arXiv subjects

Wei Xing

Publications and source records attributed to Wei Xing.

43 records · Page 3Linked to original sources

arXiv2020

Scalable Variational Gaussian Process Regression Networks

Gaussian process regression networks (GPRN) are powerful Bayesian models for multi-output regression, but their inference is intractable. To address this issue, existing methods use a fully factorized structure (or a mixture of such structures) over all the outputs and latent functions for posterior approximation, which, however, can miss the strong posterior dependencies among the latent variables and hurt the inference quality. In addition, the updates of the variational parameters are inefficient and can be prohibitively expensive for a large number of outputs. To overcome these limitations, we propose a scalable variational inference algorithm for GPRN, which not only captures the abundant posterior dependencies but also is much more efficient for massive outputs. We tensorize the output space and introduce tensor/matrix-normal variational posteriors to capture the posterior correlations and to reduce the parameters. We jointly optimize all the parameters and exploit the inherent Kronecker product structure in the variational model evidence lower bound to accelerate the computation. We demonstrate the advantages of our method in several real-world applications.

cs.LG↗

arXiv2020

Diversified Arbitrary Style Transfer via Deep Feature Perturbation

Image style transfer is an underdetermined problem, where a large number of solutions can satisfy the same constraint (the content and style). Although there have been some efforts to improve the diversity of style transfer by introducing an alternative diversity loss, they have restricted generalization, limited diversity and poor scalability. In this paper, we tackle these limitations and propose a simple yet effective method for diversified arbitrary style transfer. The key idea of our method is an operation called deep feature perturbation (DFP), which uses an orthogonal random noise matrix to perturb the deep image feature maps while keeping the original style information unchanged. Our DFP operation can be easily integrated into many existing WCT (whitening and coloring transform)-based methods, and empower them to generate diverse results for arbitrary styles. Experimental results demonstrate that this learning-free and universal method can greatly increase the diversity while maintaining the quality of stylization.

cs.CV↗

arXiv2019

Consistent Robust Adversarial Prediction for General Multiclass Classification

We propose a robust adversarial prediction framework for general multiclass classification. Our method seeks predictive distributions that robustly optimize non-convex and non-continuous multiclass loss metrics against the worst-case conditional label distributions (the adversarial distributions) that (approximately) match the statistics of the training data. Although the optimized loss metrics are non-convex and non-continuous, the dual formulation of the framework is a convex optimization problem that can be recast as a risk minimization model with a prescribed convex surrogate loss we call the adversarial surrogate loss. We show that the adversarial surrogate losses fill an existing gap in surrogate loss construction for general multiclass classification problems, by simultaneously aligning better with the original multiclass loss, guaranteeing Fisher consistency, enabling a way to incorporate rich feature spaces via the kernel trick, and providing competitive performance in practice.

stat.ML↗

arXiv2019

Deep Coregionalization for the Emulation of Spatial-Temporal Fields

Data-driven surrogate models are widely used for applications such as design optimization and uncertainty quantification, where repeated evaluations of an expensive simulator are required. For most partial differential equation (PDE) simulators, the outputs of interest are often spatial or spatial-temporal fields, leading to very high-dimensional outputs. Despite the success of existing data-driven surrogates for high-dimensional outputs, most methods require a significant number of samples to cover the response surface in order to achieve a reasonable degree of accuracy. This demand makes the idea of surrogate models less attractive considering the high computational cost to generate the data. To address this issue, we exploit the multi-fidelity nature of a PDE simulator and introduce deep coregionalization, a Bayesian non-parametric autoregressive framework for efficient emulation of spatial-temporal fields. To effectively extract the output correlations in the context of multi-fidelity data, we develop a novel dimension reduction technique, residual principal component analysis. Our model can simultaneously capture the rich output correlations and the fidelity correlations and make high-fidelity predictions with only a few expensive, high-fidelity simulation samples. We show the advantages of our model in three canonical PDE models and a fluid dynamics problem. The results show that the proposed method cannot only approximate a simulator with significantly less cost (at bout 10%-25%) but also further improve model accuracy.

physics.comp-ph↗

arXiv2018

GLStyleNet: Higher Quality Style Transfer Combining Global and Local Pyramid Features

Recent studies using deep neural networks have shown remarkable success in style transfer especially for artistic and photo-realistic images. However, the approaches using global feature correlations fail to capture small, intricate textures and maintain correct texture scales of the artworks, and the approaches based on local patches are defective on global effect. In this paper, we present a novel feature pyramid fusion neural network, dubbed GLStyleNet, which sufficiently takes into consideration multi-scale and multi-level pyramid features by best aggregating layers across a VGG network, and performs style transfer hierarchically with multiple losses of different scales. Our proposed method retains high-frequency pixel information and low frequency construct information of images from two aspects: loss function constraint and feature fusion. Our approach is not only flexible to adjust the trade-off between content and style, but also controllable between global and local. Compared to state-of-the-art methods, our method can transfer not just large-scale, obvious style cues but also subtle, exquisite ones, and dramatically improves the quality of style transfer. We demonstrate the effectiveness of our approach on portrait style transfer, artistic style transfer, photo-realistic style transfer and Chinese ancient painting style transfer tasks. Experimental results indicate that our unified approach improves image style transfer quality over previous state-of-the-art methods, while also accelerating the whole process in a certain extent. Our code is available at https://github.com/EndyWon/GLStyleNet.

cs.CV↗

arXiv2013

Modeling and Performance Analysis of Pull-Based Live Streaming Schemes in Peer-to-Peer Network

Recent years mesh-based Peer-to-Peer live streaming has become a promising way for service providers to offer high-quality live video streaming service to Internet users. In this paper, we make a detailed study on modeling and performance analysis of the pull-based P2P streaming systems. We establish the analytical framework for the pull-based streaming schemes in P2P network, give accurate models of the chunk selection and peer selection strategies, and organize them into three categories, i.e., the chunk first scheme, the peer first scheme and the epidemic scheme. Through numerical performance evaluation, the impacts of some important parameters, such as size of neighbor set, reply number, buffer size and so on are investigated. For the peer first and chunk first scheme, we show that the pull-based schemes do not perform as well as the push-based schemes when peers are limited to reply only one request in each time slot. When the reply number increases, the pull-based streaming schemes will reach close to optimal playout probability. As to the pull-based epidemic scheme, we find it has unexpected poor performance, which is significantly different from the push-based epidemic scheme. Therefore we propose a simple, efficient and easily deployed push-pull scheme which can significantly improve the playout probability.

cs.NI↗

arXiv2012

Extended Equal Service and Differentiated Service Models for Peer-to-Peer File Sharing

Peer-to-Peer (P2P) systems have proved to be the most effective and popular file sharing applications in recent years. Previous studies mainly focus on the equal service and the differentiated service strategies when peers have no initial data before their download. In an upload-constrained P2P file sharing system, we model both the equal service process and the differentiated service process when peers' initial data distribution satisfies some special conditions, and also show how to minimize the time to get the file to any number of peers. The proposed models can reveal the intrinsic relations among the initial data amount, the size of peer set and the minimum last finish time. By using the models, we can also provide arbitrary degree of differentiated service to a certain number of peers. We believe that our analysis process and achieved theoretical results could provide fundamental insights into studies on bandwidth allocation and data scheduling, and can give helpful reference both for improving system performance and building effective incentive mechanism in P2P file sharing systems.

cs.NI↗