arXiv subjects

Lingling Wu

Publications and source records attributed to Lingling Wu.

7 recordsLinked to original sources

arXiv2023

FedGSM: Efficient Federated Learning for LEO Constellations with Gradient Staleness Mitigation

Recent advancements in space technology have equipped low Earth Orbit (LEO) satellites with the capability to perform complex functions and run AI applications. Federated Learning (FL) on LEO satellites enables collaborative training of a global ML model without the need for sharing large datasets. However, intermittent connectivity between satellites and ground stations can lead to stale gradients and unstable learning, thereby limiting learning performance. In this paper, we propose FedGSM, a novel asynchronous FL algorithm that introduces a compensation mechanism to mitigate gradient staleness. FedGSM leverages the deterministic and time-varying topology of the orbits to offset the negative effects of staleness. Our simulation results demonstrate that FedGSM outperforms state-of-the-art algorithms for both IID and non-IID datasets, underscoring its effectiveness and advantages. We also investigate the effect of system parameters.

eess.SP↗

arXiv2022

Towards Efficient NLP: A Standard Evaluation and A Strong Baseline

Supersized pre-trained language models have pushed the accuracy of various natural language processing (NLP) tasks to a new state-of-the-art (SOTA). Rather than pursuing the reachless SOTA accuracy, more and more researchers start paying attention on model efficiency and usability. Different from accuracy, the metric for efficiency varies across different studies, making them hard to be fairly compared. To that end, this work presents ELUE (Efficient Language Understanding Evaluation), a standard evaluation, and a public leaderboard for efficient NLP models. ELUE is dedicated to depict the Pareto Frontier for various language understanding tasks, such that it can tell whether and how much a method achieves Pareto improvement. Along with the benchmark, we also release a strong baseline, ElasticBERT, which allows BERT to exit at any layer in both static and dynamic ways. We demonstrate the ElasticBERT, despite its simplicity, outperforms or performs on par with SOTA compressed and early exiting models. With ElasticBERT, the proposed ELUE has a strong Pareto Frontier and makes a better evaluation for efficient NLP models.

cs.CL↗

arXiv2022

A Simple Hash-Based Early Exiting Approach For Language Understanding and Generation

Early exiting allows instances to exit at different layers according to the estimation of difficulty. Previous works usually adopt heuristic metrics such as the entropy of internal outputs to measure instance difficulty, which suffers from generalization and threshold-tuning. In contrast, learning to exit, or learning to predict instance difficulty is a more appealing way. Though some effort has been devoted to employing such "learn-to-exit" modules, it is still unknown whether and how well the instance difficulty can be learned. As a response, we first conduct experiments on the learnability of instance difficulty, which demonstrates that modern neural models perform poorly on predicting instance difficulty. Based on this observation, we propose a simple-yet-effective Hash-based Early Exiting approach (HashEE) that replaces the learn-to-exit modules with hash functions to assign each token to a fixed exiting layer. Different from previous methods, HashEE requires no internal classifiers nor extra parameters, and therefore is more efficient. Experimental results on classification, regression, and generation tasks demonstrate that HashEE can achieve higher performance with fewer FLOPs and inference time compared with previous state-of-the-art early exiting methods.

cs.CL↗

arXiv2015

Isotropic Negative Thermal Expansion Metamaterials

Negative thermal expansion materials are important and desirable in science and engineering applications. However, natural materials with isotropic negative thermal expansion are rare and usually unsatisfied in performance. Here, we propose a novel method to achieve negative thermal expansion via a metamaterial approach. The metamaterial is constructed with unit cells that combine bi-material strips and anti-chiral structures. Both experimental and simulation results display isotropic negative thermal expansion properties. The coefficient of negative thermal expansion of the metamaterials is demonstrated to be dependent on the difference between the thermal expansion coefficients of two component materials in the bi-material strips, as well as on the circular node radius and the ligament length in the anti-chiral structures. The measured value of the linear negative thermal expansion coefficient reaches -68.1X10-6 1/K in an operating temperature range from 303.15 K to 773.15 K, which is among the largest achieved in experiments to date. Our findings provide a novel and practical approach to obtaining materials with tunable isotropic negative thermal expansion on any scale.

cond-mat.mtrl-sci↗

arXiv2015

Independent Manipulation of Electric and Thermal Fields with Bilayer Structure

Recently, increasing attention has been focused on the employment of transformation and metamaterial for manipulation of various physical fields, which requires complicated configuration and usually limits in single field. Here, for the first time, we propose and experimentally demonstrated bilayer structure to achieve simultaneously independent manipulation of multi-physics field (dc electric fields and thermal) by directly solving the dc electric/ thermal field equations. This structure is composed of two layers: the outer layer is made of isotropic and homogeneous material, while the inner layer is fan-shape layer. Since it is not based on TO, it can be readily experimentally fabricated with naturally occurring materials. Experimentally, we has designed, fabricated and characterized two structures simultaneously behaving as dc electric cloak/ thermal concentrator and dc electric concentrator/ thermal cloak, respectively. The simulation results agree well with the experiment ones, thus confirming the feasibility of our methodology. This job provides a novel avenue to manipulation of multiphysics fields, thus might find potential applications in various areas.

cond-mat.mtrl-sci↗

arXiv2000

Chiral Symmetry Breaking for Domain Wall Fermions in Quenched Lattice QCD

The domain wall fermion formulation exhibits full chiral symmetry for finite lattice spacing except for the effects of mixing between the domain walls. Close to the continuum limit these symmetry breaking effects should be described by a single residual mass. We determine this mass from the conservation law obeyed by the conserved axial current in quenched simulations with beta=5.7 and 6.0 and domain wall separations varying between 12 and 48 on 8^3x32 and 16^3x32 lattices. Using the resulting values for the residual mass we perform two complete and independent calculations of the pion decay constant. Good agreement is found between these two methods and with experiment.

hep-lat↗

arXiv1999

QCD hadron spectrum with domain wall fermions

We present the QCD hadron spectrum for the cases of both quenched and two-flavor dynamical domain wall fermions. We compare the results obtained using the Wilson gauge action and a renormalization group improved gauge action. Finite volume effects and the dependence on the finite extent of the fifth dimension are discussed.

hep-lat↗