arXiv subjects

Yichen Jia

Publications and source records attributed to Yichen Jia.

12 recordsLinked to original sources

arXiv2026

Leveraging Beam Search Information for Confidence Estimation in E2E ASR

To estimate confidence for end-to-end Automatic Speech Recognition (ASR) systems, recent research has proposed Confidence Estimation Modules that incorporate features from the backbone ASR model. Most existing approaches, however, are architecture-dependent. In this paper, we propose the Score-Rank Confidence Estimation Module (SR-CEM), a lightweight module that leverages beam search information to generate token- and word-level confidence scores. Specifically, SR-CEM constructs features by combining the scores and ranks of tokens within a hypothesis. Experiments show that SR-CEM achieves effective calibration on both in-domain and out-of-domain English data. On the in-domain testset, it attains a Maximum Calibration Error of 4.50% and an Expected Calibration Error of 0.30% at the token level, significantly outperforming softmax confidence (20.04% and 1.75%, respectively). At the word level, SR-CEM achieves 8.17% and 0.35%, compared to 17.91% and 1.67% from softmax confidence. Furthermore, we demonstrate its robustness across hybrid and transducer ASR architectures with different decoding strategies, as well as on Dutch, noisy and conversational speech conditions. Our main finding is that SR-CEM is particularly effective in reducing Maximum Calibration Error, which is critical for reliable downstream use of ASR outputs, while maintaining architecture independence and generality across diverse evaluation conditions.

eess.AS↗

arXiv2023

Fairy: Fast Parallelized Instruction-Guided Video-to-Video Synthesis

In this paper, we introduce Fairy, a minimalist yet robust adaptation of image-editing diffusion models, enhancing them for video editing applications. Our approach centers on the concept of anchor-based cross-frame attention, a mechanism that implicitly propagates diffusion features across frames, ensuring superior temporal coherence and high-fidelity synthesis. Fairy not only addresses limitations of previous models, including memory and processing speed. It also improves temporal consistency through a unique data augmentation strategy. This strategy renders the model equivariant to affine transformations in both source and target images. Remarkably efficient, Fairy generates 120-frame 512x384 videos (4-second duration at 30 FPS) in just 14 seconds, outpacing prior works by at least 44x. A comprehensive user study, involving 1000 generated samples, confirms that our approach delivers superior quality, decisively outperforming established methods.

cs.CV↗

arXiv2023

Surpassing millisecond coherence times in on-chip superconducting quantum memories by optimizing materials, processes, and circuit design

The performance of superconducting quantum circuits for quantum computing has advanced tremendously in recent decades; however, a comprehensive understanding of relaxation mechanisms does not yet exist. In this work, we utilize a multimode approach to characterizing energy losses in superconducting quantum circuits, with the goals of predicting device performance and improving coherence through materials, process, and circuit design optimization. Using this approach, we measure significant reductions in surface and bulk dielectric losses by employing a tantalum-based materials platform and annealed sapphire substrates. With this knowledge we predict and experimentally verify the relaxation times of aluminum- and tantalum-based transmon qubits. We additionally optimize device geometry to maximize coherence within a coaxial tunnel architecture, and realize on-chip quantum memories with single-photon Ramsey times of 2.0$-$2.7 ms, limited by their energy relaxation times of 1.0$-$1.4 ms. To our knowledge this is the highest coherence achieved in an on-chip quantum memory, and demonstrates an advancement towards a more modular and compact coaxial circuit architecture for bosonic qubits with reproducibly high coherence.

quant-ph↗

arXiv2023

Disentangling Losses in Tantalum Superconducting Circuits

Superconducting qubits are a leading system for realizing large scale quantum processors, but overall gate fidelities suffer from coherence times limited by microwave dielectric loss. Recently discovered tantalum-based qubits exhibit record lifetimes exceeding 0.3 ms. Here we perform systematic, detailed measurements of superconducting tantalum resonators in order to disentangle sources of loss that limit state-of-the-art tantalum devices. By studying the dependence of loss on temperature, microwave photon number, and device geometry, we quantify materials-related losses and observe that the losses are dominated by several types of saturable two level systems (TLSs), with evidence that both surface and bulk related TLSs contribute to loss. Moreover, we show that surface TLSs can be altered with chemical processing. With four different surface conditions, we quantitatively extract the linear absorption associated with different surface TLS sources. Finally, we quantify the impact of the chemical processing at single photon powers, the relevant conditions for qubit device performance. In this regime we measure resonators with internal quality factors ranging from 5 to 15 x 10^6, comparable to the best qubits reported. In these devices the surface and bulk TLS contributions to loss are comparable, showing that systematic improvements in materials on both fronts will be necessary to improve qubit coherence further.

quant-ph↗

arXiv2023

Chemical profiles of the oxides on tantalum in state of the art superconducting circuits

Over the past decades, superconducting qubits have emerged as one of the leading hardware platforms for realizing a quantum processor. Consequently, researchers have made significant effort to understand the loss channels that limit the coherence times of superconducting qubits. A major source of loss has been attributed to two level systems that are present at the material interfaces. We recently showed that replacing the metal in the capacitor of a transmon with tantalum yields record relaxation and coherence times for superconducting qubits, motivating a detailed study of the tantalum surface. In this work, we study the chemical profile of the surface of tantalum films grown on c-plane sapphire using variable energy X-ray photoelectron spectroscopy (VEXPS). We identify the different oxidation states of tantalum that are present in the native oxide resulting from exposure to air, and we measure their distribution through the depth of the film. Furthermore, we show how the volume and depth distribution of these tantalum oxidation states can be altered by various chemical treatments. By correlating these measurements with detailed measurements of quantum devices, we can improve our understanding of the microscopic device losses.

cond-mat.mtrl-sci↗

arXiv2022

Censored Quantile Regression Neural Networks for Distribution-Free Survival Analysis

This paper considers doing quantile regression on censored data using neural networks (NNs). This adds to the survival analysis toolkit by allowing direct prediction of the target variable, along with a distribution-free characterisation of uncertainty, using a flexible function approximator. We begin by showing how an algorithm popular in linear models can be applied to NNs. However, the resulting procedure is inefficient, requiring sequential optimisation of an individual NN at each desired quantile. Our major contribution is a novel algorithm that simultaneously optimises a grid of quantiles output by a single NN. To offer theoretical insight into our algorithm, we show firstly that it can be interpreted as a form of expectation-maximisation, and secondly that it exhibits a desirable `self-correcting' property. Experimentally, the algorithm produces quantiles that are better calibrated than existing methods on 10 out of 12 real datasets.

stat.ML↗

arXiv2022

CausalDeepCENT: Deep Learning for Causal Prediction of Individual Event Times

Deep learning (DL) has recently drawn much attention in image analysis, natural language process, and high-dimensional medical data analysis. Under the causal direct acyclic graph (DAG) interpretation, the input variables without incoming edges from parent nodes in the DL architecture maybe assumed to be randomized and independent of each other. As in a regression setting, including the input variables in the DL algorithm would reduce the bias from the potential confounders. However, failing to include a potential latent causal structure among the input variables affecting both treatment assignment and the output variable could be additional significant source of bias. The primary goal of this study is to develop new DL algorithms to estimate causal individual event times for time-to-event data, equivalently to estimate the causal time-to-event distribution with or without right censoring, accounting for the potential latent structure among the input variables. Once the causal individual event times are estimated, it would be straightforward to estimate the causal average treatment effects as the differences in the averages of the estimated causal individual event times. A connection is made between the proposed method and the targeted maximum likelihood estimation (TMLE). Simulation studies are performed to assess improvement in prediction abilities of the proposed methods by using the mean square error (MSE)-based method and rank-based $C$-Index metric. The simulation results indicate that improvement on the prediction accuracy could be substantial particularly when there is a collider among the input variables. The proposed method is illustrated with a publicly available and influential breast cancer data set. The proposed method has been implemented by using PyTorch and uploaded at https://github.com/yicjia/CausalDeepCENT.

stat.AP↗

arXiv2022

DeepCENT: Prediction of Censored Event Time via Deep Learning

With the rapid advances of deep learning, many computational methods have been developed to analyze nonlinear and complex right censored data via deep learning approaches. However, the majority of the methods focus on predicting survival function or hazard function rather than predicting a single valued time to an event. In this paper, we propose a novel method, DeepCENT, to directly predict the individual time to an event. It utilizes the deep learning framework with an innovative loss function that combines the mean square error and the concordance index. Most importantly, DeepCENT can handle competing risks, where one type of event precludes the other types of events from being observed. The validity and advantage of DeepCENT were evaluated using simulation studies and illustrated with three publicly available cancer data sets.

cs.LG↗

arXiv2020

Deep Learning for Quantile Regression under Right Censoring: DeepQuantreg

The computational prediction algorithm of neural network, or deep learning, has drawn much attention recently in statistics as well as in image recognition and natural language processing. Particularly in statistical application for censored survival data, the loss function used for optimization has been mainly based on the partial likelihood from Cox's model and its variations to utilize existing neural network library such as Keras, which was built upon the open source library of TensorFlow. This paper presents a novel application of the neural network to the quantile regression for survival data with right censoring, which is adjusted by the inverse of the estimated censoring distribution in the check function. The main purpose of this work is to show that the deep learning method could be flexible enough to predict nonlinear patterns more accurately compared to existing quantile regression methods such as traditional linear quantile regression and nonparametric quantile regression with total variation regularization, emphasizing practicality of the method for censored survival data. Simulation studies were performed to generate nonlinear censored survival data and compare the deep learning method with existing quantile regression methods in terms of prediction accuracy. The proposed method is illustrated with two publicly available breast cancer data sets with gene signatures. The method has been built into a package and is freely available at \url{https://github.com/yicjia/DeepQuantreg}.

stat.ML↗

arXiv2015

Black Arsenic-Phosphorus: Layered Anisotropic Infrared Semiconductors with Highly Tunable Compositions and Properties

Two-dimensional (2D) layered materials with diverse properties have attracted significant interest in the past decade. The layered materials discovered so far have covered a wide, yet discontinuous electromagnetic spectral range from semimetallic graphene, insulating boron nitride, to semiconductors with bandgaps from middle infrared to visible light. Here, we introduce new layered semiconductors, black arsenic-phosphorus (b-AsP), with highly tunable chemical compositions and electronic and optical properties. Transport and infrared absorption studies demonstrate the semiconducting nature of b-AsP with tunable bandgaps, ranging from 0.3 to 0.15 eV. These bandgaps fall into long-wavelength infrared (LWIR) regime and cannot be readily reached by other layered materials. Moreover, polarization-resolved infrared absorption and Raman studies reveal in-plane anisotropic properties of b-AsP. This family of layered b-AsP materials extend the electromagnetic spectra covered by 2D layered materials to the LWIR regime, and may find unique applications for future all 2D layered material based devices.

cond-mat.mtrl-sci↗

arXiv2014

Highly Anisotropic and Robust Excitons in Monolayer Black Phosphorus

Semi-metallic graphene and semiconducting monolayer transition metal dichalcogenides (TMDCs) are the two-dimensional (2D) materials most intensively studied in recent years. Recently, black phosphorus emerged as a promising new 2D material due to its widely tunable and direct bandgap, high carrier mobility and remarkable in-plane anisotropic electrical, optical and phonon properties. However, current progress is primarily limited to its thin-film form, and its unique properties at the truly 2D quantum confinement have yet to be demonstrated. Here, we reveal highly anisotropic and tightly bound excitons in monolayer black phosphorus using polarization-resolved photoluminescence measurements at room temperature. We show that regardless of the excitation laser polarization, the emitted light from the monolayer is linearly polarized along the light effective mass direction and centers around 1.3 eV, a clear signature of emission from highly anisotropic bright excitons. In addition, photoluminescence excitation spectroscopy suggests a quasiparticle bandgap of 2.2 eV, from which we estimate an exciton binding energy of around 0.9 eV, consistent with theoretical results based on first-principles. The experimental observation of highly anisotropic, bright excitons with exceedingly large binding energy not only opens avenues for the future explorations of many-electron effects in this unusual 2D material, but also suggests a promising future in optoelectronic devices such as on-chip infrared light sources.

cond-mat.mes-hall↗

arXiv2014

Rediscovering Black Phosphorus: A Unique Anisotropic 2D Material for Optoelectronics and Electronics

Anisotropy refers to the property of a material exhibiting directionally dependent features. In this paper, we introduce black phosphorous (BP), the most stable allotrope of phosphorus in layered orthorhombic structure with a bandgap of 0.3 eV in bulk, as a unique 2D material in which electrons, phonons and their interactions with photons behave in a highly anisotropic manner within the plane of the layers. The unique anisotropic nature of BP thin films is revealed using angle-resolved Raman and infrared spectroscopies, together with angle-resolved transport study. For 15 nm thick BP, we measure Hall mobility of 1000 and 600 cm^2/Vs for holes along the light (x) and heavy (y) effective mass directions, respectively, at 120 K. These BP thin films also exhibit large and anisotropic in-plane optical conductivity from 2 to 5 micrometer wavelength. Field effect transistors using 4 to 30 layers of BP (2 to 15 nm) as channel material exhibit an on-off current ratio exceeding 10^5, a field-effect mobility of 205 cm^2/Vs, and good saturation properties all at room temperature, suggesting its promising future in high performance thin film electronics. By introducing narrow bandgap BP into the 2D material family, we fill the space between semi-metallic graphene and large bandgap TMDCs, where great potentials for infrared optoelectronics lie. Most importantly, the unique anisotropic nature of this intriguing material creates unprecedented possibilities for the realization of conceptually new optoelectronic and electronic devices in which angle-dependent physical properties are highly desirable.

cond-mat.mes-hall↗