John Carmack让Ilya Sutskever推荐关于了解当下人工智能的阅读资料. Ilya Sutskever给了大约有40篇研究论文的链接.

If you really learn all of these, you’ll know 90% of what matters today. - Ilya Sutskever

如果你真的学会了这些，你就掌握了当今90%重要的内容 - Ilya Sutskever

大概在一个月以前, 推特上流传一份文件夹包含30篇资料(论文, 书籍, 博客)包含了这些资料.

阅读列表

按照发布的时间线从旧到新排列如下:

Time
Paper
1993Keeping Neural Networks Simple by Minimizing the Description Length of the Weights
2004A Tutorial Introduction to the Minimum Description Length Principle
2008Machine Super Intelligence
2011The First Law of Complexodynamics
2012ImageNet Classification with Deep Convolutional Neural Networks
2014Quantifying the Rise and Fall of Complexity in Closed Systems: The Coffee Automaton
2014Neural Turing Machines
2015The Unreasonable Effectiveness of Recurrent Neural Networks
2015Understanding LSTM Networks
2015Recurrent Neural Network Regularization
2015Deep Residual Learning for Image Recognition
2015Neural Machine Translation by Jointly Learning to Align and Translate
2015Deep Speech 2: End-to-End Speech Recognition in English and Mandarin
2016Order Matters: Sequence to Sequence for Sets
2016Multi-Scale Context Aggregation by Dilated Convolutions
2016Identity Mappings in Deep Residual Networks
2017Pointer Networks
2017Neural Message Passing for Quantum Chemistry
2017Attention Is All You Need
2017A simple neural network module for relational reasoning
2017Variational Lossy Autoencoder
2017Kolmogorov Complexity and Algorithmic Randomness
2018Relational Recurrent Neural Networks
2019GPipe: Easy Scaling with Micro-Batch Pipeline Parallelism
2020Scaling Laws for Neural Language Models
2022The Annotated Transformer
2024CS231n: Convolutional Neural Networks for Visual Recognition

计划

写30篇文章深度解析Ilya推荐的每一份资料, 并深度实践每一篇资料的概念.

为什么要做这个事

  • 现在关于贩卖人工智能焦虑的文章越来越多. 真正有价值的文章都会有特定阅读群体. 这个系列定位是所有人.
  • 互联网上的信息总是零散的多, 能够串联起来形成完成的线的文章少.
  • 中文互联网需要更多有价值的内容, 我认为这个过程是有价值的.
  • 对自己来说也是一个学习的过程, 检验自己是否真的理解. 任何经不起质疑的观点都是错的.

前提

虽然说是从零, 但是不会包括编程的教学.

  • 需要能够基本使用Python.
  • 有一定英文阅读能力

结果

看完这30篇文章并实践之后, 你能够:

  • 了解每一个篇文章提出的背景, 定义, 原理, 实现, 应用.
  • 能够通过代码实现每一个步骤, 并能论证作者的推理. 每一篇文章都会有对应的Jupter Notebook.
  • 按照时间线理解30篇文章, 所以会知道发展的脉络. 对未来的发展方向会有更清晰的理解.
  • 按照Ilya说的, 能够理解90%的人工智能.

相关链接