Machine Learning

My path learning machine learning

These resources describe my path toward learning ML (or an approximate revised version in hindsight), over approximately 6 months, in my free time as a physics postdoc. Not all steps are necessary. Not all resources need to be read in depth. This is tailored for someone with a physics background, with existing coding skills and math knowledge. 

Early-stage: basic familiarity (month 1).

• Basic neural networks: Karpathy lectures and Nielsen's book on neural networks.

• Machine learning for physicists review article (Pankaj Mehta et al), for a detailed mathematical introduction to general ML techniques. Worth a close read.

• Find which textbooks people highly recommend, and skim their tables of contents. Become familiar with all terminology and content on a superficial level.

• Twitter, for getting a sense of current state of the field.

• Systematic refresher on python basics.

Mid-stage: identifying core problems (month 2).

• Focus on identifying core problems and categorizing ways to solve them.

• Explore emerging technology.

• Begin exploring cutting-edge papers. Build up a queue for later, but get a rough sense of their content.  

• Read popular articles by trusted figures in the field, and get a sense for their perspective and their predictions.

• Interview friends, professors, and others to help identify important problems.

• Systematically refersh any missing basic knowledge or core competency (eg. statistics).

Late-stage: learning techniques, solving problems (months 3 - 6).

• Implement algorithms and networks yourself (play with transformers, implement the AlphaGo zero algorithm, etc.).

• Deep dive core recent papers in the field.

• Online courses and textbooks. 

  • (I have not thoroughly read each of these, but they represent the best I've encountered for this stage of learning.)

  • (Skim or read in depth, and dive into the papers referenced in the classes when relevant.)

  • Find a hands-on course in reinforcement learning. I did Delta Academy, which is unfortunately no longer being run. 

  • Deep reinforcement learning course CS285 at Berkeley (Sergey Levine).

  • Deep unsupervised learning course CS294 at Berkeley (Pieter Abbeel).

  • Stanford CS 25 on transformers (Chris Manning).

  • OpenAI spinning up in deep RL.

  • Deep learning book (Goodfellow).

  • Reinforcement learning textbook (Sutton and Barto).

• Work on your own independent research projects.