arXiv subjects

Yu Zhang

Publications and source records attributed to Yu Zhang.

At least 1,297 records · Page 72Linked to original sources

arXiv2021

Improving Streaming Automatic Speech Recognition With Non-Streaming Model Distillation On Unsupervised Data

Streaming end-to-end automatic speech recognition (ASR) models are widely used on smart speakers and on-device applications. Since these models are expected to transcribe speech with minimal latency, they are constrained to be causal with no future context, compared to their non-streaming counterparts. Consequently, streaming models usually perform worse than non-streaming models. We propose a novel and effective learning method by leveraging a non-streaming ASR model as a teacher to generate transcripts on an arbitrarily large data set, which is then used to distill knowledge into streaming ASR models. This way, we scale the training of streaming models to up to 3 million hours of YouTube audio. Experiments show that our approach can significantly reduce the word error rate (WER) of RNNT models not only on LibriSpeech but also on YouTube data in four languages. For example, in French, we are able to reduce the WER by 16.4% relatively to a baseline streaming model by leveraging a non-streaming teacher model trained on the same amount of labeled data as the baseline.

cs.SD↗

arXiv2021

Reinforcement Learning for Beam Pattern Design in Millimeter Wave and Massive MIMO Systems

Employing large antenna arrays is a key characteristic of millimeter wave (mmWave) and terahertz communication systems. However, due to the adoption of fully analog or hybrid analog/digital architectures, as well as non-ideal hardware or arbitrary/unknown array geometries, the accurate channel state information becomes hard to acquire. This impedes the design of beamforming/combining vectors that are crucial to fully exploit the potential of large-scale antenna arrays in providing sufficient receive signal power. In this paper, we develop a novel framework that leverages deep reinforcement learning (DRL) and a Wolpertinger-variant architecture and learns how to iteratively optimize the beam pattern (shape) for serving one or a small set of users relying only on the receive power measurements and without requiring any explicit channel knowledge. The proposed model accounts for key hardware constraints such as the phase-only, constant-modulus, and quantized-angle constraints. Further, the proposed framework can efficiently optimize the beam patterns for systems with non-ideal hardware and for arrays with unknown or arbitrary array geometries. Simulation results show that the developed solution is capable of finding near-optimal beam patterns based only on the receive power measurements.

eess.SP↗

arXiv2021

Echo State Speech Recognition

We propose automatic speech recognition (ASR) models inspired by echo state network (ESN), in which a subset of recurrent neural networks (RNN) layers in the models are randomly initialized and untrained. Our study focuses on RNN-T and Conformer models, and we show that model quality does not drop even when the decoder is fully randomized. Furthermore, such models can be trained more efficiently as the decoders do not require to be updated. By contrast, randomizing encoders hurts model quality, indicating that optimizing encoders and learn proper representations for acoustic inputs are more vital for speech recognition. Overall, we challenge the common practice of training ASR models for all components, and demonstrate that ESN-based models can perform equally well but enable more efficient training and storage than fully-trainable counterparts.

cs.CL↗

arXiv2021

Multi-Objective Meta Learning

Meta learning with multiple objectives can be formulated as a Multi-Objective Bi-Level optimization Problem (MOBLP) where the upper-level subproblem is to solve several possible conflicting targets for the meta learner. However, existing studies either apply an inefficient evolutionary algorithm or linearly combine multiple objectives as a single-objective problem with the need to tune combination weights. In this paper, we propose a unified gradient-based Multi-Objective Meta Learning (MOML) framework and devise the first gradient-based optimization algorithm to solve the MOBLP by alternatively solving the lower-level and upper-level subproblems via the gradient descent method and the gradient-based multi-objective optimization method, respectively. Theoretically, we prove the convergence properties of the proposed gradient-based optimization algorithm. Empirically, we show the effectiveness of the proposed MOML framework in several meta learning problems, including few-shot learning, neural architecture search, domain adaptation, and multi-task learning.

cs.LG↗

arXiv2021

A Better and Faster End-to-End Model for Streaming ASR

End-to-end (E2E) models have shown to outperform state-of-the-art conventional models for streaming speech recognition [1] across many dimensions, including quality (as measured by word error rate (WER)) and endpointer latency [2]. However, the model still tends to delay the predictions towards the end and thus has much higher partial latency compared to a conventional ASR model. To address this issue, we look at encouraging the E2E model to emit words early, through an algorithm called FastEmit [3]. Naturally, improving on latency results in a quality degradation. To address this, we explore replacing the LSTM layers in the encoder of our E2E model with Conformer layers [4], which has shown good improvements for ASR. Secondly, we also explore running a 2nd-pass beam search to improve quality. In order to ensure the 2nd-pass completes quickly, we explore non-causal Conformer layers that feed into the same 1st-pass RNN-T decoder, an algorithm called Cascaded Encoders [5]. Overall, we find that the Conformer RNN-T with Cascaded Encoders offers a better quality and latency tradeoff for streaming ASR.

eess.AS↗

arXiv2021

Photoproduction $γp \to K^+Λ(1520)$ in an effective Lagrangian approach

The data on differential cross sections and photon-beam asymmetries for the $γp \to K^+Λ(1520)$ reaction have been analyzed within a tree-level effective Lagrangian approach. In addition to the $t$-channel $K$ and $K^\ast$ exchanges, the $u$-channel $Λ$ exchange, the $s$-channel nucleon exchange, and the interaction current, a minimal number of nucleon resonances in the $s$ channel are introduced in constructing the reaction amplitudes to describe the data. The results show that the experimental data can be well reproduced by including either the $N(2060)5/2^-$ or the $N(2120)3/2^-$ resonance. In both cases, the contact term and the $K$ exchange are found to make significant contributions, while the contributions from the $K^\ast$ and $Λ$ exchanges are negligible in the former case and considerable in the latter case. Measurements of the data on target asymmetries are called on to further pin down the resonance contents and to clarify the roles of the $K^\ast$ and $Λ$ exchanges in this reaction.

hep-ph↗

arXiv2021

Probing the doubly-charged Higgs with Muonium to Antimuonium Conversion Experiment

The spontaneous muonium-to-antimuonium conversion is one of the interesting charged lepton flavor violation processes. MACE is the next generation experiment to probe such a phenomenon. In models with a triplet Higgs to generate neutrino masses, such as Type-II seesaw and its variant, this process can be induced by the doubly-charged Higgs contained in it. In this article, we study the prospect of MACE to probe these models via the muonium-to-antimuonium transitions. After considering the limits from $μ^+ \rightarrow e^+ γ$ and $μ^+ \rightarrow e^+ e^- e^+$, we find that MACE could probe a parameter space for the doubly-charged Higgs which is beyond the reach of LHC and other flavor experiments.

hep-ph↗

arXiv2021

Uniform quenching processes in a holographic s+p model with reentrance

We study the homogenous quenching processes in a holographic s+p model with reentrant phase transitions. We first realize the reentrant phase transition in the holographic model in probe limit and draw the phase diagram. Next, we compare the time evolution of the two condensates in two groups of numerical quenching experiments across the reentrant region, with different quenching speed as well as different width of the reentrant region, respectively. We also study the dynamical competition between the two orders in quenching processes from the normal phase to the superconductor phase.

hep-th↗

arXiv2021

Oceanic non-Kolmogorov optical turbulence and spherical wave propagation

Light propagation in turbulent media is conventionally studied with the help of the spatio-temporal power spectra of the refractive index fluctuations. In particular, for natural water turbulence several models for the spatial power spectra have been developed based on the classic, Kolmogorov postulates. However, as currently widely accepted, non-Kolmogorov turbulent regime is also common in the stratified flow fields, as suggested by recent developments in atmospheric optics. Until now all the models developed for the non-Kolmogorov optical turbulence were pertinent to atmospheric research and, hence, involved only one advected scalar, e.g., temperature. We generalize the oceanic spatial power spectrum, based on two advected scalars, temperature and salinity concentration, to the non-Kolmogorov turbulence regime, with the help of the so-called "Upper-Bound Limitation" and by adopting the concept of spectral correlation of two advected scalars. The proposed power spectrum can handle general non-Kolmogorov, anisotropic turbulence but reduces to Kolmogorov, isotropic case if the power law exponents of temperature and salinity are set to 11/3 and anisotropy coefficient is set to unity. To show the application of the new spectrum, we derive the expression for the second-order mutual coherence function of a spherical wave and examine its coherence radius (in both scalar and vector forms) to characterize the turbulent disturbance. Our numerical calculations show that the statistics of the spherical wave vary substantially with temperature and salinity non-Kolmogorov power law exponents and temperature-salinity spectral correlation coefficient. The introduced spectrum is envisioned to become of significance for theoretical analysis and experimental measurements of non-classic natural water double-diffusion turbulent regimes.

physics.ao-ph↗

arXiv2021

Learning from Home: A Mixed-Methods Analysis of Live Streaming Based Remote Education Experience in Chinese Colleges During the COVID-19 Pandemic

The COVID-19 global pandemic and resulted lockdown policies have forced education in nearly every country to switch from a traditional co-located paradigm to a pure online 'distance learning from home' paradigm. Lying in the center of this learning paradigm shift is the emergence and wide adoption of distance communication tools and live streaming platforms for education. Here, we present a mixed-methods study on live streaming based education experience during the COVID-19 pandemic. We focus our analysis on Chinese higher education, carried out semi-structured interviews on 30 students, and 7 instructors from diverse colleges and disciplines, meanwhile launched a large-scale survey covering 6291 students and 1160 instructors in one leading Chinese university. Our study not only reveals important design guidelines and insights to better support current remote learning experience during the pandemic, but also provides valuable implications towards constructing future collaborative education supporting systems and experience after pandemic.

cs.CY↗

arXiv2021

Model-based cellular kinetic analysis of SARS-CoV-2 infection: different immune response modes and treatment strategies

Increasing number in global COVID-19 cases demands for mathematical model to analyze the interaction between the virus dynamics and the response of innate and adaptive immunity. Here, based on the assumption of a weak and delayed response of the innate and adaptive immunity in SARS-CoV-2 infection, we constructed a mathematical model to describe the dynamic processes of immune system. Integrating theoretical results with clinical COVID-19 patients' data, we classified the COVID-19 development processes into three typical modes of immune responses, correlated with the clinical classification of mild & moderate, severe and critical patients. We found that the immune efficacy (the ability of host to clear virus and kill infected cells) and the lymphocyte supply (the abundance and pool of naïve T and B cell) play important roles in the dynamic process and determine the clinical outcome, especially for the severe and critical patients. Furthermore, we put forward possible treatment strategies for the three typical modes of immune response. We hope our results can help to understand the dynamical mechanism of the immune response against SARS-CoV-2 infection, and to be useful for the treatment strategies and vaccine design.

q-bio.PE↗

arXiv2021

Generative Adversarial U-Net for Domain-free Medical Image Augmentation

The shortage of annotated medical images is one of the biggest challenges in the field of medical image computing. Without a sufficient number of training samples, deep learning based models are very likely to suffer from over-fitting problem. The common solution is image manipulation such as image rotation, cropping, or resizing. Those methods can help relieve the over-fitting problem as more training samples are introduced. However, they do not really introduce new images with additional information and may lead to data leakage as the test set may contain similar samples which appear in the training set. To address this challenge, we propose to generate diverse images with generative adversarial network. In this paper, we develop a novel generative method named generative adversarial U-Net , which utilizes both generative adversarial network and U-Net. Different from existing approaches, our newly designed model is domain-free and generalizable to various medical images. Extensive experiments are conducted over eight diverse datasets including computed tomography (CT) scan, pathology, X-ray, etc. The visualization and quantitative results demonstrate the efficacy and good generalization of the proposed method on generating a wide array of high-quality medical images.

eess.IV↗

arXiv2021

Local Measurements of Shubnikov-de Haas Oscillations in Graphene Systems

Shubnikov-de Haas (SdH) oscillations, the most well-known magneto-oscillations caused by the quantization of electron energy levels in the presence of magnetic fields in two-dimensional (2D) electron systems, can be used to determine Fermi-surface properties and directly measure the Berry phase of the 2D systems. It is usually thought that transport measurements are required to measure the SdH oscillations. Contradicting this belief, we demonstrate that the SdH oscillations can be measured in graphene systems by carrying out scanning tunneling spectroscopy (STS) measurements. The energy-momentum dispersions and Berry phases of monolayer, Bernal-stacked bilayer, and ABC-stacked trilayer graphene are obtained according to the measured SdH oscillations in the STS spectra. It is possible to obtain the SdH oscillations when the size of the 2D systems is larger than the magnetic length and, importantly, no gate electrode is required in the STS measurement, therefore, the reported method in this work is applicable to a wide range of materials.

cond-mat.mes-hall↗

arXiv2020

Multiple Structural Priors Guided Self Attention Network for Language Understanding

Self attention networks (SANs) have been widely utilized in recent NLP studies. Unlike CNNs or RNNs, standard SANs are usually position-independent, and thus are incapable of capturing the structural priors between sequences of words. Existing studies commonly apply one single mask strategy on SANs for incorporating structural priors while failing at modeling more abundant structural information of texts. In this paper, we aim at introducing multiple types of structural priors into SAN models, proposing the Multiple Structural Priors Guided Self Attention Network (MS-SAN) that transforms different structural priors into different attention heads by using a novel multi-mask based multi-head attention mechanism. In particular, we integrate two categories of structural priors, including the sequential order and the relative position of words. For the purpose of capturing the latent hierarchical structure of the texts, we extract these information not only from the word contexts but also from the dependency syntax trees. Experimental results on two tasks show that MS-SAN achieves significant improvements against other strong baselines.

cs.CL↗

arXiv2020

RNN-T Models Fail to Generalize to Out-of-Domain Audio: Causes and Solutions

In recent years, all-neural end-to-end approaches have obtained state-of-the-art results on several challenging automatic speech recognition (ASR) tasks. However, most existing works focus on building ASR models where train and test data are drawn from the same domain. This results in poor generalization characteristics on mismatched-domains: e.g., end-to-end models trained on short segments perform poorly when evaluated on longer utterances. In this work, we analyze the generalization properties of streaming and non-streaming recurrent neural network transducer (RNN-T) based end-to-end models in order to identify model components that negatively affect generalization performance. We propose two solutions: combining multiple regularization techniques during training, and using dynamic overlapping inference. On a long-form YouTube test set, when the nonstreaming RNN-T model is trained with shorter segments of data, the proposed combination improves word error rate (WER) from 22.3% to 14.8%; when the streaming RNN-T model trained on short Search queries, the proposed techniques improve WER on the YouTube set from 67.0% to 25.3%. Finally, when trained on Librispeech, we find that dynamic overlapping inference improves WER on YouTube from 99.8% to 33.0%.

eess.AS↗

arXiv2020

Hierarchical Metadata-Aware Document Categorization under Weak Supervision

Categorizing documents into a given label hierarchy is intuitively appealing due to the ubiquity of hierarchical topic structures in massive text corpora. Although related studies have achieved satisfying performance in fully supervised hierarchical document classification, they usually require massive human-annotated training data and only utilize text information. However, in many domains, (1) annotations are quite expensive where very few training samples can be acquired; (2) documents are accompanied by metadata information. Hence, this paper studies how to integrate the label hierarchy, metadata, and text signals for document categorization under weak supervision. We develop HiMeCat, an embedding-based generative framework for our task. Specifically, we propose a novel joint representation learning module that allows simultaneous modeling of category dependencies, metadata information and textual semantics, and we introduce a data augmentation module that hierarchically synthesizes training documents to complement the original, small-scale training set. Our experiments demonstrate a consistent improvement of HiMeCat over competitive baselines and validate the contribution of our representation learning and data augmentation modules.

cs.CL↗

arXiv2020

Probing the $L_μ-L_τ$ gauge boson at electron colliders

We investigate the minimal $U(1)_{L_μ-L_τ}$ model with extra heavy vector-like leptons or charged scalars. By studying the kinetic mixing between $U(1)_{L_μ-L_τ}$ gauge boson $Z^\prime$ and standard model photon, which is absent at tree level and will arise at one loop level due to $μ$, $τ$ and new heavy charged leptons or scalars, the interesting behavior is shown. It can provide possibility for visible signatures of new heavy particles. We propose to search for $Z^\prime$ at electron collider experiments, such as Belle II, BESIII and future Super Tau Charm Factory (STCF), using the monophoton final state. The parameter space of $Z^\prime$ is probed, and scanned by its gauge coupling constant $g_{Z^\prime}$ and mass $m_{Z^\prime}$. We find that electron colliders have sensitivity to the previously unexplored parameter space for $Z^\prime$ with MeV-GeV mass. Future STCF experiments with $\sqrt s=2-7$ GeV can exclude the anomalous muon magnetic moment favored area when $m_{Z^\prime}<5$ GeV with the luminosity of 30 ab$^{-1}$. For $m_{Z^\prime} < 2m_μ$, $g_{Z^\prime}$ can be down to $4.2\times 10^{-5}$ at 2 GeV STCF.

hep-ph↗

arXiv2020

Fortuitous partners of antiferromagnetic and Mott states in spin-orbit-coupled Sr2IrO4: A study of Sr2Ir1-xMxO4 (M=Fe or Co)

Sr2IrO4 is an archetypal spin-orbit-coupled Mott insulator with an antiferromagnetic state below 240 K. Here we report results of our study on single crystals of Sr2Ir1-xFexO4 (0<x<0.32) and Sr2Ir1-xCoxO4 (0<x<0.22). Fe doping retains the antiferromagnetic state but simultaneously precipitates an emergent metallic state whereas Co doping causes a rapid collapse of both the antiferromagnetic and Mott states, giving rise to a confined metallic state featuring a pronounced linearity of the basal-plane resistivity up to 700 K. The results indicate tetravalent Fe4+(3d4) ions in the intermediate spin state with S=1 and Co4+(3d5) ions in the high spin state with S=5/2 substituting for Ir4+(5d5) ions in Sr2IrO4, respectively. The effective magnetic moment closely tracks the Néel temperature as doping increases, suggesting that the spin state of the dopant predominately determines the magnetic properties in doped Sr2IrO4. Furthermore, all relevant properties including charge-carrier density (e.g., 1028/m3), Sommerfeld coefficient (e.g., 19 mJ/mole K2) and Wilson ratio (e.g., 2.6), consistently demonstrates a metallic state that is both robust and highly correlated in the two systems, arising from the percolation of bound states and the weakening of structural distortions. This study strongly suggests that the antiferromagnetic and Mott states merely coexist in a fortuitous manner in Sr2IrO4.

cond-mat.str-el↗