机器学习人工学weekly-2018/3/17

1. PyTorch构架分析

PyTorch – Internal Architecture Tour

链接：http://blog.christianperone.com/2018/03/pytorch-internal-architecture-tour/

2. OpenAI新的meta-learning算法Reptile，用shortest descent算法加快learn2learn

Reptile: A Scalable Meta-Learning Algorithm

链接：https://blog.openai.com/reptile/

3. Berkeley提出RL并行化框架，大大加快了训练速度，Atari只要几分钟就训练完

Accelerated Methods for Deep Reinforcement Learning

链接：https://arxiv.org/abs/1803.02811

4. CNN再次在sequence modeling上打脸RNN

An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling

链接：https://arxiv.org/pdf/1803.01271.pdf

5. BAIR blog用GAN做字体的style transfer

Transfer Your Font Style with GANs

链接：http://bair.berkeley.edu/blog/2018/03/13/mcgan/

6. 用神经网络去做计算机内存的预读取

Learning Memory Access Patterns

链接：https://arxiv.org/pdf/1803.02329.pdf

7. Airbnb学listing embedding

Listing Embeddings for Similar Listing Recommendations and Real-time Personalization in Search

链接：https://medium.com/airbnb-engineering/listing-embeddings-for-similar-listing-recommendations-and-real-time-personalization-in-search-601172f7603e

8. 分析深度对训练速度的影响，overparameterization看来还是有好处的

Can increasing depth serve to accelerate optimization?

链接：http://www.offconvex.org/2018/03/02/acceleration-overparameterization/?utm_campaign=buffer&utm_content=buffere8a58&utm_medium=social&utm_source=twitter.com

9.（文章略老）Wolfram的神经网络后端用的MXNet，这个挺好玩

Apache MXNet in the Wolfram Language

链接：https://www.oreilly.com/ideas/apache-mxnet-in-the-wolfram-language

10. ML模拟可视化工具，包括LR, MLP，RNN和Q-learning，做的很漂亮

链接：https://www.mladdict.com/