Introduction to Machine Learning for Coders!

This is a worthwhile course, but it's important to understand its pros and cons. I had initially heard about fast.ai through its deep learning course, but decided to take this one as a refresher before I jumped in.

I completed it a few weeks ago, and so far here are my thoughts.

THE GOOD:

1). It's style of instruction is a nice break from the way most MOOC's are typically taught. Jeremy goes over material in a very straightforward manner, and he's good at breaking down topics in a way that's easy to understand. I thought the classroom format was also fun, and hearing people ask questions and listening to Jeremy's answers was very engaging. In fact, it might have been my favorite part of the class. This is a very good way to get a practitioners style of thinking about ML problems, and the videos are engaging.

2). The material is very differentiated. Most MOOC's give you a somewhat whitewashed introduction to 1 ML concept before moving onto the next one, and just repeat that sequence until you're done with the class. This course is much more opinionated, and as a result presents material with very interesting edges that are difficult to find elsewhere. I thought the PDP plots/tree interpreter utilities were very interesting, as well as the from-scratch Random Forest. So as a collection, the material represents a basket of insights that don't have a lot of direct substitutes.

3). The use of Kaggle as a teaching tool is useful, and it's fun to practice the insights from the class on your own competitions.

THE BAD

1). I don't like the FastAi library, and wish the class didn't rely on it so heavily in its teaching. In my opinion its use obscures what's actually going on, and I would have much preferred explanations and code just written in pure Pandas/Numpy/Matplotlib, and spent a lot of my time digging through the source code to make sure I understood what was actually happening.

2). On a somewhat related note, getting the fastai library up and running can be a pain, especially if you're using a Windows computer (which I am). It basically wasn't designed for Windows and you have to spend time digging through a bunch of low-level environment variables to get things working, and that sucks. I basically quit the FastAI library altogether haflway through the course and just recreated the functions on my own.

3). The examples and results given in the class are a little staged and don't replicate as well to the real world, especially on Kaggle, mostly because a bunch of people are now doing it. Don't think the methods given are going to magically unlock hidden success in ML for you, even though it feels that way during some of the examples.

4). This is inverse of point #2 above.......but the material covered is somewhat narrow. This can also be good depending on what you're looking for, but the first 3/4 of the class basically goes over Random Forests, and the last portion briefly covers a 'Hello World' version of a neural network using PyTorch and NLP using WordEmbeddings/Naive Bayes. Things like GLM's, clustering, decompositions, etc aren't touched.

Not necessarily bad, but people should be prepared.

Overall I'm glad I took the course, but found myself wishing it didn't rely on its own internal libraries for so much of its teaching.