NeurIPS (Conference and Workshop on Neural Information Processing Systems, Neural Information Processing Systems Conference) in 1986, presented at the annual closed-door forum is calculated from the California Institute of Technology and Snowbird Bell Labs organization of neural networks, originally designed for biological research and exploration artificial neural networks and complementary open interdisciplinary conference. NeurIPS General Assembly in recent years has been to machine learning, artificial intelligence and statistical paper based, is recognized as the premier meeting in the field of machine learning, which is comparable to the heat of the Olympic Games. Eye meetings offer special column this week as you continue to report 2019 News NeurIPS, so stay tuned!
Tutorials and went through the opening dinner yesterday! NeurIPS 2019 formal meeting today a great debut: eight Orals, 196 field Spotlights, 63 keynote Posters, 2 games Invited talks to bring you an academic feast! Also includes new direction Outstanding Paper Prize Paper Award and classic works show! Fast follower assistant to the scene to check it out now! (The report material captured by the eye doctor special guest session "Liu Guiliang")
Oral
Uniform convergence may be unable to explain generalization in deep learning
Author: Vaihnavh Negarajan, J.Zico K
Organization: Carnegie Mellon University, Center for Artificial Intelligence Bosch
small series Interpretation: evaluation rotation direction, the article can be negative consequences, but also a new direction was outstanding paper award!
Download the article: https: //arxiv.org/pdf/1902.04742v1.pdf
This paper shows the negative results on nature, indicating that many existing results (based on norm) performance limits the depth of learning algorithms did not reach their claims to be. The authors further demonstrated that when those investigators continue to rely on bilateral uniform convergence mechanism, they can not achieve the results they claim. While this paper does not address (also not pretend to solve) the problem of generalization depth of neural networks, but it is a "guidepost" guidelines depth study in this area to find a different direction.
Invited talk
Machine Learning Meets Single-Cell Biology：Insights and Challenges
Author: Dana Pe'er
institutions: Columbia University
Xiaobian interpretation: the top will have to encourage interbank machine learning dimensionality reduction combat, and so what!
Recent single-cell analysis technology is making a lot of data, a dataset of hundreds of thousands to millions of cells each will get out of thousands of variables. Single cell genome and the surge in single-cell imaging data created opportunities for the application of machine learning methods, can build a new map of the human cell biological significance of the discovery has huge potential. The speaker described the diversity of individual cells and cell populations composed of their function, mode of action and the impact of changes in the disease, and describes the use of these new data in health and disease computing environment, success stories and challenges currently facing .
Dual Averaging Method for Regularized Stochastic Learning and Online Optimization
Author: Lin Xiao
Organization: Microsoft Research
Xiaobian interpretation: Decade of sword! This article stand out from the 18 papers in 2009, he won the classic Paper Award!
Paper Download:
https://www.microsoft.com/en-us/research/wpcontent/uploads/2016/02/xiao10JMLR.pdf
该研究提出了用于在线最优化求解的RDA（Regularized Dual Averaging）方法，是Lin Xiao在微软10年前的研究成果。该方法是Simple Dual Averaging Scheme一个扩展，并更有效地提升了特征权重的稀疏性。
Poster
顶会Poster不轻松！相比于Oral和Spotlight，Poster这种近距离的交流显得更为亲民！NeurIPS 2019的海报展示是很吓人的。通常负责人会被一大群人层层围在海报前面，整个展厅也吵吵嚷嚷，基本离得稍远些就听不到声音了。今日份的盛况是这样的：
不过也不要担心跟不上进度，因为即便是很多不了解这个方向的人也会来到海报面前，咨询负责人：可以从头讲一遍吗！海报负责人其实也很喜欢这样的问题，他们之所以在这里，就是为了让更多人了解他们的研究。接下来小编趁着空档带大家一览几篇顶会海报大作！
AGEM:Solving Linear Inverse Problems via Deep Priors and Sampling
作者：Bichuan Guo、Yuxing Han、Jiangtao Wen
机构：清华大学
小编解读：又是一个将深度学习与贝叶斯统计相结合的示例。顶会总少不了国内理工一霸的身影！
文章下载地址：http://papers.nips.cc/paper/8345-agem-solving-linear-inverse-problems-via-deep-priors-and-sampling.pdf
作者建议在解决线性逆问题和估计其噪声参数之前，先使用降噪自动编码器（DAE）。现有的基于DAE的方法是根据经验来估算噪声参数，或将其视为可调超参数。而作者则建议使用自动编码器指导的EM，这是一种合理概率框架，可以执行具有难解的深层先验的贝叶斯推理。这样的结构可以通过Metropolis-Hastings从DAE获得有效的后验采样，从而可以使用Monte Carlo EM算法。最后文章展示了新方法在信号降噪，图像去模糊等方面具有竞争力的结果。
Reinforcement Learning on Online Markov Decision Processes with Global Concave Rewards
作者：Wang Chi Cheung
机构：新加坡国立大学
小编解读：一个人的论文，一个人的Poster！无敌是多么寂寞！
文章下载地址：
https://arxiv.org/pdf/1905.06466.pdf
我们构建了一个参与马尔可夫决策过程并在每个回合都得到结果向量的代理层。它的目标是在平均矢量结果上最大化全局凹面奖励函数。作者利用该模型在马尔可夫环境中创建了多目标优化，最大熵探索和约束优化等应用，提出了一种基于在线凸优化（OCO）工具（Agrawal和Devanur 2014）和UCRL2（Jaksch等人2010）算法，并且引入了一种新颖的梯度阈值流程，通过控制动作之间的切换可以处理微妙的权衡。通过延迟梯度更新，该流程将产生非平稳策略，从而使结果多样化来优化目标，流程与各种OCO工具兼容。
当然，其他的poster也都有很高的热度，小编就不在这里一一介绍了。感谢他们的耐心解答和辛苦付出，为计算机科学甚至其他学科领域知识的交流与传播做出了贡献！
