Paper 解读 - Arxiv 论文中文解读平台

cs.LG 1701.06538

Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer

提出稀疏门控专家混合层（MoE），实现模型容量提升超1000倍，参数达1370亿，显著优于现有技术。

Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz 等

2017-01-24 5522 引用 31

cs.LG 1701.05517

PixelCNN++: Improving the PixelCNN with Discretized Logistic Mixture Likelihood and Other Modifications

PixelCNN++采用离散Logistic混合似然，提升训练速度和生成质量，达CIFAR-10最优对数似然。

Tim Salimans, Andrej Karpathy, Xi Chen 等

2017-01-20 66

stat.ML 1701.05369

Variational Dropout Sparsifies Deep Neural Networks

提出稀疏化深度神经网络的变分Dropout，参数减少达280倍，精度影响微小。

Dmitry Molchanov, Arsenii Ashukha, Dmitry Vetrov

2017-01-19 55

cs.CR 1701.04525

Cyber-Physical Systems Security -- A Survey

提出CPS安全统一框架，涵盖威胁、漏洞、攻击与控制，结合异构组件分析安全风险。

Abdulmalik Humayed, Jingqiang Lin, Fengjun Li 等

2017-01-17 32

cs.LG 1701.04238

Thompson Sampling For Stochastic Bandits with Graph Feedback

提出Thompson Sampling扩展算法，解决图反馈下的随机决策问题，显著降低贝叶斯遗憾。

Aristide C. Y. Tossou, Christos Dimitrakakis, Devdatt Dubhashi

2017-01-16 19

stat.ME 1701.03535

Fast Bayesian Intensity Estimation for the Permanental Process

提出快速贝叶斯推断算法，基于再生核希尔伯特空间，提升Permanental过程的效率。

Christian J. Walder, Adrian N. Bishop

2017-01-13 60

cs.CL 1701.03079

RUBER: An Unsupervised Method for Automatic Evaluation of Open-Domain Dialog Systems

RUBER结合参考和无参考指标，通过嵌入和神经网络实现无监督自动评估，相关性高达0.49。

Chongyang Tao, Lili Mou, Dongyan Zhao 等

2017-01-12 63

cs.CV 1701.03077

A General and Adaptive Robust Loss Function

提出一种可自适应的鲁棒损失函数，结合概率分布优化，提升视觉任务性能。

Jonathan T. Barron

2017-01-12 45

math.OC 1701.01945

A Framework for Wasserstein-1-Type Metrics

提出统一框架扩展Wasserstein-1，结合凸性与计算效率，涵盖多种非平衡测度。

Bernhard Schmitzer, Benedikt Wirth

2017-01-08 53

cs.CV 1612.08484

An Automated CNN Recommendation System for Image Classification Tasks

提出自动CNN推荐系统，基于任务复杂度和模型能力评分，免训练快速匹配最优模型。

Song Wang, Li Sun, Wei Fan 等

2016-12-27 53

cs.CL 1612.08083

Language Modeling with Gated Convolutional Networks

提出门控卷积网络（Gated Convolutional Networks）用于语言建模，超越LSTM，达到WikiText-103新SOTA。

Yann N. Dauphin, Angela Fan, Michael Auli 等

2016-12-24 59

cs.CV 1612.06573

Detecting Unexpected Obstacles for Self-Driving Cars: Fusing Deep Learning and Geometric Modeling

结合深度学习和几何建模，自主驾驶车辆障碍检测系统在Lost and Found数据集上提升50%。

Sebastian Ramos, Stefan Gehrig, Peter Pinggera 等

2016-12-20 50

cs.CL 1612.06027

Neural Multi-Source Morphological Reinflection

提出多源形态重塑模型，基于多编码器-解码器架构，在8语种上超越单源模型，提升性能。

Katharina Kann, Ryan Cotterell, Hinrich Schütze

2016-12-19 34

cs.SD 1612.04742

Imposing higher-level Structure in Polyphonic Music Generation using Convolutional Restricted Boltzmann Machines and Constraints

使用卷积受限玻尔兹曼机和约束优化生成多声部音乐，保留结构特性。

Stefan Lattner, Maarten Grachten, Gerhard Widmer

2016-12-15 13

cs.CL 1612.04426

Improving Neural Language Models with a Continuous Cache

提出连续缓存模型，利用隐藏激活实现高效动态记忆，显著提升神经语言模型性能。

Edouard Grave, Armand Joulin, Nicolas Usunier

2016-12-14 48

math.NA 1612.02672

Convergence rate of the data-independent $P$-greedy algorithm in kernel-based approximation

数据无关P-贪婪算法在Sobolev空间核逼近中的收敛率接近最优。

Gabriele Santin, Bernard Haasdonk

2016-12-08 9

cs.CV 1612.02646

Learning Video Object Segmentation from Static Images

本方法利用静态图像训练的卷积神经网络，通过引导策略实现视频目标分割，达到了竞争性性能。

Anna Khoreva, Federico Perazzi, Rodrigo Benenson 等

2016-12-08 16

cs.CV 1612.02401

DeMoN: Depth and Motion Network for Learning Monocular Stereo

DeMoN网络通过单目图像对学习深度和相机运动，显著提升鲁棒性和精度。

Benjamin Ummenhofer, Huizhong Zhou, Jonas Uhrig 等

2016-12-08 52

cs.CV 1612.01925

FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks

FlowNet2以数据课程、堆叠加权与小位移专网，将Sintel测试EPE降至3.96并实现123fps。

Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia 等

2016-12-07 20

math.OC 1612.01554

Robustness of Control Barrier Functions for Safety Critical Control

ZBF/ZCBF结合ISS与QP，证明扰动下安全集可放宽保持，且反馈局部Lipschitz。

Xiangru Xu, Paulo Tabuada, Jessy W. Grizzle 等

2016-12-06 24