Trust snapshot

Quick read

Trust 21 - EmergingVerification L1Unclaimed author
24works
0followers
22topics
4close collaborators

Actions

Decide how to stay connected

Follow researcher0

Identity and collaboration

How to connect with this researcher

Claiming links this public author record to a researcher profile and unlocks direct collaboration workflows.

Log in to claim

Direct collaboration

Open a focused conversation when the fit is right

Claim this author entity first to unlock direct invitations.

Research graph

See the researcher in context

Open full explorer

Inspect adjacent work, topics, institutions and collaborators without jumping out to a separate graph page.

Building this graph slice

BZPEER is loading the nearby papers, people, topics and institutions for this page.

Published work

24 published item(s)

preprint2026arXiv

TTE-Flash: Accelerating Reasoning-based Multimodal Representations via Think-Then-Embed Tokens

Recent research has demonstrated that Universal Multimodal Embedding (UME) benefits significantly from Chain-of-Thought (CoT) reasoning. In this paradigm, a generative model produces explicit reasoning traces for a multimodal query, with the final representation extracted from an <eos> embedding token attending to both the query and the reasoning. Despite its effectiveness, the computational overhead of generating explicit CoT traces is often prohibitive. In this work, we propose replacing explicit CoT with latent think tokens, which are interpreted as latent variables that can produce explicit CoT traces as observed variables. By optimizing think tokens using CoT generation loss and subsequent embedding tokens using contrastive loss, we produce high-performance, reasoning-aware representations at a constant inference cost. Our study investigates two key architectural designs: 1) how think and embeddings tokens should be extracted from the same LLM backbone. 2) how the tokens should be trained as two dependent tasks. We introduce TTE-Flash-2B, a reasoning-aware multimodal representation model that outperforms its explicit-CoT counterpart on the MMEB-v2 benchmark, while producing latent think tokens that are interpretable both textually and visually. Furthermore, zero-shot evaluation across 15 video datasets reveals scaling behavior as the number of think tokens increases, and motivating a pilot study of adaptive think budget allocation based on task requirements.

preprint2023arXiv

Improving Dialogue Breakdown Detection with Semi-Supervised Learning

Building user trust in dialogue agents requires smooth and consistent dialogue exchanges. However, agents can easily lose conversational context and generate irrelevant utterances. These situations are called dialogue breakdown, where agent utterances prevent users from continuing the conversation. Building systems to detect dialogue breakdown allows agents to recover appropriately or avoid breakdown entirely. In this paper we investigate the use of semi-supervised learning methods to improve dialogue breakdown detection, including continued pre-training on the Reddit dataset and a manifold-based data augmentation method. We demonstrate the effectiveness of these methods on the Dialogue Breakdown Detection Challenge (DBDC) English shared task. Our submissions to the 2020 DBDC5 shared task place first, beating baselines and other submissions by over 12\% accuracy. In ablations on DBDC4 data from 2019, our semi-supervised learning methods improve the performance of a baseline BERT model by 2\% accuracy. These methods are applicable generally to any dialogue task and provide a simple way to improve model performance.

preprint2022arXiv

D-Flat: A Differentiable Flat-Optics Framework for End-to-End Metasurface Visual Sensor Design

Optical metasurfaces are planar substrates with custom-designed, nanoscale features that selectively modulate incident light with respect to direction, wavelength, and polarization. When coupled with photodetectors and appropriate post-capture processing, they provide a means to create computational imagers and sensors that are exceptionally small and have distinctive capabilities. We introduce D-Flat, a framework in TensorFlow that renders physically-accurate images induced by metasurface optical systems. This framework is fully differentiable with respect to metasurface shape and post-capture computational parameters and allows simultaneous optimization with respect to almost any measure of sensor performance. D-Flat enables simulation of millimeter to centimeter diameter metasurfaces on commodity computers, and it is modular in the sense of accommodating a variety of wave optics models for scattering at the metasurface and for propagation to photosensors. We validate D-Flat against symbolic calculations and previous experimental measurements, and we provide simulations that demonstrate its ability to discover novel computational sensor designs for two applications: single-shot depth sensing and single-shot spatial frequency filtering.

preprint2022arXiv

FedMCSA: Personalized Federated Learning via Model Components Self-Attention

Federated learning (FL) facilitates multiple clients to jointly train a machine learning model without sharing their private data. However, Non-IID data of clients presents a tough challenge for FL. Existing personalized FL approaches rely heavily on the default treatment of one complete model as a basic unit and ignore the significance of different layers on Non-IID data of clients. In this work, we propose a new framework, federated model components self-attention (FedMCSA), to handle Non-IID data in FL, which employs model components self-attention mechanism to granularly promote cooperation between different clients. This mechanism facilitates collaboration between similar model components while reducing interference between model components with large differences. We conduct extensive experiments to demonstrate that FedMCSA outperforms the previous methods on four benchmark datasets. Furthermore, we empirically show the effectiveness of the model components self-attention mechanism, which is complementary to existing personalized FL and can significantly improve the performance of FL.

preprint2022arXiv

Neural Program Synthesis with Query

Aiming to find a program satisfying the user intent given input-output examples, program synthesis has attracted increasing interest in the area of machine learning. Despite the promising performance of existing methods, most of their success comes from the privileged information of well-designed input-output examples. However, providing such input-output examples is unrealistic because it requires the users to have the ability to describe the underlying program with a few input-output examples under the training distribution. In this work, we propose a query-based framework that trains a query neural network to generate informative input-output examples automatically and interactively from a large query space. The quality of the query depends on the amount of the mutual information between the query and the corresponding program, which can guide the optimization of the query framework. To estimate the mutual information more accurately, we introduce the functional space (F-space) which models the relevance between the input-output examples and the programs in a differentiable way. We evaluate the effectiveness and generalization of the proposed query-based framework on the Karel task and the list processing task. Experimental results show that the query-based framework can generate informative input-output examples which achieve and even outperform well-designed input-output examples.

preprint2022arXiv

Normalized Solutions to the mixed fractional Schrodinger equations with potential and general nonlinear term

The purpose of this paper is to establish the existence of solutions with prescribed norm to a class of nonlinear equations involving the mixed fractional Laplacians. This type of equations arises in various fields ranging from biophysics to population dynamics. Due to the importance of these applications, this topic has very recently received an increasing interest. Our method is novel and our results cover all the previous ones.

preprint2022arXiv

Real-Time Robust Video Object Detection System Against Physical-World Adversarial Attacks

DNN-based video object detection (VOD) powers autonomous driving and video surveillance industries with rising importance and promising opportunities. However, adversarial patch attack yields huge concern in live vision tasks because of its practicality, feasibility, and powerful attack effectiveness. This work proposes Themis, a software/hardware system to defend against adversarial patches for real-time robust video object detection. We observe that adversarial patches exhibit extremely localized superficial feature importance in a small region with non-robust predictions, and thus propose the adversarial region detection algorithm for adversarial effect elimination. Themis also proposes a systematic design to efficiently support the algorithm by eliminating redundant computations and memory traffics. Experimental results show that the proposed methodology can effectively recover the system from the adversarial attack with negligible hardware overhead.

preprint2022arXiv

Remote Work Optimization with Robust Multi-channel Graph Neural Networks

The spread of COVID-19 leads to the global shutdown of many corporate offices, and encourages companies to open more opportunities that allow employees to work from a remote location. As the workplace type expands from onsite offices to remote areas, an emerging challenge for an online hiring marketplace is how these remote opportunities and user intentions to work remotely can be modeled and matched without prior information. Despite the unprecedented amount of remote jobs posted amid COVID-19, there is no existing approach that can be directly applied. Introducing a brand new workplace type naturally leads to the cold-start problem, which is particularly more common for less active job seekers. It is challenging, if not impossible, to onboard a new workplace type for any predictive model if existing information sources can provide little information related to a new category of jobs, including data from resumes and job descriptions. Hence, in this work, we aim to propose a principled approach that jointly models the remoteness of job seekers and job opportunities with limited information, which also suffices the needs of web-scale applications. Existing research on the emerging type of remote workplace mainly focuses on qualitative studies, and classic predictive modeling approaches are inapplicable considering the problem of cold-start and information scarcity. We precisely try to close this gap with a novel graph neural architecture. Extensive experiments on large-scale data from real-world applications have been conducted to validate the superiority of the proposed approach over competitive baselines. The improvement may translate to more rapid onboarding of the new workplace type that can benefit job seekers who are interested in working remotely.

preprint2022arXiv

The impact of filaments on dwarf galaxy properties in the Auriga simulations

With a hydrodynamical simulation using a simple galaxy formation model without taking into account feedback, our previous work has shown that dense and massive filaments at high redshift can provide potential wells to trap and compress gas, and hence affect galaxy formation in their resident low-mass haloes. In this paper, we make use of the Auriga simulations, a suite of high-resolution zoom-in hydrodynamical simulations of Milky Way-like galaxies, to study whether the conclusion still holds in the simulations with a sophisticated galaxy formation model. In agreement with the results of our previous work, we find that, comparing to their counterparts with similar halo masses in field, dwarf galaxies residing in filaments tend to have higher baryonic and stellar fractions. At the fixed parent halo mass, the filament dwarfs tend to have slightly higher star formation rates than those of field ones. But overall we do not find a clear difference in galaxy g - r colours between the filament and field populations. We also show that at high redshifts, the gas components in dwarf galaxies tend to have their spins aligned with the filaments in which they reside. Our results support a picture in which massive filaments at high redshift assist gas accretion and enhance star formation in their resident dwarf sized dark matter haloes.

preprint2022arXiv

The spatial distribution of satellites in galaxy clusters

The planar distributions of satellite galaxies around the Milky Way and Andromeda have been extensively studied as potential challenges to the standard cosmological model. Using the Sloan Digital Sky Survey and the Millennium simulation we extend such studies to the satellite galaxies of massive galaxy clusters. We find that both observations and simulations of galaxy clusters show an excess of anisotropic satellite distributions. On average, satellites in clusters have a higher degree of anisotropy than their counterparts in Milky-Way-mass hosts once we account for the difference in their radial distributions. The normal vector of the plane of satellites is strongly aligned with the host halo&#39;s minor axis, while the alignment with the large-scale structure is weak. At fixed cluster mass, the degree of anisotropy is higher at higher redshift. This reflects the highly anisotropic nature of satellites accretion points, a feature that is partly erased by the subsequent orbital evolution of the satellites. We also find that satellite galaxies are mostly accreted singly so group accretion is not the explanation for the high flattening of the planes of satellites.

preprint2021arXiv

Multi-peak solitons in nonlocal nonlinear system with sine-oscillation response

The multi-peak solitons and their stability are investigated for the nonlocal nonlinear system with the sine-oscillation response, including both the cases of positive and negative Kerr coefficients. The Hermite-Gaussian-type multi-peak solitons and the ranges of the degree of nonlocality within which the solitons exist are analytically obtained by the variational approach. This is the first time, to our knowledge at least, to discuss the solution existence range of the multi-peak solitons analytically, although approximately. The variational analytical results are confirmed by the numerical ones. The stability of the multi-peak solitons are addressed by the linear stability analysis. It is found that the upper thresholds of the peak-number of the stable solitons are five and four for the system with negative and positive Kerr coefficients, respectively.

preprint2020arXiv

A Hybrid Lagrangian/Eulerian Collocated Advection and Projection Method for Fluid Simulation

We present a hybrid particle/grid approach for simulating incompressible fluids on collocated velocity grids. We interchangeably use particle and grid representations of transported quantities to balance efficiency and accuracy. A novel Backward Semi-Lagrangian method is derived to improve accuracy of grid based advection. Our approach utilizes the implicit formula associated with solutions of Burgers&#39; equation. We solve this equation using Newton&#39;s method enabled by $C^1$ continuous grid interpolation. We enforce incompressibility over collocated, rather than staggered grids. Our projection technique is variational and designed for B-spline interpolation over regular grids where multiquadratic interpolation is used for velocity and multilinear interpolation for pressure. Despite our use of regular grids, we extend the variational technique to allow for cut-cell definition of irregular flow domains for both Dirichlet and free surface boundary conditions.

preprint2020arXiv

Adaptive Precision Training: Quantify Back Propagation in Neural Networks with Fixed-point Numbers

Adaptive Precision Training: Quantify Back Propagation in Neural Networks with Fixed-point Numbers. Recent emerged quantization technique has been applied to inference of deep neural networks for fast and efficient execution. However, directly applying quantization in training can cause significant accuracy loss, thus remaining an open challenge.

preprint2020arXiv

DeText: A Deep Text Ranking Framework with BERT

Ranking is the most important component in a search system. Mostsearch systems deal with large amounts of natural language data,hence an effective ranking system requires a deep understandingof text semantics. Recently, deep learning based natural languageprocessing (deep NLP) models have generated promising results onranking systems. BERT is one of the most successful models thatlearn contextual embedding, which has been applied to capturecomplex query-document relations for search ranking. However,this is generally done by exhaustively interacting each query wordwith each document word, which is inefficient for online servingin search product systems. In this paper, we investigate how tobuild an efficient BERT-based ranking model for industry use cases.The solution is further extended to a general ranking framework,DeText, that is open sourced and can be applied to various rankingproductions. Offline and online experiments of DeText on threereal-world search systems present significant improvement overstate-of-the-art approaches.

preprint2020arXiv

Discovery of a Universal Correlation For Long and Short GRBs and Its Application on the Study of Luminosity Function and Formation Rate

Gamma-ray bursts (GRBs) are known to be the most violent explosions in the universe, and a variety of correlations between observable GRB properties have been proposed in literature, but none of these correlations is valid for both long GRBs and short GRBs. In this paper we report the discovery of a universal correlation which is suitable for both long and short GRBs using three prompt emission properties of GRBs, i.e. the isotropic peak luminosity $L_{\rm iso}$, the peak energy of the time-integtated prompt emission spectrum $E_{\rm peak}$, and the &#34;high signal&#34; timescale $T_{\rm 0.45}$, $L_{\rm iso} \propto E_{\rm peak}^{1.94} T_{0.45}^{0.37}$. This universal correlation just involves properties of GRB prompt emission and does not require any information of afterglow phase, which can be used as a relatively unbiased redshift estimator. Here we use this correlation to estimate the pseudo-redshifts for short Gamma Ray Bursts and then use Lynden-Bell method to obtain a non-parametric estimate of their luminosity function and formation rate. The luminosity function is $ψ(L_0)\propto{L_0^{-0.63\pm{0.07}}}$ for dim SGRBs and $ψ(L_0)\propto{L_0^{-1.96\pm{0.28}}}$ for bright SGRBs, with the break point $6.95_{-0.76}^{+0.84}\times10^{50} erg/s$. The local formation rate of SGRBs is about 15 events $\rm Gpc^{-3}yr^{-1}$ . This universal correlation may have important implications for GRB physics, implying that the long and short GRBs should share similar radiation processes.

preprint2020arXiv

DWM: A Decomposable Winograd Method for Convolution Acceleration

Winograd&#39;s minimal filtering algorithm has been widely used in Convolutional Neural Networks (CNNs) to reduce the number of multiplications for faster processing. However, it is only effective on convolutions with kernel size as 3x3 and stride as 1, because it suffers from significantly increased FLOPs and numerical accuracy problem for kernel size larger than 3x3 and fails on convolution with stride larger than 1. In this paper, we propose a novel Decomposable Winograd Method (DWM), which breaks through the limitation of original Winograd&#39;s minimal filtering algorithm to a wide and general convolutions. DWM decomposes kernels with large size or large stride to several small kernels with stride as 1 for further applying Winograd method, so that DWM can reduce the number of multiplications while keeping the numerical accuracy. It enables the fast exploring of larger kernel size and larger stride value in CNNs for high performance and accuracy and even the potential for new CNNs. Comparing against the original Winograd, the proposed DWM is able to support all kinds of convolutions with a speedup of ~2, without affecting the numerical accuracy.

preprint2020arXiv

L-GALAXIES 2020: Spatially resolved cold gas phases, star formation and chemical enrichment in galactic discs

We have updated the Munich galaxy formation model, L-Galaxies, to follow the radial distributions of stars and atomic and molecular gas in galaxy discs. We include an H2-based star-formation law, as well as a detailed chemical-enrichment model with explicit mass-dependent delay times for SN-II, SN-Ia and AGB stars. Information about the star formation, feedback and chemical-enrichment histories of discs is stored in 12 concentric rings. The new model retains the success of its predecessor in reproducing the observed evolution of the galaxy population, in particular, stellar mass functions and passive fractions over the redshift range 0<=z<=3 and mass range 8<=log(M_*/Msun)<=12, the black hole-bulge mass relation at z=0, galaxy morphology as a function of stellar mass and the mass-metallicity relations of both stellar and gas components. In addition, its detailed modelling of the radial structure of discs allows qualitatively new comparisons with observation, most notably with the relative sizes and masses of the stellar, atomic and molecular components in discs. Good agreement is found with recent data. Comparison of results obtained for simulations differing in mass resolution by more than two orders of magnitude shows that all important distributions are numerically well converged even for this more detailed model. An examination of metallicity and surface-density gradients in the stars and gas indicates that our new model, with star formation, chemical enrichment and feedback calculated self-consistently on local disc scales, reproduces some but not all of the trends seen in recent many-galaxy IFU surveys.

preprint2020arXiv

Local Group Analogs in $Λ$CDM cosmological simulations

We use semi-analytic galaxy catalogs based on two high-resolution cosmological $N$-body simulations, Millennium-WMAP7 and Millennium-II, to investigate the formation of the Local Group (LG) analogs. Unlike previous studies, we use the observed stellar masses to select the LG member (Milky Way (MW) and M31) analogs, and then impose constrains using the observed separation, isolation, and kinematics of the two main member galaxies. By comparing radial and low-ellipticity orbits between the MW and M31, we find higher tangential velocity results in higher total mass, which are 4.4$^{+2.4}_{-1.5}\times$10$^{12}\rm M_{\odot}$ and 6.6$^{+2.7}_{-1.5}\times$10$^{12}\rm M_{\odot}$ for radial and low-ellipticity orbits. The orbits also influence the individual mass distribution of MW and M31 analogs. For radial orbits, the typical host halo masses of the MW and M31 are 1.5$^{+1.4}_{-0.7}\times$10$^{12}\rm M_{\odot}$ and 2.5$^{+1.3}_{-1.1}\times$10$^{12}\rm M_{\odot}$; for low-ellipticity orbits, the masses are 2.5$^{+2.2}_{-1.4}\times$10$^{12}\rm M_{\odot}$ and 3.8$^{+2.8}_{-1.8}\times$10$^{12} \rm M_{\odot}$. The LG is located primarily in filaments with tails extending toward higher densities up to $δ\sim4.5$. The dark matter velocity anisotropy parameters $β$ of both the MW and M31 analogs are close to zero in the center, increasing to 0.2--0.3 at 50--80 kpc and decreasing slowly outward. The slope is much flatter than computed from the MW satellites, and the amplitude is smaller than traced by halo stars. Values of $β$ from different tracers agree at $\sim$120 kpc where $β\sim$ 0.2. We also find that model predictions agree broadly with observations in the radial distribution and luminosity function of satellites around the MW and M31.

preprint2020arXiv

MOTS: Multiple Object Tracking for General Categories Based On Few-Shot Method

Most modern Multi-Object Tracking (MOT) systems typically apply REID-based paradigm to hold a balance between computational efficiency and performance. In the past few years, numerous attempts have been made to perfect the systems. Although they presented favorable performance, they were constrained to track specified category. Drawing on the ideas of few shot method, we pioneered a new multi-target tracking system, named MOTS, which is based on metrics but not limited to track specific category. It contains two stages in series: In the first stage, we design the self-Adaptive-matching module to perform simple targets matching, which can complete 88.76% assignments without sacrificing performance on MOT16 training set. In the second stage, a Fine-match Network was carefully designed for unmatched targets. With a newly built TRACK-REID data-set, the Fine-match Network can perform matching of 31 category targets, even generalizes to unseen categories.

preprint2020arXiv

Multivariate General Compound Point Processes in Limit Order Books

In this paper, we focus on a new generalization of multivariate general compound Hawkes process (MGCHP), which we referred to as the multivariate general compound point process (MGCPP). Namely, we applied a multivariate point process to model the order flow instead of the Hawkes process. Law of large numbers (LLN) and two functional central limit theorems (FCLTs) for the MGCPP were proved in this work. Applications of the MGCPP in the limit order market were also considered. We provided numerical simulations and comparisons for the MGCPP and MGCHP by applying Google, Apple, Microsoft, Amazon, and Intel trading data.

preprint2020arXiv

NMR and NQR studies on transition-metal arsenide superconductors LaRu2As2, KCa2Fe4As4F2, and A2Cr3As3

We report 75As-nuclear magnetic resonance (NMR) and nuclear quadrupole resonance (NQR) measurements on transition-metal arsenides LaRu2As2, KCa2Fe4As4F2, and A2Cr3As3. In the superconducting state of LaRu2As2, a Hebel- Slichter coherence peak is found in the temperature dependence of the spin-lattice relaxation rate-1/T1 just below Tc, which indicates that LaRu2As2 is a full-gap superperconducor. For KCa2Fe4As4F2, antiferromagnetic spin fluctuations are observed in the normal state. We further find that the anisotropy rate RAF = Tc1/Tab1 is small and temperature independent, implying that the low energy spin fluctuations are isotropic in spin space. Our results indicate that KCa2Fe4As4F2 is a moderately overdoped iron-arsenide high-temperature superconductor with a stoichiometric composition. For A2Cr3As3, we calculate the electric field gradient by first-principle method and assign the 75As-NQR peaks with two crystallographically different As sites, paving the way for further NMR investigation.

preprint2020arXiv

The Luminosity Distribution of Short Gamma-Ray Bursts under a Structured Jet Scenario

Luminosity of GRB 170817A is much lower than that of other sGRBs. The measurement of the superluminal movement of the radio afterglow emission confirms the presence of the relativistic jet, and the emission features can be well explained by the structured jet model. In this paper, we calculate the luminosity distribution of sGRBs and its evolution with redshift based on the structured (Gaussian) jet model, and find that the typical luminosity increase with redshift, for nearby sGRBs (such as for luminosity distance less than 200 Mpc) the typical gamma-ray luminosity is just around 10^47-10^48 erg s-1, which naturally explains the very low radiation luminosity of GRB 170817A. We derived the detection probability of sGRBs by Fermi-GBM and found that the expected detection rate of sGRBs is only about 1 yr-1 within the distance of several hundred Mpc. We explored the effect of the power-law index α of the merger time distribution on the observed characteristics and found that it had little effect on the observed luminosity and viewing-angle distributions. However, it is very interesting that, for different values of α, the distributions of the number of observed sGRBs are quite different, so it is possible to determine the value of α through observed distributions of the number of sGRBs. We used the Bayesian method to make a quantitative analysis and found that the value of α may be identified when the number of observed sGRBs with known redshifts is more than 200. Finally, we compare our results of gamma-ray luminosity distribution with sGRBs with known redshifts, and found that our results are consistent with the observation, which implies that our simulation results can reproduce the observed luminosity distribution well.

preprint2019arXiv

An almost-solvable model of complex network dynamics

We discuss a specific model, which we refer to as RandLOE, of a large multi-agent network whose dynamic is prescribed via a combination of deterministic local laws and random exogenous factors. The RandLOE approach lies outside the framework of Stochastic Differential Equations, but lends itself to analytic examination as well as to stable simulation even for relatively large networks. RandLOE is based on the logistic operator equation (LOE), which is a multidimensional dynamical system extending the classical logistic equation via an operator-algebraic interaction term. The network is defined by interpreting the LOE variable as an adjacency matrix of a complete graph. Depending on the choice of parameters, it can display a number of essentially distinct dynamical characteristics: e.g. cycles of expansion and contraction.

preprint2019arXiv

HIKER: a halo-finding method based on kernel-shift algorithm

We introduce a new halo/subhalo finder, HIKER (a Halo fInder based on KERnel-shift algorithm), which takes advantage of a machine learning method -- the mean-shift algorithm combined with the Plummer kernel function, to effectively locate density peaks corresponding to halos/subhalos in density field. Based on these density peaks, dark matter halos are identified as spherical overdensity structures, and subhalos are bound substructures with boundaries at their tidal radius. By testing HIKER code with mock halos, we show that HIKER performs excellently in recovering input halo properties. Especially, HIKER has higher accuracy in locating halo/subhalo centres than most halo finders. With cosmological simulations, we further show that HIKER reproduces the abundance of dark matter halos and subhalos quite accurately, and the HIKER halo/subhalo mass functions and $V_{max}$ functions are in good agreement with two widely used halo finders, SUBFIND and AHF.