalizing each folded convolution's per-channel output to its clean training statistics, using only forward-pass estimates. The method (i) recovers most of gradient-based TENT's accuracy gain (+20.9 vs. +24.9 points) and matches forward-only BN adaptation, while being the only method that runs on a folded integer-only model; (ii) needs to adapt only 3 of 21 layers (selected without seeing the test corruptions) to recover 93% of the benefit; (iii) survives single-sample streaming with a batch-size-scaled momentum; and (iv) generalizes across three datasets (up to 200 classes) and two architectures. We validate bit-exact int8 convolution execution and deploy on an ESP32-S3, where, measured with a Nordic PPK2 power profiler, the forward-only adaptation (a lightweight fp32 recalibration around the int8 convolutions) costs only 8.3 mJ (6.8% of inference energy) and 21.9 ms on the deployed SIMD-optimized model: forward-only adaptation is cheap on a real microcontroller.

Two-stage vision pipelines often place an enhancement network before a task network, on the assumption that a cleaner input produces a better output. We evaluate this in a firefighter assistance pipeline, where a dehazer precedes an edge detector that renders smoke-filled rooms as structural outlines. Both were designed for a Raspberry Pi 4, at 355K and 23K parameters, and quantized to UINT8 via TensorFlow Lite. The float dehazer reaches 18.60 dB peak signal-to-noise ratio (PSNR) on held-out real smoke against 13.60 dB unprocessed and 17.08 dB for an AOD-Net trained on the same data, and the edge detector reaches an F-measure at optimal dataset scale (ODS) of 0.738, outperforming an optimized Canny's result of 0.692. Dehazing improves edge extraction under dense smoke but degrades it on clear and lightly hazed frames, where the dehazer discards more detail than the haze obscures. We therefore run the dehazer only when a dark channel haze estimate exceeds a threshold, a 10.1 ms test that lets the pipeline save 469.6 ms on the dehazing stage. Averaged over four haze levels, gating is more accurate than either fixed decision, at 0.675 mean ODS against 0.664 for always dehazing and 0.630 for never dehazing. It reduces the mean per-frame time on the Raspberry Pi from 569 ms to 321 ms, and on clear frames increases the frame rate fivefold, from 1.8 to 9 frames per second.

Breast Cancer Multidisciplinary Team (MDT) meetings manage increasingly complex cases under considerable time pressure, and documentation requirements can reduce clinical efficiency and decision quality. Existing AI based MDT workflows rely on cloud-based processing, limiting their use because patient discussions contain identifiable information. We developed a fully on-device AI pipeline using open-source Automatic Speech Recognition (ASR) and Large Language Models (LLMs) that transcribes breast cancer MDT discussions, structures clinical information, and generates treatment recommendations using retrieval-augmented generation (RAG) grounded in National Institute for Health and Care Excellence (NICE) guidance. The pipeline runs on a single NVIDIA Jetson AGX Orin, ensuring that patient audio, transcripts, and outputs remain within institutional infrastructure. Evaluation included two recorded simulated MDT discussions, ten clinically validated synthetic discussions, and 1,270 acoustically augmented recordings. Optimisation of Whisper large-v3 reduced word error rate by 20.7% and 24.4% on the recorded discussions and achieved performance within 0.58% WER and 1.58% word information lost of a commercial clinical ASR benchmark on augmented audio. MedGemma-RAG identified 2.3 times more MDT-concordant interventions than a proprietary cloud comparator (p = 0.020), with no significant difference in overall accuracy. Stakeholders identified automated documentation, treatment recommendation support, and case triage as the most credible near-term applications while highlighting workflow integration, governance, and clinician trust as key implementation challenges. These findings demonstrate the feasibility of privacy-preserving, fully on-device AI for MDT documentation and guideline-informed decision support, providing a foundation for prospective clinical evaluation.

Automated plant species identification from citizen-science imagery is an established, demanding fine-grained recognition problem: large taxonomic label spaces, visually similar species, and long-tailed observations require real model capacity, while field use constrains memory, latency, and power. Model size is only part of the deployment cost: intermediate activations held in memory during inference and platformdependent execution behavior matter too, so compact recognition must be assessed on target hardware rather than through complexity metrics alone. We present BoltNet, an ultra-lightweight fully convolutional architecture combining a Spatial Redistribution Bottleneck and Logit PreSampling to improve the tradeoff between predictive performance and model size in high-cardinality classification, and report the AccuracyCompression Tradeoff as a complementary diagnostic. On Pl@ntNet300K, BoltNet reaches 0.682 F1-score with 341K parameters (1.37 MB), the highest F1-score among evaluated models below 2 MB and close to substantially larger convolutional backbones. Model-only measurements on a Raspberry Pi 5, Jetson Orin Nano, and Hailo-8 characterize execution across CPU, GPU, and NPU platforms, where BoltNet is the most consistently efficient model, with the best FPS/W on the GPU and NPU and second-best on the CPU. Results on AIDERv2 and CLRS provide secondary evidence of transfer across environmental image-classification tasks. Code available at: https://codeberg.org/danielrossi/BoltNet

Deploying three-dimensional deep learning frameworks to low-power embedded processors is bottlenecked by the unstructured nature of spatial data and the resource-intensive distance sorting algorithms often used before neural network inference. To address this gap, this paper presents a hardware-constrained workflow optimized for native execution on the Raspberry Pi 5. To account for the reality gap between noiseless, clean computer-aided design (CAD) datasets and real-world sensor data, we use physics-based simulation to construct a synthetic LiDAR dataset. Cross-dataset evaluations demonstrate a substantial drop in classification accuracy when networks trained on clean CAD data are evaluated on synthetic LiDAR sensor data, highlighting the critical need for sensor-aware training. To address the latency bottleneck of traditional geometric preprocessing on edge CPUs, we integrate an isolated, feature-driven Critical Points Layer (CPL) as a frontend filter. Our results show that the pretrained CPL deterministically compresses raw 1024-point clouds to a subset of 40 to 60 unique coordinates. When profiled on the ARM Cortex-A76 processor, the complete pipeline achieves an inference throughput of approximately 50 FPS while maintaining an instance classification accuracy of 88.36%, demonstrating the viability of deterministic real-time 3D perception at the edge.

Quantized Neural Networks~(QNN) with low-bitwidth data have proven promising in efficient storage and computation on edge devices. To mitigate accuracy degradation while maximizing speedup, layer-wise mixed-precision quantization~(MPQ) becomes a popular solution. However, existing algorithms for exploring MPQ schemes are limited in flexibility and efficiency. Comprehending the complex impacts of different MPQ schemes on post-training quantization and quantization-aware training results is a challenge for conventional methods. Furthermore, an end-to-end framework for the optimization and deployment of MPQ models is missing in existing work. To address these challenges, we propose the MiCo framework, a holistic MPQ exploration and deployment framework for edge AI applications. The framework adopts a novel optimization algorithm to search for accuracy-optimal quantization configurations under strict latency constraints. We further extended the framework to MiCoPro, which introduces a robust Hardware-Aware Proxy (HAP) model to enhance prediction accuracy and hardware versatility. By leveraging target-specific latency modeling, MiCoPro enables rapid exploration and direct deployment from PyTorch models to bare-metal C code. We demonstrate the versatility of our framework on both the BitFusion accelerator and SIMD-extended RISC-V processors, achieving up to 40% of latency reduction with less than 3% of accuracy drop.

Mobile image denoising requires both good restoration quality and low computational cost. In addition, it's annoying to collect large-scale LQ-GT clean pairs. As a result, we propose LiteKD-Net, a lightweight knowledge-distilled network for mobile image denoising. First, a physics-guided noise simulation pipeline generates paired training data by adding pixel crosstalk compared with pipelines applied to cameras. Next, we adapt the Real-ESRGAN to identity-resolution denoising and construct a lightweight Student using Lite-RRDB blocks based on depthwise separable convolutions. Third, feature-level knowledge distillation is applied to transfer the Teacher's restoration capability to the Student without introducing additional inference cost. Experiments on real-world datasets show that our model reaches great reduction in runtime and increase in the inference rate with good restoration quality. Our model also reaches the best in all metrics compared with SwinIR. These results indicate that LiteKD-Net provides a great trade-off between restoration quality and computational efficiency.

Bimanual manipulation policies trained with imitation learning are typically evaluated on workstation or datacenter-class GPUs, leaving the cost of deploying them on embedded hardware largely uncharacterized. We present a bimanual SO-101 system running entirely on an NVIDIA Jetson Orin Nano Super (8 GB), the entry-level tier of NVIDIA's embedded line, using a desktop GPU (RTX 3070) only for offline training, evaluated on pick-and-place of a deformable beanbag. First, we build a GStreamer capture pipeline backed by NVMM buffers that removes redundant host-device copies from three-camera sensing. Contrary to expectation, the conventional path fit the memory budget and dropped no frames; what zero-copy sensing recovers is CPU headroom (peak single-core utilization 98.0% to 77.0%) and worst-case latency (117.31 ms to 101.52 ms). Second, we train ACT and Diffusion Policy on identical demonstrations, each at its own reference budget (100k gradient steps for ACT, 200k for Diffusion Policy). ACT converges to a task-competent policy (19/20 trials) while Diffusion Policy does not converge to a usable one (0/10) even at twice the step count, which we attribute to differing convergence costs rather than an accuracy ceiling. Third, we convert ACT to TensorRT. FP16 reduces mean inference latency from 114.02 ms to 17.93 ms (6.4x) and INT8 to 12.65 ms (9.0x), with task success preserved at all three precisions (19/20, 18/20, 19/20). We report two findings not previously documented for ACT: TensorRT's general-purpose INT8 calibration quantizes the ResNet18 backbone but accepts zero of 145 transformer layers, explaining INT8's negligible size reduction over FP16 (0.9%) despite a further 28% latency gain; and the need for quantization is conditional on ACT's action-chunking configuration, feasible in full precision at n_action_steps = 100 but not at the per-step re-prediction temporal ensembling requires.

Physical AI policies require inference throughout their lifecycle, including model evaluation, cloud reinforcement learning rollout, edge GPU serving, and onboard deployment. Although these settings share the same checkpoint and action semantics, they often rely on separate inference programs. To unify them, we build PhyAI, a Physical AI inference engine with a single runtime that keeps architecture-specific conditioning, solver, cache, and output logic in model adapters while sharing graph execution, kernels, memory management, and parallel services. The same codebase runs vision-language-action (VLA) models and world-action models (WAMs) on single or multiple GPUs across onboard, edge, and cloud deployments. We used the adapter interface to add MiniCPM-Robot on the day of its release. PhyAI achieves 1.40x-4.65x speedups over the official implementations of pi0, pi0.5, GR00T N1.7, and MiniCPM-Robot. On Cosmos3-Nano-Policy-DROID it reduces latency from 2.46 to 1.18 s on eight H20 GPUs (CFG=2, TP=4), a 2.08x speedup. Specialized runtimes remain faster in several configurations, so our goal is one runtime with competitive latency rather than the fastest result in every case. Detailed profiles reveal why different models need different execution policies: on a Hopper-series GPU at batch size one, the pi0.5 action expert accounts for 8.8% of FLOPs but 57.2% of latency; at batch size 32 its share drops to 13.5% and throughput reaches about 100 samples/s. Cosmos3 remains generation-dominated and gains only 14.3% throughput as batch size increases from 1 to 16. We further introduce the control-time Roofline, which distinguishes inference-bound from environment-bound control; the measured pi0.5 points on four LIBERO suites are environment-bound while Cosmos3 stays inference-bound. Code and benchmarks: https://github.com/mingti-org/phyai.

Self-supervised monocular depth estimation has emerged as an appealing solution to design lightweight and effective models for deployment on computationally constrained devices due to its reduced reliance on expensive depth sensors. By eliminating the need for ground-truth annotations and leveraging the simplicity of monocular camera setups, this approach facilitates cost-effective data collection and broad applicability across fields such as computer vision and robotics. A critical challenge is achieving resource-efficient neural networks without compromising the overall performance. State-of-the-art models generally adopt depth-wise convolutions and attention mechanisms; however, these functions often incur high energy costs and face compatibility issues in embedded environments. To address this, we propose XiDepth, a lightweight architecture based on the XiNet operator block, designed to enhance feature extraction while maintaining low computational complexity and energy demand. On the KITTI dataset, XiDepth achieves state-of-the-art performance with only 0.8M parameters. Tests on a Raspberry Pi 4 further confirm its suitability for real-world embedded applications, reducing FLOPs by 40% and energy consumption by 35% compared to leading methods.

Hard-constrained recurrent physics-informed networks (HRPINNs) embed known dynamics inside a recurrent numerical integrator and restrict a neural branch to learning only the residual dynamics that the first-principles model does not capture. Kolmogorov--Arnold Networks (KANs) have been proposed as parameter-efficient replacements for multilayer perceptrons (MLPs) in such residual branches, but their learnable B-spline activations follow a markedly different execution profile. Building on prior work that characterized when a vanilla B-spline KAN matches or underperforms an MLP as an HRPINN residual branch in discovery accuracy, this paper asks whether that parameter efficiency survives deployment. Using identical trained weights, we measured execution latency, energy per integration step, and dependability under post-training quantization in the closed recurrent loop on a RISC-V RV64GC platform without vector extensions (StarFive VisionFive~2, SiFive U74). For the two accuracy-comparable pairs, the KAN residual branch executed $13.5\times$ and $8.0\times$ slower and consumed $11.3\times$ and $5.6\times$ more energy per integration step (3.7,$μ$J against 0.33,$μ$J for the smallest pair); across all four parameter-matched size tiers the ranges are $4.7\times$--$14.5\times$ and $4.7\times$--$18.7\times$. Under INT8 quantization, KAN trajectories diverged up to $43\times$ earlier than matched MLPs; the damage traces to weight quantization, not to input-side knot-interval misassignment. These results indicate that the parameter efficiency reported for KANs does not transfer to deployment cost on scalar embedded cores, and that an MLP residual branch is the more dependable default for embedded HRPINN deployment unless specific quantization co-design is used.

Deploying autonomous computer-use agents (CUAs) locally is increasingly important for privacy, cost efficiency, and practical usability, yet improving their performance under strict hardware constraints remains challenging. While recent studies show that inference-time scaling can improve frontier computer-use agents through additional computation during execution, its effectiveness for resource-constrained local models remains poorly understood. We present a systematic empirical study of inference-time scaling in local CUAs across contextual, temporal, structural, and parallel dimensions. We evaluate Qwen3-VL-8B/30B-A3B, UI-TARS-1.5-7B, and OpenCUA-7B on the OSWorld benchmark. Our results show that additional computation often yields diminishing returns while changing failure modes. Contextual scaling provides historical grounding that improves trajectory stability and task accuracy, but its gains saturate as token cost increases and failures shift from repetitive or stalled trajectories toward premature false successes. Temporal scaling similarly reduces max-step stalls, yet does not substantially improve task success, indicating that longer horizons often extend erroneous trajectories rather than correct them. We further find that structural decomposition can introduce planning and formatting overhead in local two-stage agents, while parallel scaling partially mitigates these failures at a substantial computational cost. Overall, our findings suggest that efficient local CUAs require selective compute allocation, failure-aware control mechanisms, and agentic frameworks designed around the capabilities and limitations of local models.

Airborne surveillance from low Earth orbit is hindered by two interconnected bottlenecks: nanosatellites have a limited downlink budget, yet the conventional approach still transmits terabytes of raw imagery to the ground for processing, and open satellite datasets for aircraft are scarce and severely class-imbalanced. These limitations either delay timely decision-making or prevent standard detectors from learning robust representations of rare aircraft classes. In this paper, a workflow that combines on-board inference with generative data augmentation is proposed to address both limitations jointly. Inference is executed on a 6U CubeSat equipped with a low-power edge tensor accelerator, while a diffusion model fine-tuned through low-rank adaptation generates synthetic minority-class imagery. This synthetic output is automatically annotated, pseudo-labelled, by an intermediate detector and merged with classically augmented samples. The results show that the balanced dataset increases global mean average precision from 77.9% to 82.2%, with the minority class rising from F1=0.683 to F1=0.811, and that the quantised detector fits the on-chip memory and projects 25-30 frames per second on orbit. This approach contrasts with the conventional bent-pipe architecture, in which the satellite acts as a passive data collector. Therefore, the computational tests support the proposed workflow as a decision-support tool for real-time, autonomous airborne surveillance from nanosatellites.

Deploying convolutional neural networks generated by large language models (LLMs) on real mobile hardware requires more than GPU validation accuracy: INT8 TensorFlow Lite export, delegate selection, and on-device latency jointly determine whether a model is usable. We present an automated mobile deployment pipeline that closes the loop from QLoRA fine-tuning of an architecture-generating LLM through GPU evaluation, INT8 export, and physical-device benchmarking to gated augmentation of the training corpus. The pipeline is fully scripted and runs cycle-by-cycle without manual intervention, with resume support after interruptions. We evaluate the same frozen protocol on two benchmarks, CIFAR-10 and CIFAR-100, on a Samsung SM-P613 tablet (seed 42, 20 models per cycle, cycles 0-6). On CIFAR-10, cycle 1 is gate-accepted and improves the mobile deployment score approximately 25.6x over the baseline with a mean quantized accuracy of 46.9%; later cycles raise GPU accuracy but fail the non-decreasing mobile gate. On CIFAR-100, the pre-QLoRA baseline retains the best mobile score; iterative rounds improve GPU accuracy (up to 26.2%) yet cannot surpass cycle 0 on-device, and the training pool stalls at 19 examples after the first accepted round. Together, the two studies show that closed-loop GPU fine-tuning does not guarantee monotonic mobile gains, especially on harder classification tasks, and that multi-dataset, on-device measurement is needed to stress-test deployment objectives. We release per-cycle metrics with 95% confidence intervals, all figures, and complete reproduction commands.

Deploying diabetic retinopathy (DR) screening models in primary care requires edge-efficient systems that remain accurate, safe, and reliable under domain shift. Multi-teacher knowledge distillation (KD) is a natural compression strategy, but existing approaches largely assume that all teachers provide equally trustworthy supervision. In our setting, this assumption fails: a strong CNN teacher (EfficientNet-B3, 0.876 QWK) and a weaker Transformer teacher (Swin-Base, 0.830 QWK) are complementary, yet the Transformer's logits can still mislead the student. We therefore propose OrthKD, a selective-trust distillation framework that transfers full supervision from the strong CNN, uses feature-only distillation from the weak ViT, and enforces orthogonality between teacher-specific student projections to encourage complementary rather than redundant evidence. This design preserves local lesion precision, injects global structural context, and improves robustness to distribution shift. On 132,049 retinal images, a 5.4M-parameter MobileNetV3 student reaches 0.885 QWK on EyePACS and improves zero-shot Messidor-2 performance from 0.507 to 0.728 QWK, while also achieving strong referral AUC and calibration. These results show that selectively distilling heterogeneous teachers can enable practical DR screening on resource-constrained devices.

This is an implementation and measurement study of what it costs to run a streaming speech enhancer on a CPU. We port FastEnhancer-Medium at 48 kHz to faster-enhancer.c, a C runtime with six int8 GEMM tiers selected at initialization, leaving architecture and weights untouched. One Apple M2 core reaches 0.069 real-time factor, against 0.230 for the fp32 ONNX Runtime graph on the same machine, a 3.3x speedup. A Galaxy S23+ (Snapdragon 8 Gen 2) reaches 0.096. The speedup comes from specializing every layer of the runtime around one fixed model. Activation ranges are recomputed per frame, so no calibration set is needed; the k=3 convolutions use Winograd F(2,3); cross-stage state is fp16; the GRU and the dequantization epilogues are fused; and nothing is allocated after startup. Over 824 VoiceBank-DEMAND utterances the engine tracks fp32 to within -0.006 PESQ and -0.08 dB SNR. Speed alone does not settle deployment cost. The enhancer holds a fraction of a core for as long as the microphone is open, so its real-time factor is a duty cycle. A benchmark races through a file; an audio callback does not. Pacing to the 6.67 ms deadline costs 4.2x per frame, saves 49% of the energy, and leaves the cheapest core placement missing 96% of its deadlines. All SIMD tiers within an architecture family emit byte-identical output. The runtime is released as a dependency-free library.

We present VibeVoice-ASR-BitNet, a compressed variant of VibeVoice-ASR optimized for real-time inference on edge CPUs. We apply heterogeneous quantization tailored to the computational characteristics of each stage: the VAE acoustic tokenizer uses full-pipeline INT8 quantization (I8_S) with kernel fusion and SIMD optimization, while the autoregressive language model adopts BitNet-style ternary weights (I2_S). To preserve accuracy under aggressive compression, we employ a progressive quantization-aware training strategy. For inference, we implement custom SIMD kernels and fused operators within the ggml framework targeting both ARM and x86 platforms, achieving real-time recognition (RTF < 1) on low-thread-count CPUs. VibeVoice-ASR-BitNet is 1.6--2.3x faster than Whisper.cpp at comparable model sizes (~1.6 GB), with only modest accuracy degradation compared to the FP16 baseline.

An increase in advanced Android malware requires the use of deep learning models, which can run on Android devices. But there is a trade-off between security and energy use, as strong detection models can drain the battery of devices fast. This work tests different Multi-Layer Perceptron (MLP) model configurations to balance malware detection performance and energy efficiency. In this work, we compared standard FP32 models with optimized INT8 quantized neural networks with different model depths using TUANDROMD and DREBIN datasets for both classification performance and energy consumption. The results show that INT8 quantization reduces model size by about 3.5 times with a decrease in energy consumption to 0.0189 mJ per inference, while maintaining more than 99.2% detection accuracy. We found that shallow quantized architectures, such as 3-layer and 4-layer QNNs, reduce energy costs by improving throughput and shortening the time of CPU operating in a high-power state. This work shows that efficient malware protection can be achieved on resource-constrained smartphones and provides a foundation for Green AI in mobile security.

Deploying a vision-language model with full UI understanding on end devices has long been trapped between accuracy and efficiency: on one side is the accuracy bar for OCR, screen understanding, visual question answering, and element grounding; on the other is the strict compute, memory, and power budget of mobile chips. Existing work either trades one for the other, or stops at simulation without real-device validation. We present StepX-Edge, a 0.9B-parameter on-device UI vision-language model that resolves this tension through three-layer co-design of architecture, training, and deployment. Architecturally, UI-aware Layered Visual Encoding (ULVE) and a Progressive Dimensionality Projection (PDP) connector target the extreme aspect ratios and fine-grained perception of screens, while standard full attention throughout ensures native compatibility with mainstream mobile NPU operators. For training, the five-stage StepX-Curriculum framework is designed around our observation of mutual-promotion effects among UI subtasks, so that all four capabilities grow synergistically under a tight parameter budget rather than interfering. For deployment, a module-wise differentiated two-stage PTQ-to-QAT quantization scheme keeps the post-quantization accuracy loss within 1%. StepX-Edge achieves the strongest overall UI understanding among <=1B models, surpassing all 2B-2.3B baselines on ScreenQA (88.76 F1) and Chinese OCRBench v2 (57.25), and matching 1.3B-2.3B general VLMs on RefCOCO (92.0%) and OCRBench v1 (831) with far fewer parameters. After W4A16+KV8 quantization, the model runs stably on Snapdragon 8 Gen5 devices with ~0.84 s TTFT, 98 tok/s decode, and 1.4 GB peak memory. We will open-source the training data, the full training recipe, and the quantization deployment pipeline.