arXiv subjects

Hyeonuk Kim

Publications and source records attributed to Hyeonuk Kim.

3 recordsLinked to original sources

arXiv2026

UltraViT: Latency-Optimized On-device Vision Encoder for Large Vision-Language Models

Large Vision-Language Models (LVLMs) remain bottlenecked by massive computational footprints, precluding their deployment on resource-constrained edge devices. While efforts to compress LVLMs focus heavily on vision token reduction or smaller language models, the vision encoder is largely overlooked, typically deployed as a monolithic, computationally heavy feature extractor. Moreover, there is no previous effort that designs a vision encoder for LVLMs directly optimized for on-device latency. In this paper, we present UltraViT, a vision encoder for LVLMs, explicitly designed and optimized for on-device performance. Specifically, by taking into account real on-device latencies, we systematically design a pyramidal architecture that strategically integrates and adapts heterogeneous spatial mixers at the macro-block level. Furthermore, to pre-train UltraViT, we propose a novel two-stage generative pre-training strategy: cultivating rich spatial features via dense distillation, followed by direct generative supervision from a capacity-mixed frozen LLM. Compared to standard contrastive and SSL, we show that our pre-training is much more effective for achieving high-level semantic grounding for UltraViT needed for the subsequent generative multimodal alignment of LVLM training. Extensive experiments demonstrate that our on-device latency-informed design combined with our tailored training strategy establishes a new state-of-the-art for efficient LVLM encoding, significantly outperforming existing encoder-centric baselines while operating on-device at nearly 1.7xthe speed.

cs.CV↗

arXiv2025

Energy-Efficient State Estimation with 1-Bit Sensing: A Bussgang-Kalman Framework for Internet of Things

Accurate state estimation from heavily quantized measurements is a key challenge in resource-constrained Internet of Things (IoT) sensing and tracking, where battery-powered devices may employ low-resolution analog-to-digital converters (ADCs) to simplify sensor hardware and reduce the amount of data. Existing model-based and hybrid learning-based estimators, however, typically assume high-resolution observations and therefore degrade severely under 1-bit quantization. In this paper, we study nonlinear state estimation with 1-bit observations and develop a Bussgang-aided filtering framework for IoT sensing front-ends with 1-bit quantization. For fully known system models, we propose a Bussgang-aided Kalman Filter (BKF) that explicitly incorporates quantization distortion into recursive estimation, together with a reduced-complexity variant (reduced-BKF) for computationally efficient implementation. For partially known models, we further propose Bussgang-aided KalmanNet (BKNet), a model-based deep learning architecture that combines adaptive dithering with gated recurrent units (GRUs) to mitigate severe quantization effects and model mismatch. Experiments on the Lorenz attractor and the Michigan NCLT dataset, both under 1-bit front-end quantization, demonstrate accurate and robust state estimation under highly nonlinear dynamics, imperfect models, and extreme quantization. These results support the potential of the proposed framework for reliable state estimation in resource-constrained IoT sensing and tracking applications with low-resolution front-ends.

eess.SP↗

arXiv2025

Learning MMSE Filters for OFDM Channel Estimation: Attention Transformer Gains at Linear Inference

In orthogonal frequency division multiplexing (OFDM), accurate channel estimation is crucial. Classical signal processing-based approaches, such as linear minimum mean-squared error (LMMSE) estimation, often require second-order statistics that are difficult to obtain in practice. Recent deep neural network (DNN)-based methods have been introduced to address this, but they often suffer from high inference complexity. This paper proposes an Attention-aided MMSE (A-MMSE), a model-based DNN framework that learns the linear MMSE filter via the Attention Transformer. Once trained, the A-MMSE performs channel estimation through a single linear operation, eliminating nonlinear activations during inference and thus reducing computational complexity. To improve the learning efficiency of the A-MMSE, we develop a two-stage Attention encoder that captures the frequency and temporal correlation structure of OFDM channels. We also introduce a rank-adaptive extension that adjusts the filter rank at deployment time, enabling efficient operation under resource-constrained receivers. Numerical simulations show that A-MMSE consistently outperforms baseline methods across a wide range of signal-to-noise ratio (SNR) conditions. In particular, the A-MMSE and its rank-adaptive extension provide an improved performance-complexity trade-off.

eess.SP↗