March 30, 2017
Zachariah Miller is a Senior Data Scientist at Metis in Chicago.
_____
A few days ago, I came across a question on Quora that boiled down to: "How can I learn machine learning in six months?" I started to write up a short answer, but it quickly snowballed into a huge discussion of the pedagogical approach I used and how I made the transition from physics nerd to physics-nerd-with-machine-learning-in-his-toolbelt to data scientist. Here's a roadmap highlighting major points along the way.
Machine learning is a really large and quickly evolving field. It will be overwhelming just to get started. You've most likely been jumping in at the point where you want to use machine learning to build models - you have some idea of what you want to do; but when scanning the internet for possible algorithms, there are just too many options. That’s exactly how I started, and I floundered for quite some time. With the benefit of hindsight, I think the key is to start way further upstream. You need to understand what’s happening “under the hood” of all the various machine learning algorithms before you can be ready to really apply them to ‘real’ data. So let’s dive into that.
There are 3 overarching topical skill sets that make-up data science (well, actually many more, but 3 that are the root topics):
Realistically, you have to be ready to think about the mathematics before machine learning will make any sense. For instance, if you aren’t familiar with thinking in vector spaces and working with matrices then thinking about feature spaces, decision boundaries, etc. will be a real struggle. Those concepts are the entire idea behind classification algorithms for machine learning - so if you aren’t thinking about it correctly, those algorithms will seem extraordinarily complex. Beyond that, everything in machine learning is code driven. To get the data, you'll need code. To process the data, you'll need code. To interact with the machine learning algorithms, you'll need code (even if using algorithms someone else wrote).
The place to start is learning about linear algebra. MIT has an open course on Linear Algebra. This should introduce you to all the core concepts of linear algebra, and you should pay particular attention to vectors, matrix multiplication, determinants, and Eigenvector decomposition - all of which play pretty heavily as the cogs that make machine learning algorithms go. Also, making sure you understand things like Euclidean distances will be a major positive as well.
After that, calculus should be your next focus. Here we’re most interested in learning and understanding the meaning of derivatives, and how we can use them for optimization. There are tons of great calculus resources out there, but at a minimum, you should make sure to get through all topics in Single Variable Calculus and at least sections 1 and 2 of Multivariable Calculus. This is a great place to look into Gradient Descent - a great tool for many of the algorithms used for machine learning, which is just an application of partial derivatives.
Finally, you can dive into the programming aspect. I highly recommend Python, because it is widely supported with a lot of great, pre-built machine learning algorithms. There are tons of articles out there about the best way to learn Python, so I recommend doing some googling and finding a way that works for you. Make sure to learn about plotting libraries as well (for Python start with MatPlotLib and Seaborn). Another common option is the language R. It’s also widely supported and many folks use it - I just prefer Python. If using Python, start by installing Anaconda which is a really nice compendium of Python data science/machine learning tools, including scikit-learn, a great library of optimized/pre-built machine learning algorithms in a Python accessible wrapper.
This is where the fun begins. At this point, you’ll have the background needed to start looking at some data. Most machine learning projects have a very similar workflow:
From SciKit Learn's Documentation
For this stage in your journey, I highly recommend the book “Data Science from Scratch” by Joel Grus. If you’re trying to go it alone (not using MOOCs or bootcamps), this provides a nice, readable introduction to most of the algorithms and also teaches you how to code them up. He doesn’t really address the math side of things too much…just little nuggets that scrape the surface of the topics, so I highly recommend learning the math, then diving into the book. It should also give you a nice overview on all the different types of algorithms. For instance, classification vs regression. What type of classifier? His book touches on all of these and all shows you the guts of the algorithms in Python.
The key is to break it into digest-able bits and lay out a timeline for making your goal. I admit this isn’t the most fun way to view it, because it’s not as sexy to sit down and learn linear algebra as it is to do computer vision… but this can really get you on the right track.
Start with learning the math (2–3 months)
Move into programming tutorials purely on the language you’re using… don’t get caught up in the machine learning side of coding until you feel confident writing ‘regular’ code (1 month)
Start jumping into machine learning codes, following tutorials. Kaggle is an excellent resource for some great tutorials (see the Titanic data set). Pick an algorithm you see in tutorials and look up how to write it from scratch. Really dig into it. Follow along with tutorials using pre-made datasets like this: Tutorial To Implement k-Nearest Neighbors in Python From Scratch (1–2 months)
Really jump into one (or several) short term project(s) you are passionate about, but that aren’t super complex. Don’t try to cure cancer with data (yet)… maybe try to predict how successful a movie will be based on the actors they hired and the budget. Maybe try to predict all-stars in your favorite sport based on their stats (and the stats of all the previous all stars). (1+ month)
Sidenote: Don’t be afraid to fail. The majority of your time in machine learning will be spent trying to figure out why an algorithm didn’t pan out how you expected or why I got the error XYZ… that’s normal. Tenacity is key. Just go for it. If you think logistic regression might work… try it with a small set of data and see how it does. These early projects are a sandbox for learning the methods by failing - so make use of it and give everything a try that makes sense.
Then…if you’re keen to make a living doing machine learning - BLOG. Make a website that highlights all the projects you’ve worked on. Show how you did them. Show the end results. Make it pretty. Have nice visuals. Make it digest-able. Make a product that someone else can learn from and then hope that an employer can see all the work you put in.
_____
Metis teaches part-time professional development courses that cover topics like Machine Learning and much more. Check out an on-demand Live Online Machine Learning & AI Principles sample class here.
By: Emily Wilson // February 26, 2018
By: Emily Wilson // April 16, 2018
By: Zachariah Miller // March 01, 2018