机器学习人工学weekly-2018/3/4

注意下面很多链接需要科学上网,无奈国情如此

1. RL相关

1.1 inverse RL教程,第一部分就是讲Andrew Ng 20年前的奠基论文(我没读原论文,但是惊讶的发现居然全部是用的LP解的)。这个系列值得跟一下,我记得当时看Chelsea Finn的那篇GAN和IRL的论文完全懵逼,希望看完这个系列以后能懂

Inverse Reinforcement Learning pt. I

链接:https://thinkingwires.com/posts/2018-02-13-irl-tutorial-1.html

An Outsider’s Tour of Reinforcement Learning

链接:http://www.argmin.net/2018/01/29/taxonomy/

1.3 OpenAI发布新的模拟机器人的环境以及开源的Hindsight Experience Replay算法实现。HER主要是想解决reward shaping以及sparse reward/sample complexity这几个难题,比不成功的trajectory通过变换goal的方式也拿来学习,非常好的idea

Ingredients for Robotics Research

链接:https://blog.openai.com/ingredients-for-robotics-research/

1.4 DeepMind ToMnet (Thoery of Mind),用POMDP去建模其他agent的行为,这个方向好像工作很少,我觉得以后可能会被用到multi-agent的场景

Machine Theory of Mind

链接:https://arxiv.org/pdf/1802.07740.pdf

1.5 滴滴的一篇用multi-agent RL做车辆调度论文,通过reformulation降低state space维度,并把地理位置,车辆协同等等contextual信息纳入算法,很有意思。我觉得游戏agent也可以参考

Efficient Large-Scale Fleet Management via Multi-Agent Deep Reinforcement Learning

链接:https://arxiv.org/pdf/1802.06444.pdf

1.6 DeepMind UNICORN算法,尝试解决lifelong learning多任务问题

Unicorn: Continual Learning with a Universal, Off-policy Agent

链接:https://arxiv.org/pdf/1802.08294.pdf

1.7 几个RL相关的视频

1.7.1 Joelle Pineau在NIPS上将reproducibility,这个绝对是个大问题,换几个seed结果完全变了,github上相同算法不同实施代码结果也差别巨大,囧

Reproducibility in Deep Reinforcement Learning - Prof. Pineau - NIPS2017

链接:https://www.youtube.com/watch?v=TAMer41J038

1.7.2 Vlad Mnih(DeepMind RL大神,DQN/A3C都是他带头搞出来的)在Toronto讲多任务RL学习,最新的V-Trace off-policy并行算法

Volodymyr Mnih - Efficient Multi-Task Deep Reinforcement Learning

链接:https://www.youtube.com/watch?v=TfhV51cndPY

1.7.3 Ilya Sutskever (大神不用介绍了吧)讲meta-learning

Meta Learning and Self Play

链接:Meta Learning and Self Play

2. 这个教程不错,简短又cover了很多不是deep learning的东西

A Brief Introduction to Machine Learning for Engineers

链接:https://arxiv.org/abs/1709.02840

3. URLNet用character-level和word-level CNN去鉴别某个URL是不是恶意链接,这个可以被用到Chrome等浏览器里面 :)

URLNet: Learning a URL Representation with Deep Learning for Malicious URL Detection

链接:https://arxiv.org/pdf/1802.03162.pdf

4. NVidia提出的fast photo style transfer,用unpooling layer代替WCT里面的upsampling layer,让转化后的图效果变好,速度提高了60倍,很不错的结果

A Closed-form Solution to Photorealistic Image Stylization

链接:https://arxiv.org/pdf/1802.06474.pdf

5. 百度neural voice clone论文,没怎么研究这个领域,不过之前有个朋友跟我说过这个东西的应用,比如给小孩的玩具clone一下父母的生硬,或者给别人clone去世亲人的声音等等,应该还是有些场景可以用的上

Neural Voice Cloning with a Few Samples

链接:http://research.baidu.com/neural-voice-cloning-samples/

6.Batch normalization解释的比较清楚的一篇文章,想到Ali Rahimi在NIPS上说Wouldn't you want to know what 'internal covariate shift' means? 场下大笑 :)

Training very deep networks with Batchnorm

链接:http://rohanvarma.me/Batch-Norm/

7. 用机器学习预测ICO return,这个很酷

Predicting ICO returns with machine learning

链接:https://medium.com/@MLJARofficial/predicting-ico-returns-with-machine-learning-af6108ab9e39

8. Lifelong learning终生学习survey文

Continual Lifelong Learning with Neural Networks: A Review

链接:https://arxiv.org/abs/1802.07569?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_source=The%20Wild%20Week%20in%20AI

10. 工具/教程/学习资源

10.1 spaCy教程,自然语言处理工具,可以学学

链接:https://gist.github.com/aparrish/f21f6abbf2367e8eb23438558207e1c3

10.2 Propel机器学习JS库,可以调用GPU。JS项目github上很多,这个看起来比较靠谱

链接:http://propelml.org/

10.3 PyTorch简明教程

A Promenade of PyTorch

链接:http://www.goldsborough.me/ml/ai/python/2018/02/04/20-17-20-a_promenade_of_pytorch/

作者之前也写了TF简明教程

A Sweeping Tour of TensorFlow

链接:http://www.goldsborough.me/tensorflow/ml/ai/python/2017/06/28/20-21-45-a_sweeping_tour_of_tensorflow/

10.4 用Google Sheet做一个CNN,其实基本上就是fastai课程上将的内容,可以自己练一练

Building a Deep Neural Net In Google Sheets

链接:https://towardsdatascience.com/building-a-deep-neural-net-in-google-sheets-49cdaf466da0

10.5 DeepPavalov一个开源的end-to-end对话系统,感觉类似API.AI的backend,不过chatbot这玩意儿没火起来,可惜了

链接:https://github.com/deepmipt/DeepPavlov?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_source=The%20Wild%20Week%20in%20AI

10.6 Google把之前零散发布的AI学习资源整合起来了放到一个叫Learning with Google AI的项目上了,包括TF, Cloud, ML概念等等

链接:https://ai.google/education/#?modal_active=none

10.7 TF 1.6发布

链接:https://github.com/tensorflow/tensorflow/releases/tag/v1.6.0

*不懂生物,不过这个课看起来挺适合搞生物的同学

Stanford CS522 Machine Learning Approaches to Decode the Human Genome

链接:http://cs522.stanford.edu/notes/anshulkundaje.html

**Uber也学Google/Facebook搞AI Residency啦,有兴趣的同学可以去试试

Introducing the Uber AI Residency

链接:https://eng.uber.com/uber-ai-residency/?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_source=The%20Wild%20Week%20in%20AI

  • 发表于:
  • 原文链接http://kuaibao.qq.com/s/20180304A06ZUU00?refer=cp_1026
  • 腾讯「云+社区」是腾讯内容开放平台帐号(企鹅号)传播渠道之一,根据《腾讯内容开放平台服务协议》转载发布内容。

扫码关注云+社区

领取腾讯云代金券