arXiv subjects
Publications and source records attributed to Yu Gu.
We study the one-dimensional KPZ equation on a large torus, started at equilibrium. The main results are optimal variance bounds in the super-relaxation regime and part of the relaxation regime.
A trending paradigm for multiple-choice question answering (MCQA) is using a text-to-text framework. By unifying data in different tasks into a single text-to-text format, it trains a generative encoder-decoder model which is both powerful and universal. However, a side effect of twisting a generation target to fit the classification nature of MCQA is the under-utilization of the decoder and the knowledge that can be decoded. To exploit the generation capability and underlying knowledge of a pre-trained encoder-decoder model, in this paper, we propose a generation-enhanced MCQA model named GenMC. It generates a clue from the question and then leverages the clue to enhance a reader for MCQA. It outperforms text-to-text models on multiple MCQA datasets.
We propose a general framework for creating parameterized control schemes for decentralized multi-robot systems. A variety of tasks can be seen in the decentralized multi-robot literature, each with many possible control schemes. For several of them, the agents choose control velocities using algorithms that extract information from the environment and combine that information in meaningful ways. From this basic formation, a framework is proposed that classifies each robots' measurement information as sets of relevant scalars and vectors and creates a linear combination of the measured vector sets. Along with an optimizable parameter set, the scalar measurements are used to generate the coefficients for the linear combination. With this framework and Bayesian optimization, we can create effective control systems for several multi-robot tasks, including cohesion and segregation, pattern formation, and searching/foraging.
With the development of temporal networks such as E-commerce networks and social networks, the issue of temporal link prediction has attracted increasing attention in recent years. The Temporal Link Prediction task of WSDM Cup 2022 expects a single model that can work well on two kinds of temporal graphs simultaneously, which have quite different characteristics and data properties, to predict whether a link of a given type will occur between two given nodes within a given time span. Our team, named as nothing here, regards this task as a link prediction task in heterogeneous temporal networks and proposes a generic model, i.e., Heterogeneous Temporal Graph Network (HTGN), to solve such temporal link prediction task with the unfixed time intervals and the diverse link types. That is, HTGN can adapt to the heterogeneity of links and the prediction with unfixed time intervals within an arbitrary given time period. To train the model, we design a Bi-Time-Window training strategy (BTW) which has two kinds of mini-batches from two kinds of time windows. As a result, for the final test, we achieved an AUC of 0.662482 on dataset A, an AUC of 0.906923 on dataset B, and won 2nd place with an Average T-scores of 0.628942.
We consider a nonlinear stochastic heat equation in spatial dimension $d=2$, forced by a white-in-time multiplicative Gaussian noise with spatial correlation length $\varepsilon>0$ but divided by a factor of $\sqrt{\log\varepsilon^{-1}}$. We impose a condition on the Lipschitz constant of the nonlinearity so that the problem is in the "weak noise" regime. We show that, as $\varepsilon\downarrow0$, the one-point distribution of the solution converges, with the limit characterized in terms of the solution to a forward-backward stochastic differential equation (FBSDE). We also characterize the limiting multipoint statistics of the solution, when the points are chosen on appropriate scales, in similar terms. Our approach is new even for the linear case, in which the FBSDE can be solved explicitly and we recover results of Caravenna, Sun, and Zygouras (Ann. Appl. Probab. 27(5):3050--3112, 2017).
Motivation: A perennial challenge for biomedical researchers and clinical practitioners is to stay abreast with the rapid growth of publications and medical notes. Natural language processing (NLP) has emerged as a promising direction for taming information overload. In particular, large neural language models facilitate transfer learning by pretraining on unlabeled text, as exemplified by the successes of BERT models in various NLP applications. However, fine-tuning such models for an end task remains challenging, especially with small labeled datasets, which are common in biomedical NLP. Results: We conduct a systematic study on fine-tuning stability in biomedical NLP. We show that finetuning performance may be sensitive to pretraining settings, especially in low-resource domains. Large models have potential to attain better performance, but increasing model size also exacerbates finetuning instability. We thus conduct a comprehensive exploration of techniques for addressing fine-tuning instability. We show that these techniques can substantially improve fine-tuning performance for lowresource biomedical NLP applications. Specifically, freezing lower layers is helpful for standard BERT-BASE models, while layerwise decay is more effective for BERT-LARGE and ELECTRA models. For low-resource text similarity tasks such as BIOSSES, reinitializing the top layer is the optimal strategy. Overall, domainspecific vocabulary and pretraining facilitate more robust models for fine-tuning. Based on these findings, we establish new state of the art on a wide range of biomedical NLP applications. Availability and implementation: To facilitate progress in biomedical NLP, we release our state-of-the-art pretrained and fine-tuned models: https://aka.ms/BLURB.
We consider a particle undergoing Brownian motion in Euclidean space of any dimension, forced by a Gaussian random velocity field that is white in time and smooth in space. We show that conditional on the velocity field, the quenched density of the particle after a long time can be approximated pointwise by the product of a deterministic Gaussian density and a spacetime-stationary random field $U$. If the velocity field is additionally assumed to be incompressible, then $U\equiv 1$ almost surely and we obtain a local central limit theorem.
We studied the linear electro-optic effect of chiral topological semimetal RhSi which is characterized by high-fold chiral fermions separated in energy space. We identify that the general second order conductivity $σ^{(2)}_{xyz}(ω=ω_{1}+ω_{2};ω_{1},ω_{2})$ includes a real symmetric component and an imaginary antisymmetric component, which are from the inter-band shift and intra-band injection current with frequency $ω$, respectively. The $σ^{(2)}_{xyz}$ is significantly enhanced by the high electron velocity and nontrivial band topology of chiral fermion, and modifies the phase velocity of light wave. We also predict that the electro-optic coefficient $χ_{xyz}^{(2)}(ω;ω,0)$ of chiral crystal RhSi is about 7000 pm/V at photon energy 0.01 eV and 1.1 eV, which is about 200 times that of widely used LiNbO$_{3}$ crystal. The giant electro-optic coefficient renders a relatively low half-wave voltage in order of hundreds volt, and demonstrates potential application as electro-optic crystals for the wavelength in the second telecom window of optical fiber communications.
Plans for establishing a long-term human presence on the Moon will require substantial increases in robot autonomy and multi-robot coordination to support establishing a lunar outpost. To achieve these objectives, algorithm design choices for the software developments need to be tested and validated for expected scenarios such as autonomous in-situ resource utilization (ISRU), localization in challenging environments, and multi-robot coordination. However, real-world experiments are extremely challenging and limited for extraterrestrial environment. Also, realistic simulation demonstrations in these environments are still rare and demanded for initial algorithm testing capabilities. To help some of these needs, the NASA Centennial Challenges program established the Space Robotics Challenge Phase 2 (SRC2) which consist of virtual robotic systems in a realistic lunar simulation environment, where a group of mobile robots were tasked with reporting volatile locations within a global map, excavating and transporting these resources, and detecting and localizing a target of interest. The main goal of this article is to share our team's experiences on the design trade-offs to perform autonomous robotic operations in a virtual lunar environment and to share strategies to complete the mission requirements posed by NASA SRC2 competition during the qualification round. Of the 114 teams that registered for participation in the NASA SRC2, team Mountaineers finished as one of only six teams to receive the top qualification round prize.
Aspect-based sentiment analysis (ABSA) is an emerging fine-grained sentiment analysis task that aims to extract aspects, classify corresponding sentiment polarities and find opinions as the causes of sentiment. The latest research tends to solve the ABSA task in a unified way with end-to-end frameworks. Yet, these frameworks get fine-tuned from downstream tasks without any task-adaptive modification. Specifically, they do not use task-related knowledge well or explicitly model relations between aspect and opinion terms, hindering them from better performance. In this paper, we propose SentiPrompt to use sentiment knowledge enhanced prompts to tune the language model in the unified framework. We inject sentiment knowledge regarding aspects, opinions, and polarities into prompt and explicitly model term relations via constructing consistency and polarity judgment templates from the ground truth triplets. Experimental results demonstrate that our approach can outperform strong baselines on Triplet Extraction, Pair Extraction, and Aspect Term Extraction with Sentiment Classification by a notable margin.
Pretraining large neural language models, such as BERT, has led to impressive gains on many natural language processing (NLP) tasks. However, most pretraining efforts focus on general domain corpora, such as newswire and Web. A prevailing assumption is that even domain-specific pretraining can benefit by starting from general-domain language models. In this paper, we challenge this assumption by showing that for domains with abundant unlabeled text, such as biomedicine, pretraining language models from scratch results in substantial gains over continual pretraining of general-domain language models. To facilitate this investigation, we compile a comprehensive biomedical NLP benchmark from publicly-available datasets. Our experiments show that domain-specific pretraining serves as a solid foundation for a wide range of biomedical NLP tasks, leading to new state-of-the-art results across the board. Further, in conducting a thorough evaluation of modeling choices, both for pretraining and task-specific fine-tuning, we discover that some common practices are unnecessary with BERT models, such as using complex tagging schemes in named entity recognition (NER). To help accelerate research in biomedical NLP, we have released our state-of-the-art pretrained and task-specific models for the community, and created a leaderboard featuring our BLURB benchmark (short for Biomedical Language Understanding & Reasoning Benchmark) at https://aka.ms/BLURB.
For a Brownian directed polymer in a Gaussian random environment, with $q(t,\cdot)$ denoting the quenched endpoint density and \[ Q_n(t,x_1,\ldots,x_n)=\mathbf{E}[q(t,x_1)\ldots q(t,x_n)], \] we derive a hierarchical PDE system satisfied by $\{Q_n\}_{n\geq1}$. We present two applications of the system: (i) we compute the generator of $\{μ_t(dx)=q(t,x)dx\}_{t\geq0}$ for some special functionals, where $\{μ_t(dx)\}_{t\geq0}$ is viewed as a Markov process taking values in the space of probability measures; (ii) in the high temperature regime with $d\geq 3$, we prove a quantitative central limit theorem for the annealed endpoint distribution of the diffusively rescaled polymer path. We also study a nonlocal diffusion-reaction equation motivated by the generator and establish a super-diffusive $O(t^{2/3})$ scaling.
Skyline computation is an essential database operation that has many applications in multi-criteria decision making scenarios such as recommender systems. Existing algorithms have focused on checking point domination, which lack efficiency over large datasets. We propose a grid-based structure that enables grid cell domination checks. We show that only a small constant number of cells need to be checked which is independent from the number of data points. Our structure also enables parallel processing. We thus obtain a highly efficient parallel skyline algorithm named SkyCell, taking advantage of the parallelization power of graphics processing units. Experimental results confirm the effectiveness and efficiency of SkyCell -- it outperforms state-of-the-art algorithms consistently and by up to over two orders of magnitude in the computation time.
Knowledge bases (KBs) and text often contain complementary knowledge: KBs store structured knowledge that can support long range reasoning, while text stores more comprehensive and timely knowledge in an unstructured way. Separately embedding the individual knowledge sources into vector spaces has demonstrated tremendous successes in encoding the respective knowledge, but how to jointly embed and reason with both knowledge sources to fully leverage the complementary information is still largely an open problem. We conduct a large-scale, systematic investigation of aligning KB and text embeddings for joint reasoning. We set up a novel evaluation framework with two evaluation tasks, few-shot link prediction and analogical reasoning, and evaluate an array of KB-text embedding alignment methods. We also demonstrate how such alignment can infuse textual information into KB embeddings for more accurate link prediction on emerging entities and events, using COVID-19 as a case study.
Due to the wide applications in recommendation systems, multi-class label prediction and deep learning, the Maximum Inner Product (MIP) search problem has received extensive attention in recent years. Faced with large-scale datasets containing high-dimensional feature vectors, the state-of-the-art LSH-based methods usually require a large number of hash tables or long hash codes to ensure the searching quality, which takes up lots of index space and causes excessive disk page accesses. In this paper, we relax the guarantee of accuracy for efficiency and propose an efficient method for c-Approximate Maximum Inner Product (c-AMIP) search with a lightweight iDistance index. We project high-dimensional points to low-dimensional ones via 2-stable random projections and derive probability-guaranteed searching conditions, by which the c-AMIP results can be guaranteed in accuracy with arbitrary probabilities. To further improve the efficiency, we propose Quick-Probe for quickly determining the searching bound satisfying the derived condition in advance, avoiding the inefficient incremental searching process. Extensive experimental evaluations on four real datasets demonstrate that our method requires less pre-processing cost including index size and pre-processing time. In addition, compared to the state-of-the-art benchmark methods, it provides superior results on searching quality in terms of overall ratio and recall, and efficiency in terms of page access and running time.
We study platoon control of vehicles with linear third-order longitudinal dynamics under the constant time headway policy. The controller of each follower vehicle is only based on its own velocity, acceleration, inter-vehicle distance and velocity difference with respect to its immediate predecessor, which are all obtained by on-board sensors. We develop distributed cooperative extended state observers for followers to estimate the acceleration differences between adjacent vehicles. Based on estimates of the acceleration differences, distributed cooperative control laws are designed. By using the stability theory of perturbed linear systems, we show that the control parameters can be properly designed to ensure the closed-loop and L2 string stabilities for any given positive time headway. We further show that the proposed control law based on the ideal vehicle model can guarantee the closed-loop and L2 string stabilities even if there are small model parameter uncertainties. Also, simulation results demonstrate the robustness of the proposed control law against sensing noises, input delays and parameter uncertainties.
The zero-velocity update (ZUPT) algorithm provides valuable state information to maintain the inertial navigation system (INS) reliability when stationary conditions are satisfied. Employing ZUPT along with leveraging non-holonomic constraints can greatly benefit wheeled mobile robot dead-reckoning localization accuracy. However, determining how often they should be employed requires consideration to balance localization accuracy and traversal rate for planetary rovers. To address this, we investigate when to autonomously initiate stops to improve wheel-inertial odometry (WIO) localization performance with ZUPT. To do this, we propose a 3D dead-reckoning approach that predicts wheel slippage while the rover is in motion and forecasts the appropriate time to stop without changing any rover hardware or major rover operations. We validate with field tests that our approach is viable on different terrain types and achieves a 3D localization accuracy of more than 97% over 650 m drives on rough terrain.
Existing studies on question answering on knowledge bases (KBQA) mainly operate with the standard i.i.d assumption, i.e., training distribution over questions is the same as the test distribution. However, i.i.d may be neither reasonably achievable nor desirable on large-scale KBs because 1) true user distribution is hard to capture and 2) randomly sample training examples from the enormous space would be highly data-inefficient. Instead, we suggest that KBQA models should have three levels of built-in generalization: i.i.d, compositional, and zero-shot. To facilitate the development of KBQA models with stronger generalization, we construct and release a new large-scale, high-quality dataset with 64,331 questions, GrailQA, and provide evaluation settings for all three levels of generalization. In addition, we propose a novel BERT-based KBQA model. The combination of our dataset and model enables us to thoroughly examine and demonstrate, for the first time, the key role of pre-trained contextual embeddings like BERT in the generalization of KBQA.