arXiv subjects

Li Fu

Publications and source records attributed to Li Fu.

30 records · Page 2Linked to original sources

arXiv2022

SCaLa: Supervised Contrastive Learning for End-to-End Speech Recognition

End-to-end Automatic Speech Recognition (ASR) models are usually trained to optimize the loss of the whole token sequence, while neglecting explicit phonemic-granularity supervision. This could result in recognition errors due to similar-phoneme confusion or phoneme reduction. To alleviate this problem, we propose a novel framework based on Supervised Contrastive Learning (SCaLa) to enhance phonemic representation learning for end-to-end ASR systems. Specifically, we extend the self-supervised Masked Contrastive Predictive Coding (MCPC) to a fully-supervised setting, where the supervision is applied in the following way. First, SCaLa masks variable-length encoder features according to phoneme boundaries given phoneme forced-alignment extracted from a pre-trained acoustic model; it then predicts the masked features via contrastive learning. The forced-alignment can provide phoneme labels to mitigate the noise introduced by positive-negative pairs in self-supervised MCPC. Experiments on reading and spontaneous speech datasets show that our proposed approach achieves 2.8 and 1.4 points Character Error Rate (CER) absolute reductions compared to the baseline, respectively.

eess.AS↗

arXiv2021

Self-repairing high entropy oxides

All biological organisms, from plants to living creatures, can heal minor wounds and damage. The realization of a similar self-healing capacity in inorganic materials has been a design target for many decades. This would represent a breakthrough in materials engineering, enabling many novel technological applications, since such materials would be able to resist damage caused by electromagnetic irradiation and/or mechanical impact. Here we demonstrate that a high-entropy oxide is intrinsically capable of undergoing an autonomous self-repairing process. Transmission electron microscopy revealed that the spinel structure of (AlCoCrCu0.5FeNi)3O4 can regrow and repair itself at the atomic level when damaged. Density functional theory calculations reveal that the extra enthalpy stored in the high entropy material during fabrication can be released to effectively heal macroscopic defects by regrowing into a partially ordered state. This extraordinary self-repairing phenomenon makes this new material highly desirable as a coating, enabling structures used in harsh environments to better withstand damage, such as cosmic irradiation in space, nuclear irradiation in nuclear power facilities, or tribological damage. Most importantly, our results set the general design principles for the synthesis of self-repairing materials.

cond-mat.mtrl-sci↗

arXiv2021

Incremental Learning for End-to-End Automatic Speech Recognition

In this paper, we propose an incremental learning method for end-to-end Automatic Speech Recognition (ASR) which enables an ASR system to perform well on new tasks while maintaining the performance on its originally learned ones. To mitigate catastrophic forgetting during incremental learning, we design a novel explainability-based knowledge distillation for ASR models, which is combined with a response-based knowledge distillation to maintain the original model's predictions and the "reason" for the predictions. Our method works without access to the training data of original tasks, which addresses the cases where the previous data is no longer available or joint training is costly. Results on a multi-stage sequential training task show that our method outperforms existing ones in mitigating forgetting. Furthermore, in two practical scenarios, compared to the target-reference joint training method, the performance drop of our method is 0.02% Character Error Rate (CER), which is 97% smaller than the drops of the baseline methods.

eess.AS↗

arXiv2020

Learning Enhanced Resolution-wise features for Human Pose Estimation

Recently, multi-resolution networks (such as Hourglass, CPN, HRNet, etc.) have achieved significant performance on pose estimation by combining feature maps of various resolutions. In this paper, we propose a Resolution-wise Attention Module (RAM) and Gradual Pyramid Refinement (GPR), to learn enhanced resolution-wise feature maps for precise pose estimation. Specifically, RAM learns a group of weights to represent the different importance of feature maps across resolutions, and the GPR gradually merges every two feature maps from low to high resolutions to regress final human keypoint heatmaps. With the enhanced resolution-wise features learnt by CNN, we obtain more accurate human keypoint locations. The efficacies of our proposed methods are demonstrated on MS-COCO dataset, achieving state-of-the-art performance with average precision of 77.7 on COCO val2017 set and 77.0 on test-dev2017 set without using extra human keypoint training dataset.

cs.CV↗

arXiv2020

Liquid-solid slip on charged walls: dramatic impact of charge distribution

Nanofluidic systems show great promises for applications in energy conversion, where their performance can be enhanced by nanoscale liquid-solid slip. However, efficiency is also controlled by surface charge, which is known to reduce slip. Combining molecular dynamics simulations and analytical developments, we show the dramatic impact of surface charge distribution on the slip-charge coupling. Homogeneously charged graphene exhibits a very favorable slip-charge relation (rationalized with a new theoretical model correcting some weaknesses of the existing ones), leading to giant electrokinetic energy conversion. In contrast, slip is strongly affected on heterogeneously charged surfaces, due to the viscous drag induced by counter-ions trapped on the surface. In that case slip should depend on the detailed physical chemistry of the interface controlling the fraction of bound ions. Our numerical results and theoretical models provide new fundamental insight on the molecular mechanisms of liquid-solid slip, and practical guidelines for searching new functional interfaces with optimal energy conversion properties, e.g. for blue energy or waste heat harvesting.

cond-mat.soft↗

arXiv2020

Research on Modeling Units of Transformer Transducer for Mandarin Speech Recognition

Modeling unit and model architecture are two key factors of Recurrent Neural Network Transducer (RNN-T) in end-to-end speech recognition. To improve the performance of RNN-T for Mandarin speech recognition task, a novel transformer transducer with the combination architecture of self-attention transformer and RNN is proposed. And then the choice of different modeling units for transformer transducer is explored. In addition, we present a new mix-bandwidth training method to obtain a general model that is able to accurately recognize Mandarin speech with different sampling rates simultaneously. All of our experiments are conducted on about 12,000 hours of Mandarin speech with sampling rate in 8kHz and 16kHz. Experimental results show that Mandarin transformer transducer using syllable with tone achieves the best performance. It yields an average of 14.4% and 44.1% relative Word Error Rate (WER) reduction when compared with the models using syllable initial/final with tone and Chinese character, respectively. Also, it outperforms the model based on syllable initial/final with tone with an average of 13.5% relative Character Error Rate (CER) reduction.

eess.AS↗

arXiv2019

Giant thermoelectric response of nanofluidic systems driven by water excess enthalpy

Nanofluidic systems could in principle be used to produce electricity from waste heat, but current theoretical descriptions predict a rather poor performance as compared to thermoelectric solid materials. Here we investigate the thermoelectric response of NaCl and NaI solutions confined between charged walls, using molecular dynamics simulations. We compute a giant thermoelectric response, two orders of magnitude larger than the predictions of standard models. We show that water excess enthalpy -- neglected in the standard picture -- plays a dominant role in combination with the electroosmotic mobility of the liquid-solid interface. Accordingly, the thermoelectric response can be boosted using surfaces with large hydrodynamic slip. Overall, the heat harvesting performance of the model systems considered here is comparable to that of the best thermoelectric materials, and the fundamental insight provided by molecular dynamics suggests guidelines to further optimize the performance, opening the way to recycle waste heat using nanofluidic devices.

cond-mat.stat-mech↗

arXiv2019

Space very long baseline interferometry in China

Space very long baseline interferometry (VLBI) has unique applications in high-resolution imaging of fine structure of astronomical objects and high-precision astrometry due to the key long space-Earth or space-space baselines beyond the Earth's diameter. China has been actively involved in the development of space VLBI in recent years. This review briefly summarizes China's research progress in space VLBI and the future development plan.

astro-ph.IM↗

arXiv2018

All-Silicon Topological Semimetals with Closed Nodal Line

Owing to the natural compatibility with current semiconductor industry, silicon allotropes with diverse structural and electronic properties provide promising platforms for the next-generation Si-based devices. After screening 230 all-silicon crystals in the zeolite frameworks by first-principles calculations, we disclose two structurally stable Si allotropes (AHT-Si24 and VFI-Si36) containing open channels as topological node-line semimetals with Dirac nodal points forming a nodal loop in the kz=0 plane of Brillouin zone. Interestingly, their nodal loops protected by inversion and time-reversal symmetries are robust against SU(2) symmetry breaking due to very weak spin-orbit coupling of Si. When the nodal lines are projected onto the (001) surface, flat surface bands can be observed because of the nontrivial topology of the bulk band structures. Our discoveries extend the topological physics to the three-dimensional Si materials, highlighting the possibility to realize low-cost, nontoxic and semiconductor-compatible Si-based electronics with topological quantum states.

cond-mat.mtrl-sci↗

arXiv2018

Measuring surface charge: why experimental characterization and molecular modeling should be coupled

Surface charge controls many static and dynamic properties of soft matter and micro/nanofluidic systems, but its unambiguous measurement forms a challenge. Standard characterization methods typically probe an effective surface charge, which provides limited insight into the distribution and dynamics of charge across the interface, and which cannot predict consistently all surface-charge-governed properties. New experimental approaches provide local information on both structure and transport, but models are typically required to interpret raw data. Conversely, molecular dynamics simulations have helped showing the limits of standard models and developing more accurate ones, but their reliability is limited by the empirical interaction potentials they are usually based on. This review highlights recent developments and limitations in both experimental and computational research focusing on the liquid-solid interface. Based on recent studies, we make the case that coupling of experiments and simulations is pivotal tomitigate methodological shortcomings and address open problems pertaining to charged interfaces.

cond-mat.soft↗

arXiv2018

An efficient quantum circuits optimizing scheme compared with QISKit

Recently, the development of quantum chips has made great progress-- the number of qubits is increasing and the fidelity is getting higher. However, qubits of these chips are not always fully connected, which sets additional barriers for implementing quantum algorithms and programming quantum programs. In this paper, we introduce a general circuit optimizing scheme, which can efficiently adjust and optimize quantum circuits according to arbitrary given qubits' layout by adding additional quantum gates, exchanging qubits and merging single-qubit gates. Compared with the optimizing algorithm of IBM's QISKit, the quantum gates consumed by our scheme is 74.7%, and the execution time is only 12.9% on average.

cs.ET↗

arXiv2017

What controls thermo-osmosis? Molecular simulations show the critical role of interfacial hydrodynamics

Thermo-osmotic and related thermo-phoretic phenomena can be found in many situations from biology to colloid science, but the underlying molecular mechanisms remain largely unexplored. Using molecular dynamics simulations, we measured the thermo-osmosis coefficient by both mechano-caloric and thermo-osmotic routes, for different solid-liquid interfacial energies. The simulations reveal in particular the crucial role of nanoscale interfacial hydrodynamics. For non-wetting surfaces , thermo-osmotic transport is largely amplified by hydrodynamic slip at the interface. For wetting surfaces, the position of the hydrodynamic shear plane plays a key role in determining the amplitude and sign of the thermo-osmosis coefficient. Finally, we measure a giant thermo-osmotic response of the water-graphene interface, which we relate to the very low interfacial friction displayed by this system. These results open new perspectives for the design of efficient functional interfaces for, e.g., waste heat harvesting.

cond-mat.soft↗