注意下面很多链接都需要访问外国网站,无奈国情如此。
1. 学习MXNet的资源
1.1 用MXNet根据logo图片预测公司名称
Logo detection using Apache MXNet
链接:https://www.oreilly.com/ideas/logo-detection-using-apache-mxnet?imm_mid=0fb1e0&cmp=em-data-na-na-newsltr_ai_20180212
1.2 虽然我是PyTorch/TF派,但是还是推荐Amazon李沐大神的MXNet视频教程第一季完整版
百度云链接:https://pan.baidu.com/s/1mkeilnE#list/path=%2F
2. WildML blog用RL做trading炒股票/比特币。作者之前在Google Brain当resident(可惜没留下),他的newsletter非常值得subscribe,其实我贴的很多东西也是从他那里看到的
Introduction to Learning to Trade with Reinforcement Learning
链接:http://www.wildml.com/2018/02/introduction-to-learning-to-trade-with-reinforcement-learning/
3. 三个survey,都很长
3.1 CNN做图像分类,最经典的task
Deep Convolutional Neural Networks for Image Classification: A Comprehensive Review
链接:https://www.mitpressjournals.org/doi/pdf/10.1162/neco_a_00990
3.2 CV模型可解释性的一个survey
Visual Interpretability for Deep Learning: a Survey
链接:https://arxiv.org/pdf/1802.00614.pdf
其中提到的这个LIME这个有点老,但是我觉得是个好工具,是在模型解释方面的工作。用局部线性模型去逼近classifier的decision boundary来取得解释性
blog:https://www.oreilly.com/learning/introduction-to-local-interpretable-model-agnostic-explanations-lime
repo链接:https://github.com/marcotcr/lime?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_source=The%20Wild%20Week%20in%20AI
3.3 online learning的一个巨长的survey,我不是搞这个方向也没有时间读完(?)
Online Learning: A Comprehensive Survey
链接:https://arxiv.org/pdf/1802.02871.pdf
4. Google的MobileNetV2,通过使用inverted residual connections和linear bottlenet在移动端加速,ImageNet效果比MobileNetV1, NasNet和ShuffleNet更快更准,也可以用来做feature extractor拿去做object detection和semantic segmentation
Inverted Residuals and Linear Bottlenecks: Mobile Networks for Classification, Detection and Segmentation
链接:https://arxiv.org/pdf/1801.04381.pdf
5. Facebook发布Tensor Comprehension,类似于TF XLA和TVM,把机器学习framework-level的代码自动生成优化过的高性能硬件MKL/cuDNN代码
Announcing Tensor Comprehensions
链接:https://research.fb.com/announcing-tensor-comprehensions/
6. 这个好,看舌头图片给开出重要配方。虽然算法上我觉得一般,但是是超实用的场景。我都想用这个做一个小程序 :)
链接:https://arxiv.org/pdf/1802.02203.pdf
7. neural architucture search相关
7.1 Google Brain用改进的进化算法做architure search搜索图片分类器,没有特别大的创新,但是效果还不错
Regularized Evolution for Image Classifier Architecture Search
链接:https://arxiv.org/pdf/1802.01548.pdf
7.2 Google Brain提出的Efficient NAS (ENAS),最关键的idea是站在计算图的角度来看,NAS的很多计算都是DAG这个大图的一部分,子模型的参数是可以复用的。结果就是搜索快了n倍,效果损失的不多。这个东西肯定很快就会被用到Google Cloud AutoML上了,绝对的省电省钱
Efficient Neural Architecture Search via Parameter Sharing
链接:https://arxiv.org/pdf/1802.03268.pdf
PyTorch实现
链接:https://github.com/carpedm20/ENAS-pytorch?utm_campaign=Revue%20newsletter&utm_medium=Newsletter&utm_source=The%20Wild%20Week%20in%20AI
7.3 ICLR一篇被接受的论文SMASH
SMASH: ONE-SHOT MODEL ARCHITECTURE SEARCH THROUGH HYPERNETWORKS
链接:https://openreview.net/pdf?id=rydeCEhs-
8. 神经网络pruning(此兄的文章每次都挺有深度,值得follow)
Pruning Neural Networks: Two Recent Papers
链接:http://www.inference.vc/pruning-neural-networks-two-recent-papers/
9. 实战好文,从PyTorch到CoreML再到React Native,前端后端ML一起搞定
How I Shipped a Neural Network on iOS with CoreML, PyTorch, and React Native
链接:https://attardi.org/pytorch-and-coreml
10. AI大佬们的对话
10.1 Yann LeCun, Eric Horvitz(微软)和Peter Norvig(Google/Stanford)的Reddit AMA
链接:https://www.reddit.com/r/science/comments/7yegux/aaas_ama_hi_were_researchers_from_google/
10.2 Stanford AI Lab办的AI Salon,终于放了一个视频,是Yann LeCun和Chris Manning(NLP大神)的对话,非常值得一看。不知道之后的视频会不会开放,可以去AI Salon网页跟进,他们的YouTube频道也值得订阅
What innate priors should we build into the architecture of deep learning systems?
链接:https://www.youtube.com/channel/UCpZRSrm2kmG8JvqP2Nqb0fQ/featured
*OpenAI跟一大堆机构写的一篇关于AI安全的长文。安全是个大问题
The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation
链接:https://arxiv.org/pdf/1802.07228.pdf