Hong Kong University of Science and Technology
Hong Kong University of Science and Technology Institutional RepositoryNot a member yet
162821 research outputs found
Sort by
On the Development and Evaluation of Asynchronous Reinforcement Learning for Foundation Models
No full text
Temporal Model-Based Federated Active Medical Image Classification
No full textTraditional federated learning relies on fully labeled datasets in each medical institution, which is impractical in real-world clinical scenarios. Federated Active Learning (FAL) addresses this by selecting a few informative samples for labeling, but it faces challenges such as domain shift across institutions. Besides, existing FAL methods rely on single-round model knowledge to estimate prediction-level uncertainty, ignoring uncertainty from features and model evolution during training. In this work, we propose TM-FAL, a novel framework for federated active medical image classification under domain shift. TM-FAL proposes a new uncertainty by integrating feature differences and prediction confidence from temporal local and global models to capture both local-global differences and the inherent complexity of images. Additionally, we use the prediction of the global model as pseudo labels to group images to mitigate class imbalance caused by uncertainty-based selection. Experiments on two medical image classification datasets demonstrate that TM-FAL outperforms various state-of-the-art methods.</p
Blockchain-enabled reliable outsourced decryption CP-ABE using responsive zkSNARK for mobile computing
No full textCiphertext-Policy Attribute-Based Encryption (CP-ABE) is a promising solution for access control in mobile computing. However, the heavy decryption overhead hinders its widespread adoption. A general approach to address this issue is to outsource decryption to a decryption cloud server (DCS). Existing schemes achieve verifiability but lack an effective exemption mechanism to protect honest DCS from false claims. In this paper, we propose a blockchain-enabled reliable outsourced decryption CP-ABE framework that achieves both verifiability and exemptibility without adding redundant information to the ciphertext. We use zkSNARK to verify outsourced results on blockchain efficiently and introduce a challenge-response mechanism to address the high cost of proof generation. Moreover, our framework ensures fair incentive and enables decentralized outsourcing through blockchain. Finally, we implement and evaluate our scheme on Ethereum to demonstrate its feasibility and efficiency. While maintaining almost the same decryption cost, our gas usage is 11× to 140× in the happy case and 4× to 55× in the challenge case lower than the scheme of Ge et al. (TDSC’24) in attribute numbers from 5 to 60. Building upon the proposed framework, we demonstrate its application in the data sharing of electric vehicles, enabling a more extensive use of mobile computing resources.</p
Metabolic model-guided strain design for improved succinic acid production in Yarrowia lipolytica
No full textThe oleaginous yeast Yarrowia lipolytica has emerged as a promising microbial host for the production of succinic acid (SA), a key bio-based platform chemical, owing to its metabolic versatility and robustness under industrial fermentation conditions. To enable rational engineering of Y. lipolytica toward enhanced SA production, a genome-scale metabolic model (GEM) of the industrially relevant W29 strain has been reconstructed in this study, comprising 634 genes, 1130 metabolites, and 1364 reactions distributed across eight compartments. The model achieved 88.9 % accuracy in predicting growth phenotypes on 18 carbon sources and demonstrated a robust correlation with experimental growth rates (R2 = 0.98). Leveraging in silico strain design tools, a set of knockout and overexpression targets for enhancing SA production from glycerol was identified. Simulations revealed that knocking out succinate dehydrogenase (SDH) and acetyl-CoA hydrolase (ACH) increased SA flux to 4.36 mmol/gDW/h (0.56 g/g glycerol), aligning with prior experimental studies. Overexpression of pyruvate carboxylase and TCA/glyoxylate cycle enzymes was predicted to further enhance SA yields by up to 186 %. These findings not only aligned with experimentally confirmed targets but also uncovered novel interventions, demonstrating the GEM as a robust platform for rational strain design toward enhanced production of SA and other bio-based chemicals.</p
Variability of urban riverine nutrients under coupled human-hydrological-biogeochemical framework
No full textNutrient fluxes exhibit complex dynamics under combined influences of land surface processes and human activities. In this study, we comprehensively investigate the long-term variability of riverine nutrient fluxes and uncover the underlying controlling mechanisms in a typical urban agglomeration by integrating observations with the Export Coefficient Model and the Soil and Water Assessment Tool within a coupled human-hydrological-biogeochemical framework. Our results show that the non-point sources (NPS) pollution is controlled by coupled transport processes and nutrient sources. High surface flows in urban and agricultural areas, actively combined with their respective nitrogen deposition and intensive fertilizer use, form active NPS zones. In forested regions, elevated nitrogen levels result from soil nitrate leaching in the lateral layer, where lateral flow is not a limiting factor. The elevated riverine nitrogen flux from NPS, relative to phosphorus, results from the combined effects of abundant nutrient inputs and a disproportionately higher export rate of nitrogen. In contrast, point sources are primarily driven by domestic wastewater, which establishes a persistent core-periphery pollution structure in urban agglomerations, with intensity decreasing from domestic-heavy and hybrid-sourced areas at the core to agro-centric regions at the periphery. However, this structure weakens as domestic sources decline significantly while non-domestic sources remain dynamically balanced. Regarding instream processes, nutrient transformation and removal in channels are positively influenced by organic nutrient ratios and transport distance, whereas upstream influx has a relatively minor impact. These findings provide valuable insights into riverine nutrient pollution from various sources across diverse landscapes in urban agglomerations worldwide.</p
Enhanced multi-leak detection in pressurized pipelines using super-resolution matched-field processing
No full textThe detection of a cluster of leaks in pressurized water supply pipelines is crucial as it allows for early recognition of the pipeline's poor overall condition, enabling timely interventions. This paper addresses this challenge by extending, for the first time, a high-resolution modified broadband matched-field processing (MB-MFP) framework to the multi-leak detection problem. The developed methods combine advanced signal processing with acoustic measurements to achieve super-resolution identification, where leaks spaced closer than half the minimum probing wavelength are successfully resolved. A key contribution is a methodology that enables the simultaneous estimation of the number, locations, and sizes of multiple leaks. We also show that the techniques are robust to various noise types and levels, including white noise with a signal-to-noise ratio (SNR) as low as −10 dB, as well as colored and impulse noise. The effectiveness and robustness of the proposed techniques are demonstrated through comprehensive numerical simulations and validated with laboratory experiments.</p
Enhancing interface adhesion of 3D printable concrete by biochar integration
No full textInterface adhesion is critical to ensuring the integrity of 3D printable concrete (3DPC) structures, but is susceptible to the negative effects of the printing environment and printing time interval. Here, we innovatively propose an approach to improve interface adhesion for 3DPC using engineered biochar. Our results demonstrated that biochar reduced the dynamic yield stress by 16 % and plastic viscosity by 52 % without compromising the static yield stress and improved the flow velocity gradient in the extrusion process. The characteristics of the printing process shifted from pore generation to pore removal, and interfacial porosity was reduced by 61 %, thus strengthening the interface adhesion for 3DPC. The increases in interfacial microcrack width and porosity were linearly correlated with the surface moisture loss and the growth in rigidity of the substrate layers, respectively. Biochar, as a porous biocarbon with the capacity of moisture regulation, could mitigate the moisture loss and rigidness growth for the substrate layers, reducing the interfacial microcrack width by 64.4 % and porosity by 23.8 % and enhancing bond strength by 21.3 % for 3DPC at the time interval of 60 min. This low-carbon enhancement interface adhesion would broaden the applicability and resilience of concrete printing.</p
SR-SAM: Subspace Regularization for Domain Generalization of Segment Anything Model
No full textParameter Efficient Fine-Tuning (PEFT) methods have been widely used to adapt foundation models like the Segment Anything Model (SAM) for better generalization in unseen domains. Despite their widespread use, PEFT often suffers from overfitting to the source training domain, which limits their generalization performance. To address this limitation, we propose a novel subspace regularization (SR) method for robust fine-tuning. Our approach iteratively removes the knowledge of task-specific directions, as identified by LoRA parameters learned from the source domain, from the subspace of pre-trained weights. This strategy effectively encourages the LoRA parameters to acquire a more diverse range of knowledge. In addition, we introduce an exponential moving average (EMA) LoRA module that aggregates historical updates of the LoRA parameters throughout the fine-tuning process. This aggregation enhances stability and the generalizability of the learned features by smoothing the trajectory of parameter updates. Our enhanced framework, SR-SAM, incorporates both subspace regularization and the EMA LoRA module to fine-tune the popular SAM model effectively. Experimental results on two widely used domain generalization benchmarks demonstrate that SR-SAM outperforms existing state-of-the-art methods, underscoring the effectiveness of our method. The source code is available at https://github.com/xjiangmed/SR-SAM.</p
NIFA: Low-dose CT imaging via noise intensity field aware networks
No full textComputed tomography (CT) is one of the most widely used imaging modalities in clinical practice. While profound for disease diagnosis, the extensive use of CT has contributed to the major part of population-based radiation dose, raising public concerns about the potential risk of cancer. Therefore, low-dose CT (LDCT) has attracted much attention in the past decades. LDCT lowers x-ray dose to reduce health risks during data acquisition, but it introduces excessive image noise and artifacts, compromising image quality, which hinders its clinical use. To address this problem, studies using data-driven deep neural networks to improve the LDCT image quality have been investigated. Here we tackle the LDCT challenge by leveraging the data-driven paradigm together with CT imaging physics to develop a more clinically relevant predictive model. We formulate noise in LDCT images as a noise intensity field and denoising process as intensity value regression. Based on the formulation, a noise intensity field aware (NIFA) network which separately extracts low-intensity and high-intensity information is proposed to reduce the magnitude of the intensity field while preserving texture information and anatomical details. Extensive experiments are conducted to evaluate the performance of the proposed method, including ablation studies to demonstrate the effectiveness of the innovative design.</p