Getting started in machine learning (ML) is not the easiest of paths - but neither is it the hardest. It seems that most of the difficulty in getting started is getting positive feedback from the work that you're doing. At times it feels like you've been working on something for days and yet you still have nothing to show for it. In this post I'm going to outline a brief path into practical machine learning.
Prerequisites
Most online courses list linear algebra and introductory calculus as necessary prerequisites. While these may be useful for further research, they are certainly not required for getting started. The only requirement that one needs is being mildly proficient at programming - preferably in Python. Applying machine learning is all about code but actually has surprisingly little compared to, say, web development. You therefore don't need to know that much programming to get started.
If you aren't familiar with Python or any other programming language, you should try take an online course. I'd watch the first 5 or 6 lectures (at least until you learn what a function is) and then return to the actual machine learning, where you can piece together the rest of what you need. See some recommendations at the bottom of this post.
Getting Going
Now that you're reasonably confident programming, start with the brilliant fast.ai deep learning course, taught by Jeremy Howard. This course is ridiculously good and gets you training state-of-the-art deep learning models within the first lesson. A vital note here is that no matter what don't stop: if you get stuck on some lesson (even if it's the first), just carry on. I've often been tempted to rewatch whole parts of the lesson before continuing to the next one to try and understand it all. Often, however, in the following lesson, Jeremy will go over the previous lesson and usually all the parts that I wasn't sure about are cleared up. Think of it like a real-life lecture: watch it from start to finish in one sitting, and if you lose concentration, try make up for it in the next lesson. Once you've watched all the videos, you'll have the big picture of deep learning (and hopefully a fair bit of practical experience) and can then go and rewatch videos if you want.
After every lesson you should be trying out the accompanying notebooks at the very least. Go along with the challenges Jeremy puts out each week. Soon you'll be looking for ways you can apply your new skills to all sorts of topics. Keep a record of these and when you're feeling confident enough (after about lesson 4) try build your own application from your own dataset.
I simply can't stress this enough: If you are not coding, you have not learnt anything.
A good way to stay responsible is to work in a group. Look for a local study group on the fast.ai forums that you might be able to join or start your own. Invite some friends to keep you accountable or post your work onto the forum's "share your work here" thread. The forum is a great place to find or ask any question, whether about the course or ML in general.
Continuing the Journey
I strongly believe that starting the journey - really starting it - is the hardest part. Getting going, trying out the code and not giving up for long enough until you're hooked. Once you've reached this point, however, you can start to branch out. If you felt that fast.ai was too hard then you can go through it again or if too easy then you can choose something harder.
Either way, my suggestion is to start a project. Do something that you will enjoy doing and make it good. Put it up on GitHub and get people to use it. Show your family and your friends too. After going through fast.ai, you would have seen all the amazing things that people created - now it's your turn.
The best project ideas are ones that you are personally passionate about, however, just to whet your appetite, here are some places you can look for projects to work on:
- Pick a Kaggle competition and work on it as hard as you can. Try to place in the top 10%.
- Try reimplement a recent paper.
- Get some inspiration from a GitHub search. For example, there's a whole list of ideas here.
Then if you still want to learn more, if you want to get down into the weeds of machine learning here's my suggestions. If you enjoyed fast.ai, do the machine learning course of fast.ai. After this, try part 2 of the deep learning course, which is more in depth than part 1. If fast.ai's style isn't for you and you prefer something more formal, try Stanford's CS231n, a convolutional neural network course. It has really good lecture videos, notes and assignments and covers things from a more mathematical perspective. Don't let that scare you though - it is still very approachable. Then, optionally, go through MIT OCW's multi-variable calculus and linear algebra if that's your thing and you want to get into more of the maths. During this you can skim over the Deep Learning book, which also covers most of the requisite linear algebra. Additionally try read papers that come up when you're researching.
Summary
- Get started with fast.ai. You don't need any maths knowledge - just know some programming.
- Work on mini projects whilst learning.
- Engage on forums or create/join a local study group.
- Pick a larger project that you want to work on once you feel comfortable. You will learn the most from this so do it as soon as you can.