Wenjun Huang

Wenjun Huang, Yunduan Cui, Huiyun Li et al.

This paper addresses the prediction stability, prediction accuracy and control capability of the current probabilistic model-based reinforcement learning (MBRL) built on neural networks. A novel approach dropout-based probabilistic ensembles with trajectory sampling (DPETS) is proposed where the system uncertainty is stably predicted by combining the Monte-Carlo dropout and trajectory sampling in one framework. Its loss function is designed to correct the fitting error of neural networks for more accurate prediction of probabilistic models. The state propagation in its policy is extended to filter the aleatoric uncertainty for superior control capability. Evaluated by several Mujoco benchmark control tasks under additional disturbances and one practical robot arm manipulation task, DPETS outperforms related MBRL approaches in both average return and convergence velocity while achieving superior performance than well-known model-free baselines with significant sample efficiency. The open source code of DPETS is available at https://github.com/mrjun123/DPETS.

Anzhu Yu, Wenjun Huang, Qing Xu et al.

The deep learning, which is a dominating technique in artificial intelligence, has completely changed the image understanding over the past decade. As a consequence, the sea ice extraction (SIE) problem has reached a new era. We present a comprehensive review of four important aspects of SIE, including algorithms, datasets, applications, and the future trends. Our review focuses on researches published from 2016 to the present, with a specific focus on deep learning-based approaches in the last five years. We divided all relegated algorithms into 3 categories, including classical image segmentation approach, machine learning-based approach and deep learning-based methods. We reviewed the accessible ice datasets including SAR-based datasets, the optical-based datasets and others. The applications are presented in 4 aspects including climate research, navigation, geographic information systems (GIS) production and others. It also provides insightful observations and inspiring future research directions.

Hanning Chen, Yang Ni, Wenjun Huang et al.

Vision Transformers (ViTs) have emerged as the backbone of many segmentation models, consistently achieving state-of-the-art (SOTA) performance. However, their success comes at a significant computational cost. Image token pruning is one of the most effective strategies to address this complexity. However, previous approaches fall short when applied to more complex task-oriented segmentation (TOS), where the class of each image patch is not predefined but dependent on the specific input task. This work introduces the Vision Language Guided Token Pruning (VLTP), a novel token pruning mechanism that can accelerate ViT-based segmentation models, particularly for TOS guided by multi-modal large language model (MLLM). We argue that ViT does not need to process every image token through all of its layers -- only the tokens related to reasoning tasks are necessary. We design a new pruning decoder to take both image tokens and vision-language guidance as input to predict the relevance of each image token to the task. Only image tokens with high relevance are passed to deeper layers of the ViT. Experiments show that the VLTP framework reduces the computational costs of ViT by approximately 25% without performance degradation and by around 40% with only a 1% performance drop. The code associated with this study can be found at this URL.

Wenjun Huang, Yang Ni, Arghavan Rezvani et al.

Human pose estimation (HPE) is crucial for various applications. However, deploying HPE algorithms in surveillance contexts raises significant privacy concerns due to the potential leakage of sensitive personal information (SPI) such as facial features, and ethnicity. Existing privacy-enhancing methods often compromise either privacy or performance, or they require costly additional modalities. We propose a novel privacy-enhancing system that generates privacy-enhanced portraits while maintaining high HPE performance. Our key innovations include the reversible recovery of SPI for authorized personnel and the preservation of contextual information. By jointly optimizing a privacy-enhancing module, a privacy recovery module, and a pose estimator, our system ensures robust privacy protection, efficient SPI recovery, and high-performance HPE. Experimental results demonstrate the system's robust performance in privacy enhancement, SPI recovery, and HPE.

Sanggeon Yun, Ryozo Masukawa, Minhyoung Na et al.

Autonomous systems and smart-industry deployments increasingly split computation across near-sensor, edge, and cloud resources, where tight energy, latency, and reliability budgets demand run-time adaptivity. In practice, deciding what to compute and transmit at each point is pivotal; yet as multimodal sensor suites (cameras, LiDAR/depth, etc.) proliferate at the edge, most prior approaches either (i) fuse modalities on powerful servers or (ii) apply uni-modal near-sensor filters that ignore cross-modal dependencies, leading to redundant transmissions or missed events. We present FusionSense, a fusion-aware intelligent sensing framework for energy-constrained autonomous edge systems. Lightweight near-sensor classifiers are trained via a three-step procedure: (i) a server-side fusion model learns the downstream task, (ii) filter-out-safe (FoS) labels quantify each modality's necessity relative to the fused decision, and (iii) an edge-side fusion model is compacted by injecting near-sensor predictions as auxiliary signals. The result is a run-time decision layer that jointly reduces compute and communication while scaling linearly with sensor count. On a dual-modality (RGB+Depth/LiDAR) setup with SynDrone, FusionSense sustains task quality at substantially higher data-reduction rates than uni-modal filters and delivers large end-to-end gains: up to 33x lower energy at 1% FoI prevalence, 11x at 10%, a 92.3% reduction in quality loss at a fixed 30% data reduction, and roughly 1.5x higher energy savings than the best prior filtering baseline.

Wenjun Huang, Shenghao Fu, Yian Jin et al.

RAW images captured by different camera sensors exhibit substantial domain shifts due to varying spectral responses, noise characteristics, and tone behaviors, complicating their direct use in downstream computer vision tasks. Prior methods address this problem by training domain-specific RAW-to-RAW translators for each source-target pair, but such approaches do not scale to real-world scenarios involving multiple types of commercial cameras. In this work, we introduce MERIT, the first unified framework for multi-domain RAW image translation, which leverages a single model to perform translations across arbitrary camera domains. To address domain-specific noise discrepancies, we propose a sensor-aware noise modeling loss that explicitly aligns the signal-dependent noise statistics of the generated images with those of the target domain. We further enhance the generator with a conditional multi-scale large kernel attention module for improved context and sensor-aware feature modeling. To facilitate standardized evaluation, we introduce MDRAW, the first dataset tailored for multi-domain RAW image translation, comprising both paired and unpaired RAW captures from five diverse camera sensors across a wide range of scenes. Extensive experiments demonstrate that MERIT outperforms prior models in both quality (5.56 dB improvement) and scalability (80% reduction in training iterations).

Yezi Liu, Hanning Chen, Wenjun Huang et al.

Large language models (LLMs) possess vast knowledge acquired from extensive training corpora, but they often cannot remove specific pieces of information when needed, which makes it hard to handle privacy, bias mitigation, and knowledge correction. Traditional model unlearning approaches require computationally expensive fine-tuning or direct weight editing, making them impractical for real-world deployment. In this work, we introduce LoRA-based Unlearning with Negative Examples (LUNE), a lightweight framework that performs negative-only unlearning by updating only low-rank adapters while freezing the backbone, thereby localizing edits and avoiding disruptive global changes. Leveraging Low-Rank Adaptation (LoRA), LUNE targets intermediate representations to suppress (or replace) requested knowledge with an order-of-magnitude lower compute and memory than full fine-tuning or direct weight editing. Extensive experiments on multiple factual unlearning tasks show that LUNE: (I) achieves effectiveness comparable to full fine-tuning and memory-editing methods, and (II) reduces computational cost by about an order of magnitude.

Yezi Liu, Hanning Chen, Wenjun Huang et al.

Unlearning in large foundation models (e.g., LLMs) is essential for enabling dynamic knowledge updates, enforcing data deletion rights, and correcting model behavior. However, existing unlearning methods often require full-model fine-tuning or access to the original training data, which limits their scalability and practicality. In this work, we introduce Recover-to-Forget (R2F), a novel framework for efficient unlearning in LLMs based on reconstructing full-model gradient directions from low-rank LoRA adapter updates. Rather than performing backpropagation through the full model, we compute gradients with respect to LoRA parameters using multiple paraphrased prompts and train a gradient decoder to approximate the corresponding full-model gradients. To ensure applicability to larger or black-box models, the decoder is trained on a proxy model and transferred to target models. We provide a theoretical analysis of cross-model generalization and demonstrate that our method achieves effective unlearning while preserving general model performance. Experimental results demonstrate that R2F offers a scalable and lightweight alternative for unlearning in pretrained LLMs without requiring full retraining or access to internal parameters.

Yezi Liu, Hanning Chen, Wenjun Huang et al.

Group fairness in machine learning is often enforced by adding a regularizer that reduces the dependence between model predictions and sensitive attributes. However, existing regularizers are built on heterogeneous distance measures and design choices, which makes their behavior hard to reason about and their performance inconsistent across tasks. This raises a basic question: what properties make a good fairness regularizer? We address this question by first organizing existing in-process methods into three families: (i) matching prediction statistics across sensitive groups, (ii) aligning latent representations, and (iii) directly minimizing dependence between predictions and sensitive attributes. Through this lens, we identify desirable properties of the underlying distance measure, including tight generalization bounds, robustness to scale differences, and the ability to handle arbitrary prediction distributions. Motivated by these properties, we propose a Cauchy-Schwarz (CS) fairness regularizer that penalizes the empirical CS divergence between prediction distributions conditioned on sensitive groups. Under a Gaussian comparison, we show that CS divergence yields a tighter bound than Kullback-Leibler divergence, Maximum Mean Discrepancy, and the mean disparity used in Demographic Parity, and we discuss how these advantages translate to a distribution-free, kernel-based estimator that naturally extends to multiple sensitive attributes. Extensive experiments on four tabular benchmarks and one image dataset demonstrate that the proposed CS regularizer consistently improves Demographic Parity and Equal Opportunity metrics while maintaining competitive accuracy, and achieves a more stable utility-fairness trade-off across hyperparameter settings compared to prior regularizers.

Sungheon Jeong, Ryozo Masukawa, Jihong Park et al.

While recent Large Vision-Language Models (LVLMs) exhibit strong multimodal reasoning abilities, they often produce ungrounded or hallucinated responses because they rely too heavily on linguistic priors instead of visual evidence. This limitation highlights the absence of a quantitative measure of how much these models actually use visual information during reasoning. We propose Draft and Refine (DnR), an agent framework driven by a question-conditioned utilization metric. The metric quantifies the model's reliance on visual evidence by first constructing a query-conditioned relevance map to localize question-specific cues and then measuring dependence through relevance-guided probabilistic masking. Guided by this metric, the DnR agent refines its initial draft using targeted feedback from external visual experts. Each expert's output (such as boxes or masks) is rendered as visual cues on the image, and the model is re-queried to select the response that yields the largest improvement in utilization. This process strengthens visual grounding without retraining or architectural changes. Experiments across VQA and captioning benchmarks show consistent accuracy gains and reduced hallucination, demonstrating that measuring visual utilization provides a principled path toward more interpretable and evidence-driven multimodal agent systems.

Heng Zheng, Yuling Shi, Xiaodong Gu et al.

Visual geo-localization requires extensive geographic knowledge and sophisticated reasoning to determine image locations without GPS metadata. Traditional retrieval methods are constrained by database coverage and quality. Recent Large Vision-Language Models (LVLMs) enable direct location reasoning from image content, yet individual models struggle with diverse geographic regions and complex scenes. Existing multi-agent systems improve performance through model collaboration but treat all agent interactions uniformly. They lack mechanisms to handle conflicting predictions effectively. We propose \textbf{GraphGeo}, a multi-agent debate framework using heterogeneous graph neural networks for visual geo-localization. Our approach models diverse debate relationships through typed edges, distinguishing supportive collaboration, competitive argumentation, and knowledge transfer. We introduce a dual-level debate mechanism combining node-level refinement and edge-level argumentation modeling. A cross-level topology refinement strategy enables co-evolution between graph structure and agent representations. Experiments on multiple benchmarks demonstrate GraphGeo significantly outperforms state-of-the-art methods. Our framework transforms cognitive conflicts between agents into enhanced geo-localization accuracy through structured debate.

Meiyu Qiu, Yun Li, Wenjun Huang et al.

Laryngeal cancer is a malignant disease with a high morality rate in otorhinolaryngology, posing an significant threat to human health. Traditionally larygologists manually visual-inspect laryngeal cancer in laryngoscopic videos, which is quite time-consuming and subjective. In this study, we propose a novel automatic framework via 3D-large-scale pretrained models termed 3D-LSPTM for laryngeal cancer detection. Firstly, we collect 1,109 laryngoscopic videos from the First Affiliated Hospital Sun Yat-sen University with the approval of the Ethics Committee. Then we utilize the 3D-large-scale pretrained models of C3D, TimeSformer, and Video-Swin-Transformer, with the merit of advanced featuring videos, for laryngeal cancer detection with fine-tuning techniques. Extensive experiments show that our proposed 3D-LSPTM can achieve promising performance on the task of laryngeal cancer detection. Particularly, 3D-LSPTM with the backbone of Video-Swin-Transformer can achieve 92.4% accuracy, 95.6% sensitivity, 94.1% precision, and 94.8% F_1.

Wenjun Huang, Jiakai Pan, Jiahao Tang et al.

Multimodal Large Language Models (MLLMs) have attracted much attention for their multifunctionality. However, traditional Transformer architectures incur significant overhead due to their secondary computational complexity. To address this issue, we introduce ML-Mamba, a multimodal language model, which utilizes the latest and efficient Mamba-2 model for inference. Mamba-2 is known for its linear scalability and fast processing of long sequences. We replace the Transformer-based backbone with a pre-trained Mamba-2 model and explore methods for integrating 2D visual selective scanning mechanisms into multimodal learning while also trying various visual encoders and Mamba-2 model variants. Our extensive experiments in various multimodal benchmark tests demonstrate the competitive performance of ML-Mamba and highlight the potential of state space models in multimodal tasks. The experimental results show that: (1) we empirically explore how to effectively apply the 2D vision selective scan mechanism for multimodal learning. We propose a novel multimodal connector called the Mamba-2 Scan Connector (MSC), which enhances representational capabilities. (2) ML-Mamba achieves performance comparable to state-of-the-art methods such as TinyLaVA and MobileVLM v2 through its linear sequential modeling while faster inference speed; (3) Compared to multimodal models utilizing Mamba-1, the Mamba-2-based ML-Mamba exhibits superior inference performance and effectiveness.

Hanning Chen, Wenjun Huang, Yang Ni et al.

Task-oriented object detection aims to find objects suitable for accomplishing specific tasks. As a challenging task, it requires simultaneous visual data processing and reasoning under ambiguous semantics. Recent solutions are mainly all-in-one models. However, the object detection backbones are pre-trained without text supervision. Thus, to incorporate task requirements, their intricate models undergo extensive learning on a highly imbalanced and scarce dataset, resulting in capped performance, laborious training, and poor generalizability. In contrast, we propose TaskCLIP, a more natural two-stage design composed of general object detection and task-guided object selection. Particularly for the latter, we resort to the recently successful large Vision-Language Models (VLMs) as our backbone, which provides rich semantic knowledge and a uniform embedding space for images and texts. Nevertheless, the naive application of VLMs leads to sub-optimal quality, due to the misalignment between embeddings of object images and their visual attributes, which are mainly adjective phrases. To this end, we design a transformer-based aligner after the pre-trained VLMs to re-calibrate both embeddings. Finally, we employ a trainable score function to post-process the VLM matching results for object selection. Experimental results demonstrate that our TaskCLIP outperforms the state-of-the-art DETR-based model TOIST by 3.5% and only requires a single NVIDIA RTX 4090 for both training and inference.

Sanggeon Yun, Hanning Chen, Ryozo Masukawa et al.

Introducing HyperSense, our co-designed hardware and software system efficiently controls Analog-to-Digital Converter (ADC) modules' data generation rate based on object presence predictions in sensor data. Addressing challenges posed by escalating sensor quantities and data rates, HyperSense reduces redundant digital data using energy-efficient low-precision ADC, diminishing machine learning system costs. Leveraging neurally-inspired HyperDimensional Computing (HDC), HyperSense analyzes real-time raw low-precision sensor data, offering advantages in handling noise, memory-centricity, and real-time learning. Our proposed HyperSense model combines high-performance software for object detection with real-time hardware prediction, introducing the novel concept of Intelligent Sensor Control. Comprehensive software and hardware evaluations demonstrate our solution's superior performance, evidenced by the highest Area Under the Curve (AUC) and sharpest Receiver Operating Characteristic (ROC) curve among lightweight models. Hardware-wise, our FPGA-based domain-specific accelerator tailored for HyperSense achieves a 5.6x speedup compared to YOLOv4 on NVIDIA Jetson Orin while showing up to 92.1% energy saving compared to the conventional system. These results underscore HyperSense's effectiveness and efficiency, positioning it as a promising solution for intelligent sensing and real-time data processing across diverse applications.

Wenjun Huang, Arghavan Rezvani, Hanning Chen et al.

Applications in the Internet of Things (IoT) utilize machine learning to analyze sensor-generated data. However, a major challenge lies in the lack of targeted intelligence in current sensing systems, leading to vast data generation and increased computational and communication costs. To address this challenge, we propose a novel sensing module to equip sensing frameworks with intelligent data transmission capabilities by integrating a highly efficient machine learning model placed near the sensor. This model provides prompt feedback for the sensing system to transmit only valuable data while discarding irrelevant information by regulating the frequency of data transmission. The near-sensor model is quantized and optimized for real-time sensor control. To enhance the framework's performance, the training process is customized and a "lazy" sensor deactivation strategy utilizing temporal information is introduced. The suggested method is orthogonal to other IoT frameworks and can be considered as a plugin for selective data transmission. The framework is implemented, encompassing both software and hardware components. The experiments demonstrate that the framework utilizing the suggested module achieves over 85% system efficiency in terms of energy consumption and storage, with negligible impact on performance. This methodology has the potential to significantly reduce data output from sensors, benefiting a wide range of IoT applications.

Wenjun Huang, Hanning Chen, Yang Ni et al.

Detecting marine objects inshore presents challenges owing to algorithmic intricacies and complexities in system deployment. We propose a difficulty-aware edge-cloud collaborative sensing system that splits the task into object localization and fine-grained classification. Objects are classified either at the edge or within the cloud, based on their estimated difficulty. The framework comprises a low-power device-tailored front-end model for object localization, classification, and difficulty estimation, along with a transformer-graph convolutional network-based back-end model for fine-grained classification. Our system demonstrates superior performance (mAP@0.5 +4.3%}) on widely used marine object detection datasets, significantly reducing both data transmission volume (by 95.43%) and energy consumption (by 72.7%}) at the system level. We validate the proposed system across various embedded system platforms and in real-world scenarios involving drone deployment.

Sanggeon Yun, Ryozo Masukawa, Hanning Chen et al.

The escalating challenges of managing vast sensor-generated data, particularly in audio applications, necessitate innovative solutions. Current systems face significant computational and storage demands, especially in real-time applications like gunshot detection systems (GSDS), and the proliferation of edge sensors exacerbates these issues. This paper proposes a groundbreaking approach with a near-sensor model tailored for intelligent audio-sensing frameworks. Utilizing a Fast Fourier Transform (FFT) module, convolutional neural network (CNN) layers, and HyperDimensional Computing (HDC), our model excels in low-energy, rapid inference, and online learning. It is highly adaptable for efficient ASIC design implementation, offering superior energy efficiency compared to conventional embedded CPUs or GPUs, and is compatible with the trend of shrinking microphone sensor sizes. Comprehensive evaluations at both software and hardware levels underscore the model's efficacy. Software assessments through detailed ROC curve analysis revealed a delicate balance between energy conservation and quality loss, achieving up to 82.1% energy savings with only 1.39% quality loss. Hardware evaluations highlight the model's commendable energy efficiency when implemented via ASIC design, especially with the Google Edge TPU, showcasing its superiority over prevalent embedded CPUs and GPUs.

Yang Ni, Zhuowen Zou, Wenjun Huang et al.

Drawing inspiration from the outstanding learning capability of our human brains, Hyperdimensional Computing (HDC) emerges as a novel computing paradigm, and it leverages high-dimensional vector presentation and operations for brain-like lightweight Machine Learning (ML). Practical deployments of HDC have significantly enhanced the learning efficiency compared to current deep ML methods on a broad spectrum of applications. However, boosting the data efficiency of HDC classifiers in supervised learning remains an open question. In this paper, we introduce Hyperdimensional Efficient Active Learning (HEAL), a novel Active Learning (AL) framework tailored for HDC classification. HEAL proactively annotates unlabeled data points via uncertainty and diversity-guided acquisition, leading to a more efficient dataset annotation and lowering labor costs. Unlike conventional AL methods that only support classifiers built upon deep neural networks (DNN), HEAL operates without the need for gradient or probabilistic computations. This allows it to be effortlessly integrated with any existing HDC classifier architecture. The key design of HEAL is a novel approach for uncertainty estimation in HDC classifiers through a lightweight HDC ensemble with prior hypervectors. Additionally, by exploiting hypervectors as prototypes (i.e., compact representations), we develop an extra metric for HEAL to select diverse samples within each batch for annotation. Our evaluation shows that HEAL surpasses a diverse set of baselines in AL quality and achieves notably faster acquisition than many BNN-powered or diversity-guided AL methods, recording 11 times to 40,000 times speedup in acquisition runtime per batch.

Ryozo Masukawa, Sanggeon Yun, Sungheon Jeong et al.

Traffic classification is vital for cybersecurity, yet encrypted traffic poses significant challenges. We present PacketCLIP, a multi-modal framework combining packet data with natural language semantics through contrastive pretraining and hierarchical Graph Neural Network (GNN) reasoning. PacketCLIP integrates semantic reasoning with efficient classification, enabling robust detection of anomalies in encrypted network flows. By aligning textual descriptions with packet behaviors, it offers enhanced interpretability, scalability, and practical applicability across diverse security scenarios. PacketCLIP achieves a 95% mean AUC, outperforms baselines by 11.6%, and reduces model size by 92%, making it ideal for real-time anomaly detection. By bridging advanced machine learning techniques and practical cybersecurity needs, PacketCLIP provides a foundation for scalable, efficient, and interpretable solutions to tackle encrypted traffic classification and network intrusion detection challenges in resource-constrained environments.

Wenjun Huang, Yang Ni, Hanning Chen et al.

Referring multi-object tracking (RMOT) is an emerging cross-modal task that aims to localize an arbitrary number of targets based on a language expression and continuously track them in a video. This intricate task involves reasoning on multi-modal data and precise target localization with temporal association. However, prior studies overlook the imbalanced data distribution between newborn targets and existing targets due to the nature of the task. In addition, they only indirectly fuse multi-modal features, struggling to deliver clear guidance on newborn target detection. To solve the above issues, we conduct a collaborative matching strategy to alleviate the impact of the imbalance, boosting the ability to detect newborn targets while maintaining tracking performance. In the encoder, we integrate and enhance the cross-modal and multi-scale fusion, overcoming the bottlenecks in previous work, where limited multi-modal information is shared and interacted between feature maps. In the decoder, we also develop a referring-infused adaptation that provides explicit referring guidance through the query tokens. The experiments showcase the superior performance of our model (+3.42%) compared to prior works, demonstrating the effectiveness of our designs.

Sungheon Jeong, Hanning Chen, Sanggeon Yun et al.

This paper introduces a powerful encoder that transfers CLIP`s capabilities to event-based data, enhancing its utility and expanding its applicability across diverse domains. While large-scale datasets have significantly advanced image-based models, the scarcity of comprehensive event datasets has limited performance potential in event modality. To address this challenge, we adapt CLIP`s architecture to align event embeddings with image embeddings, supporting zero-shot learning and preserving text alignment while mitigating catastrophic forgetting. Our encoder achieves strong performance in object recognition, with competitive results in zero-shot and few-shot learning tasks. Notably, it generalizes effectively to events extracted from video data without requiring additional training, highlighting its versatility. Additionally, we integrate this encoder within a cross-modality framework that facilitates interaction across five modalities-Image, Event, Text, Sound, and Depth-expanding the possibilities for cross-modal applications. Overall, this work underscores the transformative potential of a robust event encoder, broadening the scope and utility of event-based data across various fields.

Ryozo Masukawa, Sanggeon Yun, Hyunwoo Oh et al.

Recent progress in reinforcement learning with verifiable rewards (RLVR) shows that small, specialized language models (SLMs) can exhibit structured reasoning without relying on large monolithic LLMs. We introduce soft hidden-state collaboration, where multiple heterogeneous frozen SLM experts are integrated through their internal representations via a trainable attention interface. Experiments on Reasoning Gym and GSM8K show that this latent integration is competitive with strong single-model RLVR baselines. Ablations further reveal a dual mechanism of expert utilization: for simpler arithmetic domains, performance gains can largely be explained by static expert preferences, whereas more challenging settings induce increasingly concentrated and structured expert attention over training, indicating emergent specialization in how the router connects to relevant experts. Overall, hidden-state collaboration provides a compact mechanism for leveraging frozen experts, while offering an observational window into expert utilization patterns and their evolution under RLVR.

Hyunwoo Oh, Hanning Chen, Sanggeon Yun et al.

Deformable transformers deliver state-of-the-art detection but map poorly to hardware due to irregular memory access and low arithmetic intensity. We introduce QUILL, a schedule-aware accelerator that turns deformable attention into cache-friendly, single-pass work. At its core, Distance-based Out-of-Order Querying (DOOQ) orders queries by spatial proximity; the look-ahead drives a region prefetch into an alternate buffer--forming a schedule-aware prefetch loop that overlaps memory and compute. A fused MSDeformAttn engine executes interpolation, Softmax, aggregation, and the final projection (W''m) in one pass without spilling intermediates, while small tensors are kept on-chip and surrounding dense layers run on integrated GEMMs. Implemented as RTL and evaluated end-to-end, QUILL achieves up to 7.29x higher throughput and 47.3x better energy efficiency than an RTX 4090, and exceeds prior accelerators by 3.26-9.82x in throughput and 2.01-6.07x in energy efficiency. With mixed-precision quantization, accuracy tracks FP32 within <=0.9 AP across Deformable and Sparse DETR variants. By converting sparsity into locality--and locality into utilization--QUILL delivers consistent, end-to-end speedups.

Yezi Liu, Prathyush Poduval, Wenjun Huang et al.

Graph unlearning is a crucial approach for protecting user privacy by erasing the influence of user data on trained graph models. Recent developments in graph unlearning methods have primarily focused on maintaining model prediction performance while removing user information. However, we have observed that when user information is deleted from the model, the prediction distribution across different sensitive groups often changes. Furthermore, graph models are shown to be prone to amplifying biases, making the study of fairness in graph unlearning particularly important. This raises the question: Does graph unlearning actually introduce bias? Our findings indicate that the predictions of post-unlearning models become highly correlated with sensitive attributes, confirming the introduction of bias in the graph unlearning process. To address this issue, we propose a fair graph unlearning method, FGU. To guarantee privacy, FGU trains shard models on partitioned subgraphs, unlearns the requested data from the corresponding subgraphs, and retrains the shard models on the modified subgraphs. To ensure fairness, FGU employs a bi-level debiasing process: it first enables shard-level fairness by incorporating a fairness regularizer in the shard model retraining, and then achieves global-level fairness by aligning all shard models to minimize global disparity. Our experiments demonstrate that FGU achieves superior fairness while maintaining privacy and accuracy. Additionally, FGU is robust to diverse unlearning requests, ensuring fairness and utility performance across various data distributions.

Yingzhuo Jiang, Wenjun Huang, Rongdun Lin et al.

This paper tackles the challenge of learning multi-goal dexterous hand manipulation tasks using model-based Reinforcement Learning. We propose Goal-Conditioned Probabilistic Model Predictive Control (GC-PMPC) by designing probabilistic neural network ensembles to describe the high-dimensional dexterous hand dynamics and introducing an asynchronous MPC policy to meet the control frequency requirements in real-world dexterous hand systems. Extensive evaluations on four simulated Shadow Hand manipulation scenarios with randomly generated goals demonstrate GC-PMPC's superior performance over state-of-the-art baselines. It successfully drives a cable-driven Dexterous hand, DexHand 021 with 12 Active DOFs and 5 tactile sensors, to learn manipulating a cubic die to three goal poses within approximately 80 minutes of interactions, demonstrating exceptional learning efficiency and control performance on a cost-effective dexterous hand platform.

Hanning Chen, Yang Ni, Wenjun Huang et al.

Large Vision Language Models (LVLMs) have been widely adopted to guide vision foundation models in performing reasoning segmentation tasks, achieving impressive performance. However, the substantial computational overhead associated with LVLMs presents a new challenge. The primary source of this computational cost arises from processing hundreds of image tokens. Therefore, an effective strategy to mitigate such overhead is to reduce the number of image tokens, a process known as image token pruning. Previous studies on image token pruning for LVLMs have primarily focused on high level visual understanding tasks, such as visual question answering and image captioning. In contrast, guiding vision foundation models to generate accurate visual masks based on textual queries demands precise semantic and spatial reasoning capabilities. Consequently, pruning methods must carefully control individual image tokens throughout the LVLM reasoning process. Our empirical analysis reveals that existing methods struggle to adequately balance reductions in computational overhead with the necessity to maintain high segmentation accuracy. In this work, we propose LVLM_CSP, a novel training free visual token pruning method specifically designed for LVLM based reasoning segmentation tasks. LVLM_CSP consists of three stages: clustering, scattering, and pruning. Initially, the LVLM performs coarse-grained visual reasoning using a subset of selected image tokens. Next, fine grained reasoning is conducted, and finally, most visual tokens are pruned in the last stage. Extensive experiments demonstrate that LVLM_CSP achieves a 65% reduction in image token inference FLOPs with virtually no accuracy degradation, and a 70% reduction with only a minor 1% drop in accuracy on the 7B LVLM.

Wenjun Huang, Jianguo Hu

The Long Short-Term Memory (LSTM) networks have traditionally faced challenges in scaling and effectively capturing complex dependencies in visual tasks. The xLSTM architecture has emerged to address these limitations, incorporating exponential gating and a parallel matrix memory structure to enhance performance and scalability. Despite these advancements, the potential of xLSTM in visual computing has not been fully realized, particularly in leveraging autoregressive techniques for improved feature extraction. In this paper, we introduce MAL (Cluster-Masked and Multi-Task Pretraining for Enhanced xLSTM Vision Performance), a novel framework that enhances xLSTM's capabilities through innovative pretraining strategies. We propose a cluster-masked masking method that significantly improves local feature capture and optimizes image scanning efficiency. Additionally, our universal encoder-decoder pretraining approach integrates multiple tasks, including image autoregression, depth estimation, and image segmentation, thereby enhancing the model's adaptability and robustness across diverse visual tasks. Our experimental results demonstrate that MAL surpasses traditional supervised models and fully leverages the scaling potential of xLSTM, setting a new benchmark in visual task performance.