On-Device Inference

We present an RL-driven compiler that jointly optimizes ASIC architecture, memory hierarchy, and workload partitioning for AI inference across 3nm to 28nm. The design space is formulated as a single Markov Decision Process with mixed discrete-continuous actions and a unified Power-Performance-Area (PPA) objective. Soft Actor-Critic (SAC) with Mixture-of-Experts gating explores the joint space of mesh topology, per-core microarchitecture, and operator placement. We validate on two workloads, Llama 3.1 8B FP16 (high-performance mode, 29809 tokens per second at 3nm) and SmolVLM (low-power mode, less than 13 mW at all nodes, 10 MHz). Across 7 process nodes, the RL automatically adapts mesh sizes and per-tile configurations, including heterogeneous FETCH, VLEN, and memory allocation without node-specific manual retuning.

Eye tracking (ET) plays a critical role in augmented and virtual reality applications. However, rapidly deploying high-accuracy, on-device gaze estimation for new products remains challenging because hardware configurations (e.g., camera placement, camera pose, and illumination) often change across device generations. Visual foundation models (VFMs) excel on natural-image benchmarks and offer a promising path to rapid training and deployment; yet, we find that off-the-shelf VFMs still struggle to reach high accuracy on specialized near-eye infrared images. To close this gap, we introduce DistillGaze, a framework that distills a VFM using labeled synthetic data and unlabeled real data for rapid, high-accuracy on-device gaze estimation. DistillGaze proceeds in two stages. First, we adapt a VFM into a domain-specialized teacher using synthetic gaze labels and unlabeled real images. Synthetic data provide scalable, high-quality gaze supervision, while unlabeled real data bridges the synthetic-to-real domain gap. Second, we train an on-device student from both teacher guidance and self-training. Evaluated on a large-scale crowd-sourced dataset spanning more than 2,000 participants, DistillGaze reduces median gaze error by 58.6% relative to synthetic-only baselines while maintaining a lightweight 256K-parameter model suitable for real-time on-device deployment. More broadly, DistillGaze offers an efficient path to training and deploying ET models that adapt to hardware changes, and a recipe for combining synthetic supervision with unlabeled real data in on-device regression tasks.

Generative Artificial Intelligence (GenAI) features such as image editing, object removal, and prompt-guided image transformation are increasingly integrated into mobile applications. However, deploying Large Vision Models (LVMs) for such tasks on resource-constrained devices remains challenging due to their high memory and compute requirements. While Low-Rank Adapters (LoRAs) enable parameter-efficient task adaptation, existing Mobile deployment pipelines typically compile separate model binaries for each LoRA + a copy of the foundation model, resulting in redundant storage and increased runtime overhead. In this work, we present a unified framework for enabling multi-task GenAI inference on edge devices using a single shared model. Our key idea is to treat LoRA weights as runtime inputs rather than embedding them into the compiled model graph, allowing dynamic task switching at runtime without recompilation. Then, to support efficient on-device execution, we introduce QUAD (Quantization with Unified Adaptive Distillation), a quantizationaware training strategy that aligns multiple LoRA adapters under a shared quantization profile. We implement the proposed system with a lightweight runtime stack compatible with mobile NPUs and evaluate it across multiple chipsets. Experimental results demonstrate up to 6x and 4x reduction in memory footprint and latency improvements, respectively, while maintaining high visual quality across multiple GenAI tasks.

Diffusion models have made significant progress in both text-to-image (T2I) generation and text-guided image editing. However, these models are typically built with billions of parameters, leading to high latency and increased deployment challenges. While on-device diffusion models improve efficiency, they largely focus on T2I generation and lack support for image editing. In this paper, we propose DreamLite, a compact unified on-device diffusion model (0.39B) that supports both T2I generation and text-guided image editing within a single network. DreamLite is built on a pruned mobile U-Net backbone and unifies conditioning through in-context spatial concatenation in the latent space. It concatenates images horizontally as input, using a (target | blank) configuration for generation tasks and (target | source) for editing tasks. To stabilize the training of this compact model, we introduce a task-progressive joint pretraining strategy that sequentially targets T2I, editing, and joint tasks. After high-quality SFT and reinforcement learning, DreamLite achieves GenEval (0.72) for image generation and ImgEdit (4.11) for image editing, outperforming existing on-device models and remaining competitive with several server-side models. By employing step distillation, we further reduce denoising processing to just 4 steps, enabling our DreamLite could generate or edit a 1024 x 1024 image in less than 1s on a Xiaomi 14 smartphone. To the best of our knowledge, DreamLite is the first unified on-device diffusion model that supports both image generation and image editing.

Gaze gestures can provide hands free input on mobile devices, but practical use requires (i) gestures users can learn and recall and (ii) recognition models that are efficient enough for on-device deployment. We present an end-to-end pipeline using commodity ARKit head/eye transforms and a scaffolded guidance-to-recall protocol grounded in learning theory. In a pilot feasibility study (N=4 participants; 240 trials; controlled single-session setting), we benchmark a compact time-series model (TinyHAR) against deeper baselines (DeepConvLSTM, SA-HAR) on 5-way gesture recognition and 4-way user identification. TinyHAR achieves strong performance in this pilot benchmark (Macro F1 = 0.960 for gesture recognition; Macro F1 = 0.997 for user identification) while using only 46k parameters. A modality analysis further indicates that head pose dynamics are highly informative for mobile gaze gestures, highlighting embodied head--eye coordination as a key design consideration. Although the small sample size and controlled setting limit generalizability, these results indicate a potential direction for further investigation into on-device gaze gesture recognition.

Diffusion Transformers (DiT) have established a new state-of-the-art in high-fidelity image synthesis; however, their massive computational complexity and memory requirements hinder local deployment on resource-constrained edge devices. In this paper, we introduce EdgeDiT, a family of hardware-efficient generative transformers specifically engineered for mobile Neural Processing Units (NPUs), such as the Qualcomm Hexagon and Apple Neural Engine (ANE). By leveraging a hardware-aware optimization framework, we systematically identify and prune structural redundancies within the DiT backbone that are particularly taxing for mobile data-flows. Our approach yields a series of lightweight models that achieve a 20-30% reduction in parameters, a 36-46% decrease in FLOPs, and a 1.65-fold reduction in on-device latency without sacrificing the scaling advantages or the expressive capacity of the original transformer architecture. Extensive benchmarking demonstrates that EdgeDiT offers a superior Pareto-optimal trade-off between Frechet Inception Distance (FID) and inference latency compared to both optimized mobile U-Nets and vanilla DiT variants. By enabling responsive, private, and offline generative AI directly on-device, EdgeDiT provides a scalable blueprint for transitioning large-scale foundation models from high-end GPUs to the palm of the user.

This paper presents a lightweight K-Means anomaly detection model and a distributed model-sharing workflow designed for resource-constrained microcontrollers (MCUs). Using real power measurements from a mini-fridge appliance, the system performs on-device feature extraction, clustering, and threshold estimation to identify abnormal appliance behavior. To avoid retraining models on every device, we introduce the Distributed Internet of Learning (DIoL), which enables a model trained on one MCU to be exported as a portable, text-based representation and reused directly on other devices. A two-device prototype demonstrates the feasibility of the "Train Once, Share Everywhere" (TOSE) approach using a real-world appliance case study, where Device A trains the model and Device B performs inference without retraining. Experimental results show consistent anomaly detection behavior, negligible parsing overhead, and identical inference runtimes between standalone and DIoL-based operation. The proposed framework enables scalable, low-cost TinyML deployment across fleets of embedded devices.

This paper presents a fully autonomous Tiny Machine Learning (TinyML) Z-Score-based anomaly detection system deployed on a low-power microcontroller for real-time monitoring of appliance behavior using power side-channel data. Unlike existing Internet of Things (IoT) anomaly detection approaches that rely on offline training or cloud-assisted analytics, the proposed system performs both model training and inference directly on a resource-constrained microcontroller without external computation or connectivity. The system continuously samples current consumption, computes Root Mean Square (RMS) values on-device, and derives statistical parameters during an initial training phase. Anomalies are detected using lightweight Z-Score thresholds, enabling interpretable and computationally efficient inference suitable for embedded deployment. The architecture was implemented on an STM32-based platform and evaluated using a 14-day dataset collected from a household mini-fridge under normal operation and controlled anomaly conditions. Results demonstrate perfect detection performance, with Precision and Recall of 1.00, inference latencies on the order of tens of microseconds, and a total memory footprint of approximately 3.3 KB SRAM and 63 KB Flash. These results confirm that robust and fully autonomous TinyML anomaly detection can be achieved on low-cost microcontrollers. Future work includes extending the framework to incorporate additional lightweight models and multi-device learning scenarios.

This work presents a lightweight super-resolution (LiteSR) neural network for depth and intensity images acquired from a consumer-grade single-photon avalanche diode (SPAD) array with a 48x32 spatial resolution. The proposed framework reconstructs high-resolution (HR) images of size 256x256. Both synthetic and real datasets are used for performance evaluation. Extensive quantitative metrics demonstrate high reconstruction fidelity on synthetic datasets, while experiments on real indoor and outdoor measurements further confirm the robustness of the proposed approach. Moreover, the SPAD sensor is interfaced with an Arduino UNO Q microcontroller, which receives low-resolution (LR) depth and intensity images and feeds them into a compressed, pre-trained deep learning (DL) model, enabling real-time SR video streaming. In addition to the 256x256 setting, a range of target HR resolutions is evaluated to determine the maximum achievable upscaling resolution (512x512) with LiteSR, including scenarios with noise-corrupted LR inputs. The proposed LiteSR-embedded system co-design provides a scalable, cost-effective solution to enhance the spatial resolution of current consumer-grade SPAD arrays to meet HR imaging requirements.

Real-time 30-to-60 fps video frame interpolation on mobile neural processing units (NPUs) requires each synthesized frame within 33.3 ms. We show that mainstream flow-based video frame interpolation faces three structural deployment barriers on mobile NPUs: spatial sampling operators exceed the frame budget or lack hardware support, iterative flow refinement collapses under 8-bit integer post-training quantization, and memory-bound operators dominate the inference graph. ANVIL addresses these barriers by reusing motion vectors from the H.264/AVC decoder to prealign input frames, removing learned optical flow, spatial sampling, and iterative accumulation from the accelerator graph. The remaining residual is refined by a convolution-dominated network composed almost entirely of compute-bound operators. On a Snapdragon 8 Gen 3 device, ANVIL achieves 12.8 ms 1080p inference at 8-bit integer precision; an open-source Android player sustains 28.4 ms median end-to-end latency over 30-minute continuous playback. Per-operator causal analysis identifies quantized accumulation on recurrent flow states as a key mechanism behind integer quantization failure in iterative methods. The current design targets H.264/AVC playback with decoder-exposed motion vectors.

Always-on sensing is essential for next-generation edge/wearable AI systems, yet continuous high-fidelity RGB video capture remains prohibitively expensive for resource-constrained mobile and edge platforms. We present a new paradigm for efficient streaming video understanding: grayscale-always, color-on-demand. Through preliminary studies, we discover that color is not always necessary. Sparse RGB frames suffice for comparable performance when temporal structure is preserved via continuous grayscale streams. Building on this insight, we propose ColorTrigger, an online training-free trigger that selectively activates color capture based on windowed grayscale affinity analysis. Designed for real-time edge deployment, ColorTrigger uses lightweight quadratic programming to detect chromatic redundancy causally, coupled with credit-budgeted control and dynamic token routing to jointly reduce sensing and inference costs. On streaming video understanding benchmarks, ColorTrigger achieves 91.6% of full-color baseline performance while using only 8.1% RGB frames, demonstrating substantial color redundancy in natural videos and enabling practical always-on video sensing on resource-constrained devices.

Machine learning models are widely integrated into modern mobile apps to analyze user behaviors and deliver personalized services. Ensuring low-latency on-device model execution is critical for maintaining high-quality user experiences. While prior research has primarily focused on accelerating model inference with given input features, we identify an overlooked bottleneck in real-world on-device model execution pipelines: extracting input features from raw application logs. In this work, we explore a new direction of feature extraction optimization by analyzing and eliminating redundant extraction operations across different model features and consecutive model inferences. We then introduce AutoFeature, an automated feature extraction engine designed to accelerate on-device feature extraction process without compromising model inference accuracy. AutoFeature comprises three core designs: (1) graph abstraction to formulate the extraction workflows of different input features as one directed acyclic graph, (2) graph optimization to identify and fuse redundant operation nodes across different features within the graph; (3) efficient caching to minimize operations on overlapping raw data between consecutive model inferences. We implement a system prototype of AutoFeature and integrate it into five industrial mobile services spanning search, video and e-commerce domains. Online evaluations show that AutoFeature reduces end-to-end on-device model execution latency by 1.33x-3.93x during daytime and 1.43x-4.53x at night.

Earth observation (EO) missions traditionally rely on transmitting raw or minimally processed imagery from satellites to ground stations for computationally intensive analysis. This paradigm is infeasible for CubeSat systems due to stringent constraints on the onboard embedded processors, energy availability, and communication bandwidth. To overcome these limitations, the paper presents a TinyML-based Convolutional Neural Networks (ConvNets) model optimization and deployment pipeline for onboard image classification, enabling accurate, energy-efficient, and hardware-aware inference under CubeSat-class constraints. Our pipeline integrates structured iterative pruning, post-training INT8 quantization, and hardware-aware operator mapping to compress models and align them with the heterogeneous compute architecture of the STM32N6 microcontroller from STMicroelectronics. This Microcontroller Unit (MCU) integrates a novel Arm Cortex-M55 core and a Neural-ART Neural Processing Unit (NPU), providing a realistic proxy for CubeSat onboard computers. The paper evaluates the proposed approach on three EO benchmark datasets (i.e., EuroSAT, RS_C11, MEDIC) and four models (i.e., SqueezeNet, MobileNetV3, EfficientNet, MCUNetV1). We demonstrate an average reduction in RAM usage of 89.55% and Flash memory of 70.09% for the optimized models, significantly decreasing downlink bandwidth requirements while maintaining task-acceptable accuracy (with a drop ranging from 0.4 to 8.6 percentage points compared to the Float32 baseline). The energy consumption per inference ranges from 0.68 mJ to 6.45 mJ, with latency spanning from 3.22 ms to 30.38 ms. These results fully satisfy the stringent energy budgets and real-time constraints required for efficient onboard EO processing.

Non-linear activation functions play a pivotal role in on-device inference and training, as they not only consume substantial hardware resources but also impose a significant impact on system performance and energy efficiency. In this work, we propose Distribution-Aware Piecewise Activation (DAPA), a differentiable and hardware-friendly activation function for Transformer architectures by exploiting the distribution of pre-activation data. DAPA employs a non-uniform piecewise approximation that allocates finer segments to high-probability regions of the distribution, improving generalizability over prior piecewise linear methods. The resulting approximation is further quantized using Distribution-Weighted Mean Square Error to reduce latency and resource utilization for hardware deployment. Our HLS implementation demonstrates that DAPA speeds up GELU computation by 16$\times$ and decreases DSP utilization by 16$\times$ while maintaining comparable or better performance across vision Transformers and GPT-2 models.

Mobile robotic manipulation--the ability of robots to navigate spaces and interact with objects--is a core capability of physical AI. Foundation models have led to breakthroughs in their performance, but at a significant computational cost. We present the first measurement study of mobile robotic manipulation workloads across onboard, edge, and cloud GPU platforms. We find that the full workload stack is infeasible to run on smaller onboard GPUs, while larger onboard GPUs drain robot batteries several hours faster. Offloading alleviates these constraints but introduces its own challenges, as additional network latency degrades task accuracy, and the bandwidth requirement makes naive cloud offloading impractical. Finally, we quantify opportunities and pitfalls of sharing compute across robot fleets. We believe our measurement study will be crucial to designing inference systems for mobile robots.

Realizing on-device ML-based gesture detection under tight real-time performance, energy and memory constraints is challenging, especially when considering mobile devices with varying battery-power levels. Existing EdgeAI deployments typically rely on a single fixed detector, limiting optimization opportunities. We present Scale-Gest, a novel run-time adaptive gesture detection framework that expands the detector space into a dense family of tiny-YOLO architectures. We introduce multiple novel device-calibrated ACE (Accuracy-Complexity-Energy) profiles by analyzing different model-resolution-stride operating points. A lightweight run-time controller selects an appropriate ACE mode under user-defined and battery constraints, while a motion-aware hand-gesture-tracking ROI gate crops the input for reduced complexity detection. To evaluate performance of our system in real-world car driving scenarios, we introduce a temporally-annotated Driver Simulated Gesture (DSG-18) dataset. Scale-Gest maintains event-level F1 while significantly reducing energy and latency compared to single-detector approaches. On a battery-powered laptop running gesture streams, our ACE controller reduces per-frame energy by 4x (from 6.9 mJ to 1.6 mJ) while maintaining high gesture-detection performance (event-level F1 = 0.8-0.9) and low mean latency (6 ms).

Embodied AI agents increasingly require parallel execution of multiple tasks, such as manipulation, conversation, and memory construction, from shared observations under distinct time constraints. Recent Mixture-of-Transformers (MoT) Vision-Language-Action Models (VLAs) support such heterogeneous outputs, yet existing inference systems incur redundant computation and resource contention when serving concurrent tasks on-device. We identify isolated KV cache management as the root cause. To address this, we propose unified KV cache management, an inference design that treats the KV cache as a first-class shared resource across tasks and control cycles. This abstraction enables two key optimizations: cross-task KV sharing eliminates redundant prefill of shared observations, while cross-frame continuous batching batches unfinished language requests from different frames and advances them within a per-frame decoding budget. We implement this design in OxyGen and evaluate $π_{0.5}$ on NVIDIA GeForce RTX 4090 and Jetson AGX Thor. OxyGen achieves up to 3.7$\times$ speedup over isolated execution, delivering over 200 tokens/s language throughput and 70 Hz action frequency simultaneously on RTX 4090. We also validate the design on Xiaomi-Robotics-0 and StarVLA across different backbones and expert configurations. We further demonstrate the inference gains on a humanoid robot with on-board Jetson AGX Thor.

Space missions increasingly deploy high-fidelity sensors that produce data volumes exceeding onboard buffering and downlink capacity. This work evaluates FPGA acceleration of neural networks (NNs) across four space use cases on the AMD ZCU104 board. We use Vitis AI (AMD DPU) and Vitis HLS to implement inference, quantify throughput and energy, and expose toolchain and architectural constraints relevant to deployment. Vitis AI achieves up to 34.16$\times$ higher inference rate than the embedded ARM CPU baseline, while custom HLS designs reach up to 5.4$\times$ speedup and add support for operators (e.g., sigmoids, 3D layers) absent in the DPU. For these implementations, measured MPSoC inference power spans 1.5-6.75 W, reducing energy per inference versus CPU execution in all use cases. These results show that NN FPGA acceleration can enable onboard filtering, compression, and event detection, easing downlink pressure in future missions.

Real-time, on-device segmentation is critical for latency-sensitive and privacy-aware applications such as smart glasses and Internet-of-Things devices. We introduce PicoSAM3, a lightweight promptable visual segmentation model optimized for edge and in-sensor execution, including deployment on the Sony IMX500 vision sensor. PicoSAM3 has 1.3M parameters and combines a dense CNN architecture with region of interest prompt encoding, Efficient Channel Attention, and knowledge distillation from SAM2 and SAM3. On COCO and LVIS, PicoSAM3 achieves 65.45% and 64.01% mIoU, respectively, outperforming existing SAM-based and edge-oriented baselines at similar or lower complexity. The INT8 quantized model preserves accuracy with negligible degradation while enabling real-time in-sensor inference at 11.82ms latency on the IMX500, fully complying with its memory and operator constraints. Ablation studies show that distillation from large SAM models yields up to +14.5% mIoU improvement over supervised training and demonstrate that high-quality, spatially flexible promptable segmentation is feasible directly at the sensor level.

3D Gaussian Splatting (3DGS) has emerged as a powerful representation for high-quality rendering across a wide range of applications.However, its high computational demands and large storage costs pose significant challenges for deployment on mobile devices. In this work, we propose a mobile-tailored real-time Gaussian Splatting method, dubbed Mobile-GS, enabling efficient inference of Gaussian Splatting on edge devices. Specifically, we first identify alpha blending as the primary computational bottleneck, since it relies on the time-consuming Gaussian depth sorting process. To solve this issue, we propose a depth-aware order-independent rendering scheme that eliminates the need for sorting, thereby substantially accelerating rendering. Although this order-independent rendering improves rendering speed, it may introduce transparency artifacts in regions with overlapping geometry due to the scarcity of rendering order. To address this problem, we propose a neural view-dependent enhancement strategy, enabling more accurate modeling of view-dependent effects conditioned on viewing direction, 3D Gaussian geometry, and appearance attributes. In this way, Mobile-GS can achieve both high-quality and real-time rendering. Furthermore, to facilitate deployment on memory-constrained mobile platforms, we also introduce first-order spherical harmonics distillation, a neural vector quantization technique, and a contribution-based pruning strategy to reduce the number of Gaussian primitives and compress the 3D Gaussian representation with the assistance of neural networks. Extensive experiments demonstrate that our proposed Mobile-GS achieves real-time rendering and compact model size while preserving high visual quality, making it well-suited for mobile applications.

Autonomous navigation typically relies on power-intensive processors, limiting accessibility in low-cost robotics. Although microcontrollers offer a resource-efficient alternative, they impose strict constraints on model complexity. We present TinyNav, an end-to-end TinyML system for real-time autonomous navigation on an ESP32 microcontroller. A custom-trained, quantized 2D convolutional neural network processes a 20-frame sliding window of depth data to predict steering and throttle commands. By avoiding 3D convolutions and recurrent layers, the 23k-parameter model achieves 30 ms inference latency. Correlation analysis and Grad-CAM validation indicate consistent spatial awareness and obstacle avoidance behavior. TinyNav demonstrates that responsive autonomous control can be deployed directly on highly constrained edge devices, reducing reliance on external compute resources.

The recent developments in machine learning have highlighted a conflict between online platforms and their users in terms of privacy. The importance of user privacy and the struggle for power over user data has been intensified as regulators and operators attempt to police online platforms. As users have become increasingly aware of privacy issues, client-side data storage, management, and analysis have become a favoured approach to large-scale centralised machine learning. However, state-of-the-art machine learning methods require vast amounts of labelled user data, making them unsuitable for models that reside client-side and only have access to a single user's data. State-of-the-art methods are also computationally expensive, which degrades the user experience on compute-limited hardware and also reduces battery life. A recent alternative approach has proven remarkably successful in classification tasks across a wide variety of data -- using a compression-based distance measure (called normalised compression distance) to measure the distance between generic objects in classical distance-based machine learning methods. In this work, we demonstrate that the normalised compression distance is actually not a metric; develop it for the wider context of kernel methods to allow modelling of complex data; and present techniques to improve the training time of models that use this distance measure. We demonstrate that the normalised compression distance works as well as and sometimes better than other metrics and kernels -- while requiring only marginally more computational costs and in spite of the lack of formal metric properties. The end results is a simple model with remarkable accuracy even when trained on a very small number of samples allowing for models that are small and effective enough to run entirely on a client device using only user-supplied data.

Deep learning (DL) models can now run on microcontrollers (MCUs). Through neural architecture search (NAS), we can search DL models that meet the constraints of MCUs. Among various constraints, energy and latency costs of the model inference are critical metrics. To predict them, existing research relies on coarse proxies such as multiply-accumulations (MACs) and model's input parameters, often resulting in inaccurate predictions or requiring extensive data collection. In this paper, we propose InstMeter, a predictor leveraging MCUs' clock cycles to accurately estimate the energy and latency of DL models. Clock cycles are fundamental metrics reflecting MCU operations, directly determining energy and latency costs. Furthermore, a unique property of our predictor is its strong linearity, allowing it to be simple and accurate. We thoroughly evaluate InstMeter under different scenarios, MCUs, and software settings. Compared with state-of-the-art studies, InstMeter can reduce the energy and latency prediction errors by $3\times$ and $6.5\times$, respectively, while requiring $100\times$ and $10\times$ less training data. In the NAS scenario, InstMeter can fully exploit the energy budget, identifying optimal DL models with higher inference accuracy. We also evaluate InstMeter's generalization performance through various experiments on three ARM MCUs (Cortex-M4, M7, M33) and one RISC-V-based MCU (ESP32-C3), different compilation options (-Os, -O2), GCC versions (v7.3, v10.3), application scenarios (keyword spotting, image recognition), dynamic voltage and frequency scaling, temperatures (21°C, 43°C), and software settings (TFLMv2.4, TFLMvCI). We will open our source codes and the MCU-specific benchmark datasets.

We present TVBC (Time-Varying Biquad Cascade), an interpretable, low-latency speech enhancement model for real-time, on-device assistive hearing. A lightweight neural controller predicts, in real time, the coefficients of a differentiable cascade of 35 second-order IIR filters (biquads), so the model tracks non-stationary noise while keeping a fully interpretable processing chain: every spectral modification is an explicit, adjustable equalizer curve rather than an opaque `black-box' transform. Because the biquad cascade carries the signal processing, the controller can be made very small, driving the cascade with only 24k parameters at a 10.7ms algorithmic latency, within hearing-aid budgets, and running entirely on-device so that audio never leaves the device. We also expose the suppression-versus-preservation trade-off as an explicit control: it can be set during training through the loss weighting, and adjusted at inference, with no retraining, by mixing the noisy input with the denoised output. On hearing-aid metrics (HASPI/HASQI) the 24k model stays within about 0.02 of DFNet3 (2.3M parameters, almost two orders of magnitude larger) while using about 29X fewer multiply-accumulates, although larger black-box models still lead on reference metrics such as PESQ. We present TVBC as a proof of concept for a compact, interpretable, and controllable denoiser for on-device assistive hearing.

Vision-Language-Action (VLA) models provide a unified framework for perception, language conditioning, and action generation, but many existing systems remain difficult to deploy in embedded robotic settings because of their computational requirements and inference latency. In this paper, we present LiteVLA-Edge, a deployment-oriented VLA pipeline for fully on-device inference on Jetson Orin-class hardware. Our approach combines supervised image-to-action fine-tuning in FP32 with post-training 4-bit GGUF quantization and GPU-accelerated inference through the llama.cpp runtime. Under our deployment configuration, LiteVLA-Edge achieves a mean end-to-end latency of 150.5,ms (approximately 6.6,Hz) while operating entirely offline within a ROS~2-integrated perception--reasoning--action pipeline. Rather than introducing a new policy objective, our contribution is a practical systems path for executing compact multimodal control models locally on embedded hardware while preserving modular interfaces between perception, reasoning, and actuation. These results establish timing feasibility for reactive language-conditioned control and provide a reproducible baseline for future task-level evaluation of on-device VLAs in robotics.

Virtual try-on (VTON) has recently achieved impressive visual fidelity, but most existing systems require uploading personal photos to cloud-based GPUs, raising privacy concerns and limiting on-device deployment. To address this, we present Mobile-VTON, a high-quality, privacy-preserving framework that enables fully offline virtual try-on on commodity mobile devices using only a single user image and a garment image. Mobile-VTON introduces a modular TeacherNet-GarmentNet-TryonNet (TGT) architecture that integrates knowledge distillation, garment-conditioned generation, and garment alignment into a unified pipeline optimized for on-device efficiency. Within this framework, we propose a Feature-Guided Adversarial (FGA) Distillation strategy that combines teacher supervision with adversarial learning to better match real-world image distributions. GarmentNet is trained with a trajectory-consistency loss to preserve garment semantics across diffusion steps, while TryonNet uses latent concatenation and lightweight cross-modal conditioning to enable robust garment-to-person alignment without large-scale pretraining. By combining these components, Mobile-VTON achieves high-fidelity generation with low computational overhead. Experiments on VITON-HD and DressCode at 1024 x 768 show that it matches or outperforms strong server-based baselines while running entirely offline. These results demonstrate that high-quality VTON is not only feasible but also practical on-device, offering a secure solution for real-world applications. Code and project page are available at https://zhenchenwan.github.io/Mobile-VTON/.

A recent report on "Learning the natural history of human disease with generative transformers" created an opportunity to assess the engineering challenge of delivering user-facing Generative AI applications in privacy-sensitive domains. The application of these models, particularly for personalized healthcare tasks like predicting individual morbidity risk, is typically constrained by data privacy concerns. This project was accordingly designed as an in-browser model deployment exercise (an "App") testing the architectural boundaries of client-side inference generation (no downloads or installations). We relied exclusively on the documentation provided in the reference report to develop the model, specifically testing the "R" component of the FAIR data principles: Findability, Accessibility, Interoperability, and Reusability. The successful model deployment, leveraging ONNX and a custom JavaScript SDK, establishes a secure, high-performance architectural blueprint for the future of private generative AI in medicine.

Always-on keyword spotting (KWS) demands on-device adaptation to cope with user- and environment-specific distribution shifts under tight latency and energy budgets. This paper proposes, for the first time, coupling weight adaptation (i.e., on-device training) with architectural adaptation, in the form of online structured channel pruning, for personalized on-device KWS. Starting from a state-of-the-art self-learning personalized KWS pipeline, we compare data-agnostic and data-aware pruning criteria applied on in-field pseudo-labelled user data. On the HeySnips and HeySnapdragon datasets, we achieve up to 9.63x model-size compression with respect to unpruned baselines at iso-task performance, measured as the accuracy at 0.5 false alarms per hour. When deploying our adaptation pipeline on a Jetson Orin Nano embedded GPU, we achieve up to 1.52x/1.57x and 1.64x/1.77x latency and energy-consumption improvements during online training/inference compared to weights-only adaptation.

This work presents a wearable human activity recognition (HAR) system that performs real-time inference directly inside a shank-mounted inertial measurement unit (IMU) to support low-latency control of a lower-limb exoskeleton. Unlike conventional approaches that continuously stream raw inertial data to a microcontroller for classification, the proposed system executes activity recognition at the sensor level using the embedded Machine Learning Core (MLC) of the STMicroelectronics LSM6DSV16X IMU, allowing the host microcontroller to remain in a low-power state and read only the recognized activity label from IMU registers. While the system generalizes to multiple human activities, this paper focuses on three representative locomotion modes - stance, level walking, and stair ascent - using data collected from adult participants. A lightweight decision-tree model was configured and deployed for on-sensor execution using ST MEMS Studio, enabling continuous operation without custom machine learning code on the microcontroller. During operation, the IMU asserts an interrupt when motion or a new classification is detected; the microcontroller wakes, reads the MLC output registers, and forwards the inferred mode to the exoskeleton controller. This interrupt-driven, on-sensor inference architecture reduces computation and communication overhead while preserving battery energy and improving robustness in distinguishing level walking from stair ascent for torque-assist control.

Deploying object detection on microcontrollers (MCUs) enables intelligent edge devices but current models cannot learn new object categories after deployment. Existing continual learning methods require storing raw images far exceeding MCU memory budgets of tens of kilobytes. We present Latent Replay Detection (LRD), the first framework for continual object detection under MCU memory constraints. Our key contributions are: 1. Task-Adaptive Compression: Unlike fixed PCA, we propose learnable compression with FiLM (Feature-wise Linear Modulation) conditioning, where task specific embeddings modulate the compression to preserve discriminative features for each task's distribution; 2. Spatial-Diverse Exemplar Selection: Traditional sampling ignores spatial information critical for detection - we select exemplars maximizing bounding box diversity via farthest-point sampling in IoU space, preventing localization bias in replay; 3. MCU-Deployable System: Our latent replay stores 150 bytes per sample versus >10KB for images, enabling a 64KB buffer to hold 400+ exemplars. Experiments on CORe50 (50 classes, 5 tasks) demonstrate that LRD achieves mAP@50 on the initial task and maintains strong performance across subsequent tasks - a significant improvement over naive fine-tuning while operating within strict MCU constraints. Our task-adaptive FiLM compression and spatial diverse exemplar selection work synergistically to preserve detection capabilities. Deployed on STM32H753ZI, ESP32-S3, and MAX78000 MCUs, LRD achieves 4.9-97.5ms latency per inference within a 64KB memory budget-enabling practical continual detection on edge devices for the first time.