1
$\begingroup$

I am familiar with few deep learning models and I understand how (little bit why) CNN/RNN works. But I still cannot make sense of new research papers. I want to dive deeper into field of deep learning specifically reinforcement learning. How should I approach learning mathematics as a self-learner?

My Background:

  • I'm familiar with high school mathematics (basic integration/differentiation/limits etc.).
  • I know just enough of deep learning to understand what different layers/optimization functions means.
  • I have participated in one kaggle competition Deepfake Detection Challenge and got 132th place on leaderboard.
  • I have fairly good grasp on programming.

What Do I want to achieve:

  • I want to be able to read research papers and understand what they're trying to convey with intuition.
  • Able to reason about new reinforcement learning algorithms.
  • Design new algorithms

Currently What I'm planning to do (in order):

  • calculus from CALCULUS WITH ANALYTIC GEOMETRY
  • Linear algebra book by Gilbert S.
  • Pattern Recognition and Machine Learning Book by Christopher Bishop
  • Probability (suggestion from comments)
  • Statistics (suggestion from comments)
  • Optimization (suggestion from comments)
  • -> All with help from various MIT OCW and youtube videos

I value every suggestion and I'm ready to devote however much time it takes. Also I understand that I'm aiming too high here so I'm open to any advice related.

Improve this question
$\endgroup$
6
  • 1
    $\begingroup$ Add optimization to your list ... $\endgroup$
    – kjetil b halvorsen ♦
    Commented Feb 20, 2021 at 15:25
  • $\begingroup$ Add probability and statistics to your list. $\endgroup$
    – bogovicj
    Commented Feb 20, 2021 at 15:34
  • 1
    $\begingroup$ Hi: it's hard to read-learn things without a "reason" per-se so another approach is to try to read a good deep-learning or reinforcement learning book and then, every time you get stuck, go to the relevant topic that's sticking and read a book or chapter from it. i'm not sure if this is your preferred way but the other way, where you teach yourself topics can be difficult also because you don't know "why" you're learning them. I want to someday do what you're doing so, if you find a good deep or re-inforcement learning book and can let me know through this thread, it's appreciated. $\endgroup$
    – mlofton
    Commented Feb 20, 2021 at 18:07
  • $\begingroup$ I like the idea of reading while needed. Also even if I read all of this mathematics (from my plan), I'll have to stick to this approach as there will always be something missing practically. So now I have to rephrase my question as minimal maths required to read 90% of research papers without going back to any text/reference. I'm particularly interested in bottom up approach because 1. That's how it all originally developed 2. I don't have any hard deadlines. Definitely I'll be posting everything from all comments into an organized thread comment or question footer (I'll mention you). $\endgroup$
    – nkit
    Commented Feb 20, 2021 at 19:40
  • $\begingroup$ I would start with Bishop's "neural networks for pattern recognition". It is a bit dated, but the fundamentals are there. It is accessible, and you can see what you need to learn as you go. If, in addition, you learn some library (e.g. PyTorch), you can try things out as you go. $\endgroup$
    – jpmuc
    Commented Feb 20, 2021 at 21:37

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.