Jeremy Howard never ceases to amaze me. I have been a Machine Learning practitioner for the last 10-11 years and have taken the Fast AI course from 2017 and yet every time I come back to the course, I learn something new. This is the 4th time I am taking this course and I thought I would document everything digitally instead of doing it on pen and paper this time.
I have started with the Lesson 0 and would try to put in my notes and some extra comments from my side for every lesson. I am not sure how much time this is going to take with a world full of distractions, my day job building recommendation systems, understanding how LLM’s are going to change the world, helping (and getting helped) in getting my team to understand some of the basic concepts of Statistics which go a long way while working on Receommendation System stuff. But then, what the hell. Let me put some social pressure on myself and start this. As always, one of my cats Sobuj is taking the course alongside me. He last took David Silver’s Reinforcement Learning course with me back in 2018-19.
Lesson 0 - This is a lesson on how to do Fast AI which is fascinating to my mind because I have taken the course so many times that when I hear this lecture, I connect with so many things which are spoken about.
Here are my takeaways. I will add my own comments on this. Here is a link to this video.
A lot of this video is based on Radek’s (fastai alum) book.
Putting social pressure on yourself is alright. Writing all this as a part of substack is one way to do it for me. In one of my courses, I did it in a group. That work’s too.
Fast AI courses are great ways to learn basics of computer science, programming, calculus, linear algebra etc.
Lots of people spend time preparing to take the course. This is a never ending pursuit. Get into the course and on the way as you get stuck, read about the concept needed to understand the course material
A couple of good courses recommended as reference are CS50 for computer science and the missing semester course at MIT. I have not taken CS50 but would highly recommend the missing semester at MIT course. As a non-CS grad, it helped me scale up my game by a lot
Share the work which you do on blogs, repos etc. This is exactly what I am doing right now. Also, do some project. It doesn’t need to be unique. It can be something that has been done by someone else or it can be unique. I want to do something for my pets. I will do the same as the course goes along.
The sequence to do a Fast AI lesson is :
The two ways to do the course are -
Through Notebook server(Collab, Kaggle, Sagemaker etc)
Through a full blown linux server where you have create your own set up.
Read the Fast AI code and write lots of code. Become an even better developer than you are now
Start blogging about what you learn. While Jeremy recommends fastpages, at the time of my writing this blog, you might need to use Quarto
Machine learning code is notoriously hard to write and debug. I can confirm this from my own experience. A simple baseline is a good way to start.
Joining a Kaggle competition is a great way to skill up. I have been guilty of not doing so since 2014
Get involved in the Fast AI forums. I would highly recommend this. Most often than not you will get answers for eveything that you get stuck with.
Thats it then! I will be writing blogs ending in FAIxx for this course as I am reading a few other things at the same time. Stay tuned for some of my writings on Federated Learning, Causal Inference, Basic Probablity and Statistics, Recommender Systems and ofcouse LLM’s now.