Search for: All records

Large language models (LLMs) have achieved remarkable performance on various natural language tasks. However, they are trained on static corpora and their knowledge can become outdated quickly in the fast-changing world. This motivates the development of knowledge editing (KE) to update specific knowledge in LLMs without changing unrelated others or compromising their pre-trained capabilities. Previous efforts sought to update a small amount of parameters of a LLM and proved effective for making selective updates. Nonetheless, the edited LLM often exhibits degraded ability to reason about the new knowledge. In this work, we identify a key issue: \textit{heterogeneous token overfitting} (HTO), where the LLM overfits different tokens in the provided knowledge at varying rates. To tackle this, we propose {\NAME}, a token-level smoothing method that mitigates HTO by adaptively refining the target distribution. Theoretically, {\NAME} offers better parameter updates with negligible computation overhead. It also induces an implicit DPO but does not require preference data pairs. Extensive experiments across four editing methods, two LLMs, and diverse scenarios demonstrate the effectiveness and versatility of our method. 

This work proposes a class of differentially private mechanisms for linear queries, in par- ticular range queries, that leverages corre- lated input perturbation to simultaneously achieve unbiasedness, consistency, statisti- cal transparency, and control over utility re- quirements in terms of accuracy targets ex- pressed either in certain query margins or as implied by the hierarchical database struc- ture. The proposed Cascade Sampling al- gorithm instantiates the mechanism exactly and efficiently. Our theoretical and empir- ical analysis demonstrates that we achieve near-optimal utility, effectively compete with other methods, and retain all the favorable statistical properties discussed earlier. 

This work investigates the steady-state nonlinear dynamics of a large-deformation flexible beam model under oscillatory flow. A flexible beam dynamics model combined with hydrodynamic loading is employed using large deformation beam theory. The equations of motion discretised using the high-order finite element method (FEM) are solved in the time domain using the efficient Galerkin averaging-incremental harmonic balance (EGA-IHB) method. The arc-length continuation method and Hsu’s method trace stable and unstable solutions. The numerical results are in accordance with the physical experimental results and reveal multiple resonance phenomena. Low-order resonances exhibit hardening due to geometric nonlinearity, while higher-order resonances transition from softening to hardening influenced by inertia and geometric nonlinearity. A strong coupling between tensile and bending deformation is observed. The axial deformation is dominated by inertia, while bending resonance is influenced by an interplay between inertia, structure stiffness, and fluid drag. Finally, the effects of two dimensionless parameters, Keulegan and Carpenter number (KC) and Cauchy number (Ca), on the response of the flexible beam are discussed. 

Metric magnitude is a measure of the “size” of point clouds with many desirable geometric properties. It has been adapted to various mathematical contexts and recent work suggests that it can enhance machine learning and optimization algorithms. But its usability is limited due to the computational cost when the dataset is large or when the computation must be carried out repeatedly (e.g. in model training). In this paper, we study the magnitude computation problem, and show efficient ways of approximating it. We show that it can be cast as a convex optimization problem, but not as a submodular optimization. The paper describes two new algorithms – an iterative approximation algorithm that converges fast and is accurate, and a subset selection method that makes the computation even faster. It has been previously proposed that magnitude of model sequences generated during stochastic gradient descent is correlated to generalization gap. Extension of this result using our more scalable algorithms shows that longer sequences in fact bear higher correlations. We also describe new applications of magnitude in machine learning – as an effective regularizer for neural network training, and as a novel clustering criterion. 

Abstract Climate change‐induced heat stress has significant effects on human health, and is influenced by a wide variety of factors. Most assessments of future heat‐related risks however are based on coarse resolution projections of heat hazards and overlook the contribution of relevant factors other than climate change to the negative impacts on health. Research highlights sociodemographic disparities related to heat stress vulnerability, especially among older adults, women and individuals with low socioeconomic status, leading to higher morbidity and mortality rates. There is thus an urgent need for detailed, local information on demographic characteristics underlying vulnerability with refined spatial resolution. This study aims to address the research gaps by presenting a new population projection exercise at high‐resolution based on the Bayesian modeling framework for the case study of Madrid, using demographic data under the scenarios compatible with the Shared Socioeconomic Pathways. We examine the spatial and temporal distribution of population subgroups at the intra‐urban level within Madrid. Our findings reveal a concentration of vulnerable populations, as measured by their age, sex and educational attainment level in some of the city's most disadvantaged neighborhoods. These vulnerable clusters are projected to widen in the future unless a sustainable trajectory is realized, driving vulnerability dynamics toward a more uniform and resilient change. These results can guide local adaptation efforts and support climate justice initiatives to protect vulnerable communities in urban environments. 

THIS PAPER IS UNDER REVISION 

Designing a senior-level course that involves problem-based learning, including project completion task, is laborious and challenging. A well-designed project motivates the students to be self-learners and prepares them for future industrial or academic endeavors. The COVID-19 pandemic brought many challenges when instructions were forced to move either online or to a remote teaching/learning environment. Due to this rapid transition, delivery modes in teaching and learning modalities faced disruption making course design more difficult. The senior level Flight Controls course AME - 4513 is designed with Unmanned Aerial Systems (UAS) related projects for the students to have a better understanding of UAS usage on various applications in support of Advanced Technological Education (ATE) program. The purpose of this paper is to present the UAS lab modules in a junior level robotics lab, AME - 4802, which preceded the Flight Controls course in the school of Aerospace and Mechanical Engineering at the University of Oklahoma. Successfully completing the course project requires independent research and involves numerical simulations of UAS. The Robotics Lab course focuses on hands-on projects of robotic systems with an emphasis on semi-autonomous mobile robots, including an UAS introduction module. - The UAS module in the Robotics Lab class is introduced in Spring 2020. Therefore, most of the students enrolled in the Spring 2020 Robotics Lab course have introductory knowledge about the UAS system when taking the Fall 2020 Flight Control course. In addition, Spring 2020 Robotics Lab was affected due to COVID-19. - The UAS module was not introduced in 2019 Spring Robotics lab. Thus, the students enrolled in Fall 2019 Flight Controls course did not have prior knowledge on the UAS system. - We thus present the implementation of UAS module in a junior level robotics lab which preceded the senior level Flight Controls course in following Fall semester, when the same instructor taught the course.