arXiv subjects

Yue Liu

Publications and source records attributed to Yue Liu.

At least 343 records · Page 19Linked to original sources

arXiv2020

Video Moment Retrieval via Natural Language Queries

In this paper, we propose a novel method for video moment retrieval (VMR) that achieves state of the arts (SOTA) performance on R@1 metrics and surpassing the SOTA on the high IoU metric (R@1, IoU=0.7). First, we propose to use a multi-head self-attention mechanism, and further a cross-attention scheme to capture video/query interaction and long-range query dependencies from video context. The attention-based methods can develop frame-to-query interaction and query-to-frame interaction at arbitrary positions and the multi-head setting ensures the sufficient understanding of complicated dependencies. Our model has a simple architecture, which enables faster training and inference while maintaining . Second, We also propose to use multiple task training objective consists of moment segmentation task, start/end distribution prediction and start/end location regression task. We have verified that start/end prediction are noisy due to annotator disagreement and joint training with moment segmentation task can provide richer information since frames inside the target clip are also utilized as positive training examples. Third, we propose to use an early fusion approach, which achieves better performance at the cost of inference time. However, the inference time will not be a problem for our model since our model has a simple architecture which enables efficient training and inference.

cs.CV↗

arXiv2020

A Blockchain-based Platform Architecture for Multimedia Data Management

Massive amounts of multimedia data (i.e., text, audio, video, graphics and animation) are being generated everyday. Conventionally, multimedia data are managed by the platforms maintained by multimedia service providers, which are generally designed using centralised architecture. However, such centralised architecture may lead to a single point of failure and disputes over royalties or other rights. It is hard to ensure the data integrity and track fulfilment of obligations listed on the copyright agreement. To tackle these issues, in this paper, we present a blockchain-based platform architecture for multimedia data management. We adopt self-sovereign identity for identity management and design a multi-level capability-based mechanism for access control. We implement a proof-of-concept prototype using the proposed approach and evaluate it using a use case. The results show that the proposed approach is feasible and has scalable performance.

cs.CR↗

arXiv2020

Adding Seemingly Uninformative Labels Helps in Low Data Regimes

Evidence suggests that networks trained on large datasets generalize well not solely because of the numerous training examples, but also class diversity which encourages learning of enriched features. This raises the question of whether this remains true when data is scarce - is there an advantage to learning with additional labels in low-data regimes? In this work, we consider a task that requires difficult-to-obtain expert annotations: tumor segmentation in mammography images. We show that, in low-data settings, performance can be improved by complementing the expert annotations with seemingly uninformative labels from non-expert annotators, turning the task into a multi-class problem. We reveal that these gains increase when less expert data is available, and uncover several interesting properties through further studies. We demonstrate our findings on CSAW-S, a new dataset that we introduce here, and confirm them on two public datasets.

cs.CV↗

arXiv2020

Online NEAT for Credit Evaluation -- a Dynamic Problem with Sequential Data

In this paper, we describe application of Neuroevolution to a P2P lending problem in which a credit evaluation model is updated based on streaming data. We apply the algorithm Neuroevolution of Augmenting Topologies (NEAT) which has not been widely applied generally in the credit evaluation domain. In addition to comparing the methodology with other widely applied machine learning techniques, we develop and evaluate several enhancements to the algorithm which make it suitable for the particular aspects of online learning that are relevant in the problem. These include handling unbalanced streaming data, high computation costs, and maintaining model similarity over time, that is training the stochastic learning algorithm with new data but minimizing model change except where there is a clear benefit for model performance

cs.LG↗

arXiv2020

Orbital Stability of smooth solitary waves for the Degasperis-Procesi Equation

The Degasperis-Procesi equation is the integrable Camassa-Holm-type model which is an asymptotic approximation for the unidirectional propagation of shallow water waves. This work establishes the orbital stability of localized smooth solitary waves to the Desgasperis-Procesi (DP) equation on the real line. %extending our previous work on their spectral stability \cite{LLW}. The main difficulty stems from the fact that the translation symmetry for the DP equation gives rise to a conserved quantity equivalent to the $L^2$-norm, which by itself can not bound the higher-order nonlinear terms in the Lagrangian. The remedy is to observe that, given a sufficiently smooth initial condition satisfying a measurable constraint, the $L^\infty$ orbital norm of the perturbation is bounded above by a function of its $L^2$ orbital norm, yielding the orbital stability in the $L^2\cap L^\infty$ space.

math.AP↗

arXiv2020

Design Patterns for Blockchain-based Self-Sovereign Identity

Self-sovereign identity is a new identity management paradigm that allows entities to really have the ownership of their identity data and control their use without involving any intermediary. Blockchain is an enabling technology for building self-sovereign identity systems by providing a neutral and trustable storage and computing infrastructure and can be viewed as a component of the systems. Both blockchain and self-sovereign identity are emerging technologies which could present a steep learning curve for architects. We collect and propose 12 design patterns for blockchain-based self-sovereign identity systems to help the architects understand and easily apply the concepts in system design. Based on the lifecycles of three main objects involved in self-sovereign identity, we categorise the patterns into three groups: key management patterns, decentralised identifier management patterns, and credential design patterns. The proposed patterns provide a systematic and holistic guide for architects to design the architecture of blockchain-based self-sovereign identity systems.

cs.CR↗

arXiv2020

Real-time Human Activity Recognition Using Conditionally Parametrized Convolutions on Mobile and Wearable Devices

Recently, deep learning has represented an important research trend in human activity recognition (HAR). In particular, deep convolutional neural networks (CNNs) have achieved state-of-the-art performance on various HAR datasets. For deep learning, improvements in performance have to heavily rely on increasing model size or capacity to scale to larger and larger datasets, which inevitably leads to the increase of operations. A high number of operations in deep leaning increases computational cost and is not suitable for real-time HAR using mobile and wearable sensors. Though shallow learning techniques often are lightweight, they could not achieve good performance. Therefore, deep learning methods that can balance the trade-off between accuracy and computation cost is highly needed, which to our knowledge has seldom been researched. In this paper, we for the first time propose a computation efficient CNN using conditionally parametrized convolution for real-time HAR on mobile and wearable devices. We evaluate the proposed method on four public benchmark HAR datasets consisting of WISDM dataset, PAMAP2 dataset, UNIMIB-SHAR dataset, and OPPORTUNITY dataset, achieving state-of-the-art accuracy without compromising computation cost. Various ablation experiments are performed to show how such a network with large capacity is clearly preferable to baseline while requiring a similar amount of operations. The method can be used as a drop-in replacement for the existing deep HAR architectures and easily deployed onto mobile and wearable devices for real-time HAR applications.

cs.CV↗

arXiv2020

Stable Prediction via Leveraging Seed Variable

In this paper, we focus on the problem of stable prediction across unknown test data, where the test distribution is agnostic and might be totally different from the training one. In such a case, previous machine learning methods might exploit subtly spurious correlations in training data induced by non-causal variables for prediction. Those spurious correlations are changeable across data, leading to instability of prediction across data. By assuming the relationships between causal variables and response variable are invariant across data, to address this problem, we propose a conditional independence test based algorithm to separate those causal variables with a seed variable as priori, and adopt them for stable prediction. By assuming the independence between causal and non-causal variables, we show, both theoretically and with empirical experiments, that our algorithm can precisely separate causal and non-causal variables for stable prediction across test data. Extensive experiments on both synthetic and real-world datasets demonstrate that our algorithm outperforms state-of-the-art methods for stable prediction.

cs.LG↗

arXiv2020

Predicting the Porosity Formed in Freeze Casting by Artificial Neural Network

Freeze casting has been increasingly applied to process various porous materials. A linear relationship between the final porosity and the initial solid material fraction in the suspension was reported by other researchers. However, the relationship of the volume fraction between the porosity and the solid material shows high divergence among different materials or frozen solvents, as the nature of materials significantly affects the pores formed in freeze casting. Here, we proposed an artificial neural network (ANN) to evaluate the porosity in freeze casting process. After well training the ANN model on experimental data, a porosity value can be predicted from four inputs which describe the most influential process conditions. The error range, reliability and optimization of the model were also analyzed and discussed in this study. The results proved that this method effectively summarizes a general rule for diverse materials in one model, which is difficult to be realized by linear fitting. Finally, a user-friendly mini program based on a well-trained ANN model is also provided to predict the porosity level for customized freeze-casting experiments.

physics.app-ph↗

arXiv2020

Design-Pattern-as-a-Service for Blockchain-based Self-Sovereign Identity

Self-sovereign identity (SSI) is considered to be a "killer application" of blockchain. However, there is a lack of systematic architecture designs for blockchain-based SSI systems to support methodical development. An aspect of such gap is demonstrated in current solutions, which are considered coarse grained and may increase data security risks. In this paper, we first identify the lifecycles of three major SSI objects (i.e., key, identifier, and credential) and present fine-grained design patterns critical for application development. These patterns are associated with particular state transitions, providing a systematic view of system interactions and serving as a guidance for effective use of these patterns. Further, we present an SSI platform architecture, which advocates the notion of Design-Pattern-as-a-Service. Each design pattern serves as an API by wrapping the respective pattern code to ease application development and improve scalability and security. We implement a prototype and evaluate it on feasibility and scalability.

cs.SE↗

arXiv2020

RecipeGPT: Generative Pre-training Based Cooking Recipe Generation and Evaluation System

Interests in the automatic generation of cooking recipes have been growing steadily over the past few years thanks to a large amount of online cooking recipes. We present RecipeGPT, a novel online recipe generation and evaluation system. The system provides two modes of text generations: (1) instruction generation from given recipe title and ingredients; and (2) ingredient generation from recipe title and cooking instructions. Its back-end text generation module comprises a generative pre-trained language model GPT-2 fine-tuned on a large cooking recipe dataset. Moreover, the recipe evaluation module allows the users to conveniently inspect the quality of the generated recipe contents and store the results for future reference. RecipeGPT can be accessed online at https://recipegpt.org/.

cs.CL↗

arXiv2020

A Note on Parallel Distinguishability of two Quantum Operations

We consider a homogeneous system of linear equations of the form $A_α^{\otimes N} {\bf x} = 0$ arising from the distinguishability of two quantum operations by $N$ uses in parallel, where the coefficient matrix $A_α$ depends on a real parameter $α$. It was conjectured by Duan et al. that the system has a non-trivial nonnegative solution if and only if $α$ lies in a certain interval $R_N$ depending on $N$. We affirm the necessity part of the conjecture and establish the sufficiency of the conjecture for $N\leq 10$ by presenting explicit non-trivial nonnegative solutions for the linear system.

quant-ph↗

arXiv2019

Spectral stability of smooth solitary waves for the Degasperis-Procesi Equation

The Degasperis-Procesi equation is an approximating model of shallow-water wave propagating mainly in one direction to the Euler equations. Such a model equation is analogous to the Camassa-Holm approximation of the two-dimensional incompressible and irrotational Euler equations with the same asymptotic accuracy, and is integrable with the bi-Hamiltonian structure. In the present study, we establish existence and spectral stability results of localized smooth solitons to the Degasperis-Procesi equation on the real line. The stability proof relies essentially on refined spectral analysis of the linear operator corresponding to the second-order variational derivative of the Hamiltonian of the Degasperis-Procesi equation.

math.AP↗

arXiv2019

Spots, strips, and spiral waves in models for static and motile cells

The polarization and motility of eukaryotic cells depends on assembly and contraction of the actin cytoskeleton and its regulation by proteins called GTPases. The activity of GTPases causes assembly of filamentous actin (by GTPases Cdc42, Rac), resulting in protrusion of the cell edge. Mathematical models for GTPase dynamics address the spontaneous formation of patterns and nonuniform spatial distributions of such proteins in the cell. Here we revisit the wave-pinning model for GTPase-induced cell polarization, together with a number of extensions proposed in the literature. These include introduction of sources and sinks of active and inactive GTPase (by the group of A. Champneys), and negative feedback from F-actin to GTPase activity. We discuss these extensions singly and in combination, in 1D, and 2D static domains. We then show how the patterns that form (spots, waves, and spirals) interact with cell boundaries to create a variety of interesting and dynamic cell shapes and motion.

q-bio.CB↗

arXiv2019

Characterizing and Predicting Repeat Food Consumption Behavior for Just-in-Time Interventions

Human beings are creatures of habit. In their daily life, people tend to repeatedly consume similar types of food items over several days and occasionally switch to consuming different types of items when the consumptions become overly monotonous. However, the novel and repeat consumption behaviors have not been studied in food recommendation research. More importantly, the ability to predict daily eating habits of individuals is crucial to improve the effectiveness of food recommender systems in facilitating healthy lifestyle change. In this study, we analyze the patterns of repeat food consumptions using large-scale consumption data from a popular online fitness community called MyFitnessPal (MFP), conduct an offline evaluation of various state-of-the-art algorithms in predicting the next-day food consumption, and analyze their performance across different demographic groups and contexts. The experiment results show that algorithms incorporating the exploration-and-exploitation and temporal dynamics are more effective in the next-day recommendation task than most state-of-the-art algorithms.

cs.IR↗

arXiv2019

Estimating Glycemic Impact of Cooking Recipes via Online Crowdsourcing and Machine Learning

Consumption of diets with low glycemic impact is highly recommended for diabetics and pre-diabetics as it helps maintain their blood glucose levels. However, laboratory analysis of dietary glycemic potency is time-consuming and expensive. In this paper, we explore a data-driven approach utilizing online crowdsourcing and machine learning to estimate the glycemic impact of cooking recipes. We show that a commonly used healthiness metric may not always be effective in determining recipes suitable for diabetics, thus emphasizing the importance of the glycemic-impact estimation task. Our best classification model, trained on nutritional and crowdsourced data obtained from Amazon Mechanical Turk (AMT), can accurately identify recipes which are unhealthful for diabetics.

cs.CY↗

arXiv2019

Causal Discovery by Kernel Intrinsic Invariance Measure

Reasoning based on causality, instead of association has been considered as a key ingredient towards real machine intelligence. However, it is a challenging task to infer causal relationship/structure among variables. In recent years, an Independent Mechanism (IM) principle was proposed, stating that the mechanism generating the cause and the one mapping the cause to the effect are independent. As the conjecture, it is argued that in the causal direction, the conditional distributions instantiated at different value of the conditioning variable have less variation than the anti-causal direction. Existing state-of-the-arts simply compare the variance of the RKHS mean embedding norms of these conditional distributions. In this paper, we prove that this norm-based approach sacrifices important information of the original conditional distributions. We propose a Kernel Intrinsic Invariance Measure (KIIM) to capture higher order statistics corresponding to the shapes of the density functions. We show our algorithm can be reduced to an eigen-decomposition task on a kernel matrix measuring intrinsic deviance/invariance. Causal directions can then be inferred by comparing the KIIM scores of two hypothetic directions. Experiments on synthetic and real data are conducted to show the advantages of our methods over existing solutions.

cs.LG↗

arXiv2019

Cell size, mechanical tension, and GTPase signaling in the Single Cell

Cell polarization requires redistribution of specific proteins to the nascent front and back of a eukarytotic cell. Among these proteins are Rac and Rho, members of the small GTPase family that regulate the actin cytoskeleton. Rac promotes actin assembly and protrusion of the front edge, whereas Rho activates myosin-driven contraction at the back. Mathematical models of cell polarization at many levels of detail have appeared. One of the simplest based on "wave-pinning", consists of a pair of reaction-diffusion equations for a single GTPase. Mathematical analysis of wave-pinning so far is largely restricted to static domains in one spatial dimension. Here we extend the analysis to cells that change in size, showing that both shrinking and growing cells can lose polarity. We further consider the feedback between mechanical tension, GTPase activation, and cell deformation in both static, growing, shrinking, and moving cells. Special cases (spatially uniform cell chemistry, absence or presence of mechanical feedback) are analyzed, and the full model is explored by simulations in 1D. We find a variety of novel behaviors, including "dilution-induced" oscillations of Rac activity and cell size, as well as gain or loss of polarization and motility in the model cell.

q-bio.CB↗