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
Federated Cubic Regularized Newton Learning with sparsification-amplified differential privacy
No full textThis paper explores the cubic-regularized Newton method within a federated learning framework while addressing two major concerns: privacy leakage and communication bottlenecks. We propose the Differentially Private Federated Cubic Regularized Newton (DP-FCRN) algorithm, which leverages second-order techniques to achieve lower iteration complexity than first-order methods. We incorporate noise perturbation during local computations to ensure privacy. Furthermore, we employ sparsification in uplink transmission, which not only reduces the communication costs but also amplifies the privacy guarantee. Specifically, this approach reduces the necessary noise intensity without compromising privacy protection. We analyze the convergence properties of our algorithm and establish the privacy guarantee. Finally, we validate the effectiveness of the proposed algorithm through experiments on a benchmark dataset.</p
VIX Term Structure in the Rough Heston Model via Markovian Approximation
No full textWe model the VIX term structure using the rough Heston model. Since the direct numerical modeling of the rough Heston model is computationally inefficient, we adopt a Markovian approximation approach. Building on the Markovian framework, we eliminate the need for simulation by exploiting an analytical expression for VIX. The resulting formula for squared VIX under the Markovian approximation provides an analytical approximation to its counterpart under the rough Heston model. Another efficiency in the calibration procedure is achieved by exploiting the analytical gradient formulas of squared VIX. Empirically, using an extensive dataset of daily VIX term structures, we show that the rough Heston model outperforms various competing Heston-type models with jumps in both in-sample and out-of-sample fit and yields more reliable estimates of spot volatility, validating that rough volatility is preferred to jumps in modeling VIX term structure.</p
A Compact SMA-Driven Push–Pull Module With Precise Linear Motion
No full textShape memory alloy (SMA) actuators have demonstrated great value in soft robotic applications; however, traditional tendon-based actuation methods suffer from hysteresis that affects control accuracy. In this article, we propose a SMA-driven push–pull module by using antagonistic actuation and rod connection, and further integrate a linear thin film potentiometer within a compact size, which can achieve precise pulling and pushing motions based on the model-based closed-loop control. The compact SMA module with a large stroke can fulfill the submillimeter position accuracy for pulling and pushing motions at varying speeds and loads. An application to a soft origami continuum robot proves that the proposed SMA module can solve the hysteresis problem, simultaneously achieving precise control of bending and contraction/expansion (RMSE ≤ 0.85 mm). It is also demonstrated that the SMA module can be MR-compatible and function well under MRI feedback. This study will contribute to the advancement of SMA actuators in precision motion control and soft robotic applications.</p
Turbulent intermittency in broadband rotor noise
No full textTurbulent flow in the boundary layer of rotor blades causes unsteady surface pressure and produces broadband noise. Turbulent intermittency affects surface pressure and creates sporadic high-strength noise sources. In this study, we experimentally demonstrate that the statistical results of rotor noise amplitude follow a log-normal distribution. A wavelet-based beamforming method is used to capture transient noise sources on a 0.22 m-diameter rotor. We suggest that noise sources at different frequencies can be regarded as statistically homogeneous, and a distribution function exists to unify the strength distributions of noise sources at different frequencies. The similarity of rotor noise is identified by the unified distribution. Results show that the mean and variance of the transient noise source strength are controlled by the rotating speed and the surface pressure spectrum in the form of μ≈Cμ+50log10Ω and σ2∝(Cσ−50log10ω)2. Here Cμ is related to the unsteady surface pressure spectra, Cσ is a constant, and ω and Ω are the angular frequency and rotating speed, respectively. A dimensionless statistical variable X is proposed to quantify the effect of turbulent intermittency on the noise generation process. Our statistical analysis is validated by experimental results and provides new insights into aeroacoustics.</p
Coupling data assimilation and machine learning to improve land surface conditions and near-surface temperature and humidity forecasts
No full textEnhancing the representation of land surface conditions and improving the accuracy of near-surface weather forecasts remain critical challenges for numerical weather prediction (NWP). This study coupled a hybrid data assimilation-machine learning framework (DL) with the Weather Research and Forecasting (WRF) model to quantify the impacts of incorporating soil moisture (SM) and vegetation data on land surface initialization and near-surface weather forecast accuracy. This was achieved by integrating satellite-based leaf area index (LAI) and multi-source SM data into the WRF model in the Southern Great Plains (SGP) of the United States. The results indicate that optimizing LAI and SM significantly improves the simulation of water table depth, evapotranspiration (ET), air temperature and humidity in the WRF model. In addition to SM, LAI optimization provides additional benefits to the WRF model in dry years. A series of comparison experiments were conducted across both dry and wet years to evaluate the accuracy of air temperature and humidity forecasts. The optimized vegetation and SM conditions from the DL method were used as initial conditions for the early days of the forecast period. The results confirm that the DL method effectively refines the land surface initial conditions at the beginning of the forecast period. This effect improves the estimation of near-surface atmospheric conditions (e.g., air temperature and humidity) and alters precipitation patterns during the forecast period. In addition, the integration of LAI and SM is more effective in improving forecasts in wet/normal years than dry years. Analysis of the forecast results illustrates that the DL method can optimize initial conditions and improve near-surface weather forecasts over the next month.</p
AIvilization: Explore the Future of Human-AI Coexistence, Co-building a Digital Sandbox for Agent Civilization
No full textInvestigating the Factors Influencing Aristolochic Acid Toxicity and the Development of Remediation Method
No full textSmart Meter Condition Monitoring in a Metropolitan City with Data Analytics and Machine Learning
No full text
DTCO-based Hybrid Rail 8T Complementary FET SRAM Design Towards Advanced Node
No full textAs technology scaling advances toward the sub-1nm era, Complementary Field-Effect Transistor (CFET) architectures offer a promising path to sustain SRAM density and performance. The inherent 3D integration capability of CFET technology proves particularly advantageous for implementing 8T SRAM configurations at advanced nodes, overcoming the structural limitations of conventional 6T designs. However, the footprint of the current 8T CFET SRAM is still much larger than 6T SRAM. This work proposes a CFET-compatible hybrid-rail 8T SRAM design that addresses interconnect and layout challenges, achieving a 27% reduction in cell area compared to conventional dual-port 8T CFET SRAM. Furthermore, it demonstrates reduced parasitic effects on the bitline (BL) and wordline (WL). Through a Design Technology Co-Optimization approach (DTCO), SPICE simulations reveal that the hybrid-rail 8T SRAM achieves significant improvements in terms of power and delay. The results show that hybrid-rail 8T SRAM offers superior area efficiency, energy savings, and performance, establishing it as a compelling solution for high-density, low-power memory in advanced-node applications.<br/